Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

docs(marketing): add partner program packets and the $THREE utility season screenshot set

Nicholas12e658beac4a
diff --git a/marketing/partner-packets/README.md b/marketing/partner-packets/README.md
new file mode 100644
index 000000000..e1a9d2e9b
--- /dev/null
+++ b/marketing/partner-packets/README.md
@@ -0,0 +1,75 @@
+# Partner packets
+
+Forwardable request packets for partner and program opportunities that had no packet. Each one follows
+the rule in the [marketing operating plan](../OPERATING-PLAN.md): one-sentence pitch, 100-word abstract,
+working link, screenshot, bio where relevant, two proposed dates where relevant, the exact requested
+action, and approved relationship wording. Every packet also names the verified intake URL, the single
+step the owner takes, and each metric with its source and capture date.
+
+Researched, verified, and captured on **2026-09-16**. Sending, submitting, and posting remain owner
+actions; nothing here has been sent.
+
+## Index
+
+| Packet | Target | Deadline | The owner's one step | Status |
+|---|---|---|---|---|
+| [IBM Champions nomination](ibm-champions-nomination.md) | IBM Champions class of 2027 (individual) | Not published; 2027 applications are open now. Last cycle closed Nov 21, so treat **2026-11-20** as the latest safe date | The nominee submits the candidate form at `ibm.biz/beanIBMChampion` with the drafted answers | Ready after the nominee confirms name, title, country, and authorship of the npm connectors |
+| [IBM G2 review](ibm-g2-review.md) | G2, IBM watsonx.ai | None published | The real watsonx.ai user writes and submits the review in their own words with their own console screenshot | **Needs reviewer confirmation.** Hosted Granite is not configured today; past tense only |
+| [OpenAI share-your-story](openai-share-your-story.md) | OpenAI Stories form | None on form; pipeline target 2026-10-09 | After the App Directory review returns, fill name, email, city, state and submit the four answers | Drafted; hold for the review outcome |
+| [Google Cloud AI Agents](google-cloud-ai-agents.md) | Partner Network, AI Agents program, agent finder, Marketplace | None published; pipeline target 2026-09-25 | Enroll the company at `partners.cloud.google.com/enrollment` | Enrollment ready; the A2A listing is blocked by the Vertex billing denial and Agent Card gaps |
+| [ElevenLabs Startup Grants](elevenlabs-startup-grants.md) | ElevenLabs Startup Grants | Rolling; closing date not computable; pipeline target 2026-09-24 | Confirm headcount under 25, then submit the application | Ready if under 25 employees; Demo Day 2026-10-21 is already cast |
+| [Quicknode field note](quicknode-rpc-failover-field-note.md) | Quicknode Startup Program | Attach to the ask below | Re-check `/api/healthz` and `/api/version`, then attach | Ready |
+| [Quicknode spotlight ask](quicknode-spotlight-ask.md) | Quicknode Feature Friday or founder spotlight | None published; pipeline target 2026-10-09 | Reply in the existing program thread with the email and the field note | Ready; program contact name is not in the repo |
+| [Anthropic](anthropic.md) | MCP Registry, Claude plugin marketplace, Connectors Directory, Claude for Startups | None | Run the staged MCP Registry publish of three servers | Registry ready; plugin ready after re-validation; Connectors needs two owner decisions; no partnership exists |
+
+## Findings that change what we can claim
+
+These came out of verification and contradict current pipeline rows. They are repeated in the packet each
+one affects.
+
+- **Granite is not running in production.** `/api/ibm/galaxy` returns `watsonx_not_configured`; there are no
+ `WATSONX_*` variables on Cloud Run (absent since at least 2026-07-29, commit `2effca36a`). "Run Granite in
+ production" in [docs/partners/opportunities.md](../../docs/partners/opportunities.md) is stale.
+- **Vertex AI is billing-denied on the production project.** Logs on 2026-09-16 read "Lightning dunning
+ decision is deny for project: projects/93741856042". This blocks Google's Model Garden default
+ requirement for A2A agent listings and any Vertex Claude lane.
+- **Claude is the configured default brain, not the served one.** No Anthropic key, Vertex Claude flags off.
+- **ElevenLabs runs only with a user's own key.** No platform key; the Voice Lab copy implies platform
+ metering that cannot currently run.
+- **The Claude connector submission sheet is stale.** `/api/mcp` now has 61 tools, including a copy-trading
+ setup tool and an NFT mint, both review risks under the directory's financial-asset rule.
+- **`/ibm/hello` imitates ibm.com page chrome** (IBM wordmark and navigation). Its screenshot was
+ deliberately left out of the IBM packets; sending IBM a three.ws page styled as ibm.com invites a
+ brand-use objection.
+
+## Images
+
+All captured with Playwright from live pages on 2026-09-16 at 1600 by 900 and viewed to confirm they
+render real content.
+
+| File | Page | Used in |
+|---|---|---|
+| [ibm-community-group.png](images/ibm-community-group.png) | IBM Community, Three.ws User Group home (cropped to 720 px tall to remove the cookie banner) | IBM Champions |
+| [openai-page.png](images/openai-page.png) | `https://three.ws/openai` | OpenAI share-your-story |
+| [forge.png](images/forge.png) | `https://three.ws/forge` | Google Cloud |
+| [voice-lab-elevenlabs.png](images/voice-lab-elevenlabs.png) | `https://three.ws/voice`, "Use Your Own ElevenLabs Key" section | ElevenLabs |
+| [docs-solana-failover.png](images/docs-solana-failover.png) | `https://three.ws/docs/solana`, failover section | Quicknode |
+| [docs-mcp.png](images/docs-mcp.png) | `https://three.ws/docs/mcp` | Anthropic |
+
+The G2 packet has no image by design: G2 wants the reviewer's own screenshot of watsonx.ai.
+
+## Keeping these current
+
+Figures go stale fast. Re-run these on the day a packet is sent and update the numbers in place:
+
+```bash
+curl -s https://three.ws/api/platform/stats
+curl -s https://three.ws/api/version
+curl -s https://three.ws/api/healthz
+curl -s https://three.ws/api/tts/catalog
+curl -s https://three.ws/api/ibm/galaxy
+```
+
+Generation counts come from the production `forge_creations` table and are not public. When a packet is
+sent, record the date and any receipt in the packet's header and move the matching row in
+[opportunities.csv](../growth/opportunities.csv).
diff --git a/marketing/partner-packets/anthropic.md b/marketing/partner-packets/anthropic.md
new file mode 100644
index 000000000..3896a54e2
--- /dev/null
+++ b/marketing/partner-packets/anthropic.md
@@ -0,0 +1,170 @@
+# Anthropic: an honest packet
+
+**There is no Anthropic partnership, program membership, or co-marketing agreement.** Anthropic is not on
+`/partners`, and nothing in the repo records a contact at Anthropic. This packet covers the three
+self-serve distribution surfaces that are ready, the public co-marketing routes that actually exist
+(verified 2026-09-16), and a draft for the one route three.ws can use today.
+
+**Owner's one step:** publish the three staged servers to the official MCP Registry
+(`mcp-publisher login github`, review, then run
+[`prompts/store-submissions/_generated/mcp-registry-republish.sh`](../../prompts/store-submissions/_generated/mcp-registry-republish.sh)).
+It is the only surface with no eligibility or policy question left open; the other two each need an owner
+decision first, listed below.
+
+---
+
+## How three.ws actually depends on Claude (checked 2026-09-16)
+
+| Fact | Evidence |
+|---|---|
+| New agents default to an Anthropic brain | [`src/manifest.js`](../../src/manifest.js) sets `brain.provider: 'anthropic'` for a saved agent with no brain configured; all 8 agent templates in [`src/templates.js`](../../src/templates.js) use `claude-sonnet-4-6`; agent delegation defaults to `claude-haiku-4-5-20251001` |
+| The platform can serve Claude through Vertex AI on GCP credits | [`api/_lib/vertex-claude.js`](../../api/_lib/vertex-claude.js), `VERTEX_CLAUDE_PRIMARY` puts Vertex Claude at the head of the chain |
+| **Production does not serve Claude by default today** | Cloud Run has no `ANTHROPIC_API_KEY`; `VERTEX_CLAUDE_ENABLED=0` and `VERTEX_CLAUDE_PRIMARY=0`; Vertex is billing-denied on the project (logs 2026-09-16T16:15Z). A Claude-configured agent is served by the free fallback chain unless the user brings their own key |
+| The MCP servers are built for Claude clients | [three.ws/docs/mcp](https://three.ws/docs/mcp) opens with "lets Claude and other MCP-compatible AI systems interact with your three.ws account" |
+| Claude Code plugins and skills exist | [`.claude-plugin/marketplace.json`](../../.claude-plugin/marketplace.json) lists 4 plugins; `.agents/skills/` holds the skills pack |
+
+Say "built for Claude" and "agents default to Claude models", never "powered by Claude" or "runs on
+Claude", until a Claude lane serves production traffic.
+
+---
+
+## Dispatch checklist: the three ready surfaces
+
+### 1. Official MCP Registry
+
+- **Intake:** `mcp-publisher` CLI against https://registry.modelcontextprotocol.io
+- **Current state (queried 2026-09-16):** 72 distinct `io.github.nirholas/*` server names are already
+ live. The three staged servers, `herald-mcp`, `knock-mcp`, and `home-mcp`, are **not** in the registry.
+- [ ] `mcp-publisher login github` as the `nirholas` GitHub account
+- [ ] Read `prompts/store-submissions/_generated/mcp-registry-republish.sh` and confirm it publishes only
+ those three
+- [ ] Run it
+- [ ] Re-query `https://registry.modelcontextprotocol.io/v0/servers?search=nirholas` and record the three
+ new entries in [TRACKER.md](../../prompts/store-submissions/_generated/TRACKER.md)
+
+### 2. Claude plugin marketplace (community catalog)
+
+- **Intake:** https://platform.claude.com/plugins/submit (200 on 2026-09-16), which feeds
+ https://github.com/anthropics/claude-plugins-community. The curated official marketplace has no
+ public submission ([SUBMISSION.md](../../marketplace/plugins/three-ws-3d/SUBMISSION.md)).
+- **Submit only `three-ws-3d`.** The repo's own marketplace file carries four plugins; one of them is
+ third-party coin-launchpad tooling, which falls under the owner's commit and promotion gate for other
+ crypto projects and should not be submitted to Anthropic.
+- [ ] `claude plugin validate ./marketplace/plugins/three-ws-3d --strict` (the connector docs ask for this
+ before submitting)
+- [ ] Re-drive the install test in SUBMISSION.md section 3; the recorded evidence dates from 2026-07-08
+- [ ] Submit the form with the plugin path and repository URL
+- [ ] Record the submission in TRACKER.md
+
+### 3. Claude Connectors Directory
+
+- **Intake:** remote MCP servers submit **only** through the submission portal in Claude.ai
+ organization settings, which requires a **Team or Enterprise** organization and an Owner or Directory
+ permission ([submission docs](https://claude.com/docs/connectors/building/submission), read 2026-09-16).
+ The "public form" fallback in the older submission sheet applies to desktop extensions (MCPB), not to
+ remote servers. Escalations: `[email protected]`.
+- **Review criteria that bite** ([review criteria](https://claude.com/docs/connectors/building/review-criteria)):
+ connectors are not accepted if they "Transfer money, cryptocurrency, or other financial assets" or
+ "Generate images, video, or audio via AI models".
+- **The submission sheet is stale.** [claude-submission.md](../../prompts/store-submissions/_generated/claude-submission.md)
+ (2026-09-03) says `/api/mcp` has 53 tools; a live `tools/list` on 2026-09-16 returned **61 tools**
+ (45 read-only, 4 destructive, all titled). It also says the market tools "do not move funds or execute
+ trades", but the live list now includes a copy-trading setup tool (`copy_subscribe`) and an on-chain NFT
+ mint (`mint_3d_asset`). `/api/mcp-3d` has 36 tools including `persona_tip` and `persona_send`, which move
+ real funds.
+- **Discovery behaviour:** `POST /api/mcp` `tools/list` with `Accept: application/json` returns 200
+ without credentials; with the streamable-HTTP header `Accept: application/json, text/event-stream` it
+ returns 401 with an OAuth `WWW-Authenticate` challenge. Confirm which the portal sends before relying
+ on credential-free tool sync.
+- [ ] **Owner decision:** get a Team or Enterprise Claude organization with Owner access, or stop here
+- [ ] **Owner decision:** submit a scoped server with no fund-moving, trading, or minting tools. The
+ cleanest candidate is the free, keyless `/api/mcp-studio` (3D generation only, no payment surface),
+ but ask `[email protected]` first whether 3D model (GLB) generation counts as media generation
+- [ ] Regenerate claude-submission.md against the live server
+- [ ] Generate a real `MCP_REVIEW_SECRET` and decide reviewer credentials (tracker item 3)
+- [ ] Submit in the portal; record the submission id
+
+---
+
+## Public Anthropic co-marketing routes (researched 2026-09-16)
+
+| Route | Is it real and self-serve? | Fit for three.ws today |
+|---|---|---|
+| **Claude for Startups** | Yes. https://claude.com/programs/startups, apply at https://claude.com/form/startups-application. The page offers credits and priority rate limits to venture-backed startups ("Mention your investor when you apply"), plus community and events for founders | **Best route.** Credits would directly fund the Claude lane production lacks. Funding status is not recorded in the repo, so credit eligibility is unverified; the community side does not depend on it |
+| **Community meetups and demo nights** | Yes. The startups page links "Meetups and demo nights organized by founders building with Claude" at https://luma.com/claudecommunity and Anthropic-hosted events at https://www.anthropic.com/events | Good fit with the in-world event format; needs a named host and date |
+| **Customer stories** | Page exists at https://claude.com/customers with published stories. **No public submission form was found**; stories appear to originate with Anthropic's own teams | Not available as a cold submission. Revisit if an Anthropic account team exists |
+| **Official Claude plugin marketplace** | No public submission; Anthropic reaches out | Not available |
+| **Connectors Directory and MCP Registry listings** | Yes, covered above | Distribution, not co-marketing |
+
+---
+
+## Draft: Claude for Startups application
+
+The form fields were not opened (it is an external form). The program page says applicants should mention
+investors; third-party guides, not Anthropic, say applicants need a Claude Console account, company
+email, website, and a description. Use these blocks for whatever the form asks.
+
+**Company description**
+
+> three.ws is an open-source platform for 3D AI agents. It turns a text prompt into a rigged,
+> animation-ready 3D character, gives it a brain, a voice, and an on-chain identity, and embeds it on any
+> website with one tag. An agent with no brain configured defaults to a Claude model, our agent templates use Claude Sonnet, our MCP servers are documented for
+> Claude clients first, and we ship Claude Code plugins and skills.
+
+**What will you build with Claude**
+
+> Make Claude the default brain that actually serves every three.ws agent in production, not only the
+> configured default. Today agents are configured for Claude Sonnet but fall back to free models because
+> the platform has no funded Claude lane. Credits and priority rate limits would let every embedded agent
+> think, call tools, and speak through Claude, and let us publish a scoped 3D connector to the Claude
+> Connectors Directory.
+
+**Traction** (re-read on submission day)
+
+> 74,846 avatars, 3,817 agents, and 632 embedded widgets (public stats, 2026-09-16); 39,675 3D
+> generations started since June 2026; 72 servers on the official MCP Registry; OpenAI Select Partner
+> and IBM Business Partner.
+
+**Investors:** owner supplies. Not recorded in the repo.
+
+---
+
+## Packet fields
+
+**One-sentence pitch:** a 3D agent platform whose agents are configured for Claude and whose MCP servers
+are built for Claude clients, asking for the startup credits that would let Claude actually serve them.
+
+**100-word abstract:** three.ws turns a text prompt into a rigged 3D AI agent embeddable on any website.
+An agent with no brain configured defaults to a Claude model, its eight agent templates use Claude Sonnet, and its MCP
+servers and Claude Code plugins are built for Claude clients, with 72 servers already on the official MCP
+Registry. Production cannot yet serve that default: there is no funded Claude lane, so agents fall back
+to free models. A Claude for Startups award would close that gap and support a scoped submission of the
+free 3D connector to the Claude Connectors Directory. There is no existing Anthropic partnership.
+
+**Working link:** https://three.ws/docs/mcp (200 on 2026-09-16)
+
+**Screenshot:** [images/docs-mcp.png](images/docs-mcp.png), captured 2026-09-16. Alt text: "three.ws MCP
+Integration docs explaining how Claude can list, render, validate, and inspect 3D avatars."
+
+**Founder bio:** not in the repo; owner supplies.
+
+**Two proposed dates:** not applicable for the application. If hosting a community demo night, propose
+dates only after a host and venue (in-world or physical) are named.
+
+**Requested action:** accept three.ws into Claude for Startups.
+
+**Approved relationship wording:** none exists. Allowed: "three.ws agents default to Claude models",
+"MCP servers built for Claude", "listed on the official MCP Registry". Not allowed: "Anthropic partner",
+"in partnership with Anthropic", "powered by Claude" (production does not serve Claude by default), or any
+Anthropic logo lockup. After acceptance: "a member of Claude for Startups".
+
+## Metrics used
+
+| Metric | Value | Source | Captured |
+|---|---|---|---|
+| Servers on the official MCP Registry | 72 distinct names | Registry API, paginated search for `nirholas` | 2026-09-16 |
+| Staged, unpublished registry servers | 3 | Same query, plus TRACKER.md | 2026-09-16 |
+| `/api/mcp` tools | 61 (45 read-only, 4 destructive, 61 titled) | Live `tools/list` | 2026-09-16 |
+| `/api/mcp-3d` tools | 36 (16 read-only, 2 destructive) | Live `tools/list` | 2026-09-16 |
+| Avatars / agents / widgets | 74,846 / 3,817 / 632 | `https://three.ws/api/platform/stats` | 2026-09-16T16:02Z |
+| Claude-served production traffic | **Not measurable: no Claude lane is enabled** | Cloud Run env | 2026-09-16 |
diff --git a/marketing/partner-packets/elevenlabs-startup-grants.md b/marketing/partner-packets/elevenlabs-startup-grants.md
new file mode 100644
index 000000000..08c0e5a6d
--- /dev/null
+++ b/marketing/partner-packets/elevenlabs-startup-grants.md
@@ -0,0 +1,133 @@
+# ElevenLabs Startup Grants
+
+**Target:** ElevenLabs Startup Grants program
+**Program page:** https://elevenlabs.io/startup-grants
+**Verified intake:** https://elevenlabs.io/grants-application (terms page with an "I agree" checkbox and
+"Proceed to application"; read 2026-09-16). The application questions sit behind that checkbox and
+were not opened, because ticking it is accepting terms on the company's behalf. The program page says
+the application covers "the startup you're building, your team, how you expect to grow".
+**Deadline:** rolling. The terms say the program "will remain open for applications for a period of six
+(6) months, subject to extension or early termination", with no start date given, so the closing date
+cannot be computed. Decisions within about one week. Pipeline target **2026-09-24**.
+**Owner's one step:** confirm the employee count is under 25, then tick the terms box and submit the
+answers below from a three.ws business email.
+
+---
+
+## Eligibility checklist (from the live grant page and terms, 2026-09-16)
+
+| Requirement (quoted or paraphrased from ElevenLabs) | three.ws position | Verdict |
+|---|---|---|
+| "fewer than twenty-five (25) employees, determined at the time of application" | Headcount is not recorded anywhere in the repo | **Owner must confirm** |
+| "Monetized product use case" with a business or monetization strategy | Voice Lab publishes prices: $0.50 per voice clone and $0.30 per 1,000 spoken characters on the ElevenLabs lane, charged to a prepaid credit balance ([`api/tts/eleven.js`](../../api/tts/eleven.js) meters every platform-key synthesis) | Meets |
+| "No short-term or one-off projects" | Platform live since at least June 2026 with public changelog | Meets |
+| Valid business email | `[email protected]` or `[email protected]` | Meets |
+| "No projects for minors: ... children aged 18 or under" | Not built for children, but the [Terms of Service](https://three.ws/legal/tos) allow users **13 and older** with guardian consent under 18 | **Grey area. Disclose the 13+ policy honestly if asked; do not describe the audience as adults-only** |
+| "No agencies or consulting firms" | Product company | Meets |
+| Existing enterprise customers ineligible | No ElevenLabs enterprise contract recorded in the repo | Meets as far as the repo shows |
+| One application per company; previous grant recipients ineligible | No prior grant recorded | Meets as far as the repo shows |
+| Not located in or organized under a sanctioned jurisdiction | Terms of Service are governed by Delaware law | Likely meets; owner confirms entity |
+| Use must comply with the ElevenLabs Prohibited Use Policy | Voice cloning is exposed to users; the policy was not reviewed line by line here | Owner reads https://elevenlabs.io/use-policy before applying |
+
+## Two findings the application must not paper over
+
+**1. There is no platform ElevenLabs key in production today.** `GET https://three.ws/api/tts/catalog`
+on 2026-09-16 returned the ElevenLabs lane as `available: false`, `byok: true`, reason "Add your own
+ElevenLabs key below", and the Cloud Run service carries no `ELEVENLABS_API_KEY`. Voice Lab's header
+counted "357 voices across 3 providers" (the free Edge, Gemini, and NVIDIA lanes). So ElevenLabs is fully
+integrated in code and usable today only with a user's own key. The grant would directly switch on the
+platform lane. Say that; it is a stronger and truer pitch than implying current ElevenLabs volume.
+
+**Side effect to fix separately:** the Voice Lab copy says ElevenLabs calls "on the three.ws account ...
+are metered to your credit balance". With no platform key that path cannot run, so the copy overstates
+what a visitor can do today.
+
+**2. Demo Day 2026 is already cast.** https://demoday.elevenlabs.io/ shows the "2nd annual Startup Grants
+Virtual Demo Day" on **Wednesday 2026-10-21, 6:00 PM CEST**, where "11 grant recipients will showcase".
+A grant accepted in late September is very unlikely to present in this edition. Ask to be considered
+for the next one; do not promise Demo Day in any announcement.
+
+---
+
+## Application answers
+
+**Company one-liner**
+
+> three.ws turns a text prompt or photo into a rigged 3D AI agent that speaks, gestures, and reacts on
+> any web page.
+
+**What are you building?**
+
+> three.ws is an open-source platform for embodied AI agents. A user describes a character, gets a
+> textured and rigged 3D avatar, gives it a mind and a voice, and embeds it on any site with one web
+> component. ElevenLabs is integrated into that agent flow: speech synthesis through our TTS proxy,
+> instant voice cloning from a short recording, the public Voice Library, and per-agent voices, with
+> Flash v2.5 as the default model for low latency and Turbo v2.5 and Multilingual v2 selectable. The
+> audio drives the visible performance: the avatar's mouth follows the speech in real time through
+> viseme lip-sync, so the voice becomes a face and body rather than an audio track beside a character.
+
+**How do you use (or plan to use) ElevenLabs?**
+
+> Today ElevenLabs runs on a bring-your-own-key basis: a user pastes their ElevenLabs key and every
+> synthesis and clone runs on their account. The integration is complete in production code (shared
+> client, cached voice catalog, cloning through the official SDK, voice deletion, per-agent voice
+> keys). The grant would fund the platform lane, so any visitor can hear an ElevenLabs voice on a
+> 3D agent without their own account. We would use it for three things: default voices for public
+> demo agents, voice cloning for creators, and a public browser demo where a cloned voice drives a
+> rigged avatar in real time.
+
+**Business model and growth**
+
+> Voice is already priced: $0.50 per clone and $0.30 per 1,000 characters on the ElevenLabs lane,
+> charged to a prepaid credit balance, alongside paid 3D generation and pay-per-call tools for
+> agents. Across the platform, 74,846 avatars and 3,817 agents exist and 632 widgets are embedded on
+> other sites (public stats, 2026-09-16). Growth comes from embeds: every embedded agent that speaks
+> with a premium voice is a metered, repeat use.
+
+**Team:** owner writes (names, roles, and headcount are not in the repo).
+
+**Anything else**
+
+> We would welcome consideration for a future Startup Grants Demo Day with the jump from voice agent
+> to visible digital human.
+
+---
+
+## Packet fields
+
+**One-sentence pitch:** ElevenLabs voices that drive a visible 3D body, embeddable on any web page,
+from a platform whose integration is finished and waiting on a platform key.
+
+**100-word abstract:** three.ws turns a prompt or photo into a rigged, embeddable 3D AI agent. Its
+production code integrates ElevenLabs speech synthesis, instant voice cloning, the Voice Library, and
+per-agent voices, defaulting to Flash v2.5 for real-time lip-synced avatars. Today the lane runs only
+with a user's own ElevenLabs key, because the platform has no ElevenLabs account of its own. A Startup
+Grant would switch on the platform lane for every visitor, power voiced demo agents, and fund a
+public browser demo where a cloned voice drives a rigged avatar. Voice is already monetized at $0.30
+per 1,000 characters and $0.50 per clone.
+
+**Working link:** https://three.ws/voice (200 on 2026-09-16)
+
+**Screenshot:** [images/voice-lab-elevenlabs.png](images/voice-lab-elevenlabs.png), captured 2026-09-16.
+Alt text: "The 'Use Your Own ElevenLabs Key' section of three.ws Voice Lab, listing per-clone and
+per-character pricing and an API key field."
+
+**Founder bio:** not in the repo; owner supplies for the Team field.
+
+**Two proposed dates:** not applicable (rolling application, no meeting requested).
+
+**Requested action:** award a Startup Grant to three.ws.
+
+**Approved relationship wording:** there is **no relationship with ElevenLabs**. Say "three.ws integrates
+the ElevenLabs API". After acceptance, say "three.ws received an ElevenLabs Startup Grant" and nothing
+stronger. Note that the grant terms license ElevenLabs to use the three.ws name and logo for promotion.
+
+## Metrics used
+
+| Metric | Value | Source | Captured |
+|---|---|---|---|
+| ElevenLabs lane state | `available: false`, BYOK only | `https://three.ws/api/tts/catalog` | 2026-09-16 |
+| Voices available without a key | 357 across 3 providers | Voice Lab page header | 2026-09-16 |
+| ElevenLabs pricing on platform | $0.50 per clone, $0.30 per 1,000 characters | Voice Lab page copy | 2026-09-16 |
+| Avatars / agents / embedded widgets | 74,846 / 3,817 / 632 | `https://three.ws/api/platform/stats` | 2026-09-16T16:02Z |
+| ElevenLabs characters synthesized or clones made | **Not measured.** No usage export exists and BYOK calls bill the user's own account | n/a | n/a |
diff --git a/marketing/partner-packets/google-cloud-ai-agents.md b/marketing/partner-packets/google-cloud-ai-agents.md
new file mode 100644
index 000000000..fa3b09fb1
--- /dev/null
+++ b/marketing/partner-packets/google-cloud-ai-agents.md
@@ -0,0 +1,171 @@
+# Google Cloud: Partner Network, AI Agents program, agent finder, Marketplace
+
+**Target:** Google Cloud Marketplace for agent builders, the Google Cloud AI Agents program, and the
+AI agent finder
+**Source of the path:** https://cloud.google.com/marketplace/sell ("Agent builders" tab, read in a real
+browser 2026-09-16)
+**Deadline:** none published. Pipeline target **2026-09-25** ([opportunities.csv](../growth/opportunities.csv)).
+**Owner's one step:** enroll the company in Google Cloud Partner Network at
+https://partners.cloud.google.com/enrollment, signed in with the Google account that should hold the
+Partner Administrator role. Everything after that needs a Partner Network login, which is why it is
+the one step.
+
+---
+
+## Read this first: two findings that change the ask
+
+**1. Vertex AI is denied on the production project today.** Cloud Run logs for `three-ws-api` on
+2026-09-16 at 16:14Z and 16:15Z read `vertex imagen failed, falling back: Lightning dunning decision is
+deny for project: projects/93741856042` (that is `aerial-vehicle-466722-p5`), and the Vertex reference
+image lane returned 403. Google's AI agent listing requirements say an A2A agent must "use Google
+foundation or 3rd party models hosted in Model Garden as a default"
+([requirements](https://docs.cloud.google.com/marketplace/docs/partners/ai-agents)). Production does not
+meet that today: Vertex is billing-denied and `VERTEX_CLAUDE_PRIMARY` is `0` on the service, so the
+default brain is the free third-party chain. **Clear the billing hold and make a Model Garden model the
+default before submitting the A2A agent type.** Partner Network enrollment itself does not depend on
+it, which is why enrollment is still the first step.
+
+**2. The public Agent Card is not yet a valid Marketplace Agent Card.** Read from
+`https://three.ws/.well-known/agent-card.json` on 2026-09-16:
+
+| Requirement | Current card | Gap |
+|---|---|---|
+| A2A Agent Card specification ([a2a-protocol.org](https://a2a-protocol.org/dev/specification/)) | Has `name`, `description`, `url`, `provider`, `version` 1.5.1, `capabilities`, `defaultInputModes`, `defaultOutputModes`, 10 `skills` | Missing `protocolVersion`, `preferredTransport`, and `securitySchemes`/`security` (uses the older `authentication` key) |
+| The card's `url` is the agent's A2A endpoint | `https://three.ws/` (the website) | The A2A JSON-RPC handler lives at `/api/agents/a2a-paid`; `POST /api/a2a` returns 404 |
+| Stored as a JSON file in Cloud Storage and attached in Producer Portal ([Agent Card doc](https://docs.cloud.google.com/marketplace/docs/partners/ai-agents/agent-card)) | Served from the website only | Upload a Marketplace copy to a Cloud Storage bucket in the project |
+| Customers are billed through Google Cloud Marketplace | Declares the x402 payments extension with `"required": true` | A Marketplace customer cannot be required to settle each call on-chain; the Marketplace variant must make that extension optional or omit it |
+| Passes the Agent ScoreCard thresholds | Never run | Run validation in Producer Portal after upload |
+
+Fixing the card is engineering work that needs no permission; it is listed here so the owner does not
+submit the current file.
+
+---
+
+## The verified intake path, in order
+
+From the "Agent builders" tab of the sell page and the linked docs, all resolved 2026-09-16:
+
+1. **Enroll in Partner Network:** https://partners.cloud.google.com/enrollment
+2. **Become a Marketplace vendor:** in Partner Network Hub choose "Initiate onboarding your product to
+ Marketplace", submit the architecture diagram and business inputs, then accept the
+ [Marketplace Vendor Agreement](https://cloud.google.com/terms/marketplace-vendor-agreement) (needs
+ Partner Administrator). Complete the project info form the Marketplace team sends to get Producer
+ Portal access. ([vendor steps](https://docs.cloud.google.com/marketplace/docs/partners/offer-products))
+3. **Apply for the Google Cloud AI Agents program:**
+ https://partners.cloud.google.com/resources/google-cloud-ai-agent-program (Google sign-in required;
+ the program page content could not be read logged out)
+4. **Onboard the agent in Producer Portal:** create the AI agent, attach the Agent Card from Cloud
+ Storage, add product details, add pricing (about 4 business days of review), integrate account
+ creation and Google sign-in, then submit for publication.
+ ([onboarding](https://docs.cloud.google.com/marketplace/docs/partners/ai-agents))
+5. **Agent finder discoverability:** https://cloud.withgoogle.com/agentfinder/ (linked from the sell
+ page under "Enhance discoverability"; the page does not publish a separate submission form, so ask
+ the assigned partner manager how an approved listing is routed into it)
+
+**Listing type decision.** Google offers "AI Agents as a Service (A2A protocol)" and "Software as a
+Service (SaaS)" for agents using other protocols. The A2A type is the stronger fit and the one that
+reaches Gemini Enterprise, but it carries the Model Garden default requirement above. SaaS is the
+fallback if the billing hold cannot be cleared quickly.
+
+**Pricing model decision.** Google supports Free, Subscription, Usage-based, and Combined. Recommend
+**Free** for the first listing: it matches the free, keyless generation lane that already exists, and
+it avoids rebuilding x402 per-call billing onto Marketplace metering before there is demand.
+
+---
+
+## Forwardable packet
+
+**One-sentence pitch:** three.ws is a production 3D agent platform that already runs on Cloud Run,
+Cloud Scheduler, and a Cloud Run GPU fleet, and wants to bring an A2A agent that turns text into rigged,
+embeddable 3D characters to Google Cloud Marketplace and the AI agent finder.
+
+**100-word abstract:** three.ws turns a text prompt or photo into a textured, rigged, animation-ready
+3D character and embeds it anywhere with one web component. Production runs on Google Cloud: one Cloud
+Run API service scaling from 6 to 100 instances, eight Cloud Run GPU services for 3D generation,
+rigging, and motion, and 120 Cloud Scheduler jobs. Since June 2026 it has started 39,675 generations,
+33,941 of them completed. It already publishes an Agent Card with ten skills and supports A2A. We want
+to enroll in Partner Network, join the AI Agents program, and list a Marketplace A2A agent that
+enterprise teams can discover through the agent finder.
+
+**Working link:** https://three.ws/forge (200 on 2026-09-16; the free lane needs no account)
+
+**Screenshot:** [images/forge.png](images/forge.png), captured 2026-09-16. Alt text: "The three.ws Forge
+page: a prompt box, quality and engine selectors, and a Generate button, free with no sign-up."
+
+**Two proposed dates** (intro call with the assigned partner manager after enrollment):
+**2026-09-30** or **2026-10-02**, any time 9 AM to 12 PM US Pacific. Proposals only; nothing is booked.
+
+**Requested action:** after enrollment, assign a partner manager and confirm the AI Agents program
+intake, including how an approved agent is routed into the AI agent finder and the Gemini Enterprise
+validation path.
+
+**Approved relationship wording** (from [docs/listings.md](../../docs/listings.md) and
+[docs/partners.md](../../docs/partners.md)): three.ws is a **member of Google Cloud for Web3 Startups**
+and runs its production stack on Google Cloud. It is **not yet a Google Cloud Partner** (no Partner
+Network enrollment) and has **no Marketplace listing**. The `/partners` card chip "Cloud" describes the
+technical dependency, not a program tier. Do not write "Google Cloud partner" or "partnership" in the
+outreach, even though the blog post's URL slug says `google-cloud-partnership`; its title is "three.ws
+Joins Google Cloud for Web3 Startups".
+
+---
+
+## Marketplace business case (draft for the "business inputs" step)
+
+**Product.** A 3D agent platform. The Marketplace offer is a hosted A2A agent that generates a textured
+3D model from text or images, rigs humanoid characters for animation, validates and inspects GLB files,
+and returns an embeddable viewer. Its skills are already declared in the public Agent Card:
+`list-avatars`, `get-avatar`, `search-avatars`, `render-avatar`, `delete-avatar`, `validate-model`,
+`inspect-model`, `optimize-model`, `inspect-glb-a2a`, `x402-service-catalog`. The Marketplace variant
+should drop `delete-avatar` from a first listing unless account linking is complete, and scope out
+`x402-service-catalog`.
+
+**Customer problem.** Enterprise product, marketing, and commerce teams want 3D assets and embodied
+agents but do not have 3D pipelines. Agents in Gemini Enterprise answer in text; this agent returns a
+3D object a user can inspect, place in AR, or embed.
+
+**Why Google Cloud Marketplace.** Production already runs on Google Cloud, so the customer's data and
+compute path does not change vendors. Marketplace procurement and commit drawdown remove the purchasing
+step that blocks enterprise pilots.
+
+**Architecture (for the diagram).**
+- `three-ws-api` on Cloud Run, us-central1: serves the frontend, route table, and every API handler;
+ min 6, max 100 instances (read from the service on 2026-09-16; live revision
+ `three-ws-api-00437-lsv`, commit `58224bd69`, per `https://three.ws/api/version`).
+- GPU generation fleet on Cloud Run: seven services on NVIDIA L4 (`model-trellis`, `model-hunyuan3d`,
+ `model-hunyuan3d-21`, `model-rig`, `model-text2motion`, `model-triposg`, `model-triposr`) and one on
+ NVIDIA RTX PRO 6000 (`model-hunyuan3d-21-rtx`), plus CPU services for background removal, remeshing,
+ segmentation, stylization, and reconstruction.
+- `three-ws-multiplayer` on Cloud Run for the live 3D world.
+- Cloud Scheduler: 120 jobs in us-central1, synced from the 117 cron definitions in `vercel.json`.
+- Cloud Build with a pinned build service account; runbook in
+ [docs/ops/gcp-production.md](../../docs/ops/gcp-production.md).
+- Vertex AI: Gemini and Imagen lanes, plus Claude on Vertex as an optional transport. **Currently
+ billing-denied**; restore before claiming it in the business case.
+
+**Evidence of production readiness.**
+
+| Metric | Value | Source | Captured |
+|---|---|---|---|
+| 90-day uptime, external probe every 5 minutes | 99.85% overall; Platform API 99.78% | `https://three.ws/status` | 2026-09-16 |
+| Generations started / completed | 39,675 / 33,941 since 2026-06-11 | Production database, `forge_creations` | 2026-09-16 |
+| Generations in the last 30 days | 22,638 | Same | 2026-09-16 |
+| Avatars / agents / embedded widgets | 74,846 / 3,817 / 632 | `https://three.ws/api/platform/stats` | 2026-09-16T16:02Z |
+| Agent Card skills | 10 | `https://three.ws/.well-known/agent-card.json` | 2026-09-16 |
+
+**Go to market.** A Free listing first, measured on activated Marketplace accounts and completed
+generations per account. After the first ten active accounts, propose a usage-based plan and a joint
+startup success story. Existing distribution that can point at the listing: the OpenAI Select Partner
+connector, the IBM Community user group, and the public docs.
+
+**What we would ask Google for after acceptance.** AI agent finder placement, Gemini Enterprise
+validation, and consideration for a Google Cloud startup success story once there is a measured
+Marketplace outcome ([growth register](../growth/opportunities.md) sets that trigger).
+
+---
+
+## Not verifiable from outside
+
+- The AI Agents program page content, eligibility, and application questions (behind Google sign-in).
+- Whether Partner Network enrollment has any tier or revenue threshold for the AI Agents program.
+- Whether the Google Cloud for Web3 Startups membership carries a partner-manager contact who can
+ shortcut step 1; nothing in the repo names one.
diff --git a/marketing/partner-packets/ibm-champions-nomination.md b/marketing/partner-packets/ibm-champions-nomination.md
new file mode 100644
index 000000000..d6e718547
--- /dev/null
+++ b/marketing/partner-packets/ibm-champions-nomination.md
@@ -0,0 +1,130 @@
+# IBM Champions 2027: self-nomination packet
+
+**Target:** IBM Champions program, class of 2027 (individual recognition, not a company program)
+**Program page:** https://www.ibm.com/community/champions-program/
+**Verified intake:** https://ibm.biz/beanIBMChampion, which redirects to the live Airtable form
+"IBM Champion Program - Candidate Application"
+(https://airtable.com/appuwf3eOGdO6x1oS/pag5EKa36AiYAo12z/form). Read on 2026-09-16: the form header
+says it is for "new applicants who wish to be considered for the 2027 IBM Champions program".
+**Deadline:** not published. No class-of-2027 announcement blog was found on 2026-09-16. The previous
+cycle ran 2025-09-15 to 2025-11-21 ([announcement](https://community.ibm.com/community/user/blogs/libby-ingrassia/2025/09/15/ibm-champions-nominations)).
+Treat **2026-11-20** as the latest safe date; submitting now also starts Rising Champions badge credit
+for this calendar year, per the form.
+**Owner's one step:** the nominee opens the form and submits the answers below under their own name.
+
+---
+
+## Stop and confirm before sending (three facts the repo cannot supply)
+
+1. **Who the nominee is.** The repo records **nich** as the person behind the IBM work: IBM Community
+ handle `nich8` (blog author), `nichxbt (three.ws)` (thread author), and `nich` / `[email protected]` as
+ the named contact on IBM's own event page. Legal first and last name, job title, and country of
+ residence are **not recorded anywhere in the repo**. The nominee types those.
+2. **Whether the npm connectors are the nominee's code.** Both IBM connector packages list
+ `nirholas` as author on the npm registry. Nothing in the repo proves `nirholas` and `nich` are the
+ same person. Cite the packages as the nominee's work only if that is true; otherwise cite them as
+ "three.ws team" work and drop them from the open-source answer.
+3. **The Granite production claim is not currently true.** See "Evidence gaps". Do not write "runs
+ Granite in production" in the bio or anywhere else on the form.
+
+IBM employees are ineligible (the form says so). The nominee is not an IBM employee per the repo.
+
+---
+
+## Forwardable summary
+
+**One-sentence pitch.** A three.ws builder who runs a dedicated IBM Community user group, led a live
+build session on IBM's Global AI and Data Science community, and publishes open-source watsonx.ai
+connectors for other developers.
+
+**100-word abstract.** Since June 2026 the nominee has advocated for IBM watsonx.ai and Granite in
+public, on IBM's own properties. They were the listed contact for "Building 3D AI Agents Live", a live
+technical session on IBM's Global AI and Data Science community calendar. They post in the Three.ws
+User Group on IBM Community, which has 11 threads, 6 blogs and 18 members, including a feature tour
+blog and a hands-on Forge walkthrough. The team they build with publishes two Apache-2.0 MCP connectors
+that let any MCP client reach watsonx.ai and Granite, downloaded 977 times in the 30 days to
+2026-09-11. They also organized the group's first in-world meetup.
+
+**Working link:** https://community.ibm.com/community/user/groups/community-home?communitykey=e71510cc-d953-408f-9a1c-019f5c0a7016
+
+**Screenshot:** [images/ibm-community-group.png](images/ibm-community-group.png), the public group home
+captured logged-out on 2026-09-16 (cookie banner cropped out). Alt text: "IBM Community page for the
+Three.ws User Group showing 11 threads, 6 blogs, 2 library entries and 18 members."
+
+**Two proposed dates:** not applicable (rolling form, no meeting requested).
+
+**Requested action:** consider the nominee for the IBM Champions class of 2027.
+
+**Approved relationship wording** (from [docs/ibm.md](../../docs/ibm.md)): three.ws is an **IBM Business
+Partner**. That is a designation, not an endorsement. The `/api/ibm/*` showcase and the
+`@three-ws/ibm-watsonx-mcp` connector are community-built, not IBM products, not official partnership
+deliverables, and not endorsed by IBM. The nomination is for a person; never phrase it as "three.ws
+is nominated".
+
+---
+
+## Form answers, in form order
+
+Every claim carries a public URL. Metrics were captured 2026-09-16 unless another date is stated.
+
+| Form field | Answer |
+|---|---|
+| First Name / Last Name | Nominee types their legal name (not in repo) |
+| Primary Email | `[email protected]` if that is the nominee's inbox; it is the contact IBM's event page shows. Do not use an ibm.com address |
+| IBM ID | The nominee's IBMid login for community.ibm.com, if different |
+| Enterprise email | Same as primary if on the three.ws domain |
+| Country of Residence | Nominee selects (not in repo) |
+| Organization in a different country? | Nominee selects |
+| Company | three.ws |
+| Job Title | Nominee types (not in repo) |
+| Company website | https://three.ws |
+| Business relationship to IBM | **IBM Business Partner** |
+| Activities enjoyed most | Select: Speaking at events, conferences, user groups, or meetups; Hosting/organizing/supporting user groups or meetups; Writing technical articles or blogs; Participating in open-source projects and writing useful code; Contributing and/or moderating online communities (IBM Community) |
+| Primary Product Skills | IBM watsonx.ai; IBM Granite (if the list offers Granite separately). Do not select watsonx Orchestrate: the Agent Connect listing is not live |
+| Social profiles | IBM Community; GitHub (https://github.com/nirholas/three.ws, only if item 2 above is confirmed); X (https://x.com/trythreews, the company account) |
+| User group role | **Leader** only if the nominee is a group leader in the group admin panel. The public page shows Jessica Swanson (IBM) plus one second leader whose name is hidden to logged-out visitors, so this could not be verified from outside. Otherwise select **Volunteer** |
+| Data privacy checkbox | Nominee confirms personally |
+
+### Nominee bio (paste, then add name and title)
+
+> [Nominee name] builds 3D AI agents at three.ws, an IBM Business Partner, and works in public on how
+> developers can combine IBM watsonx.ai and Granite with embodied, browser-native agents. In June 2026
+> they were the contact for "Building 3D AI Agents Live: From Prompt to Embeddable Agent in Minutes",
+> a live build session on IBM's Global AI and Data Science community. They post tutorials and feature
+> walkthroughs in the Three.ws User Group on IBM Community and organized its first in-world meetup in
+> August 2026. The team they build with maintains two Apache-2.0 MCP connectors that let Claude,
+> Cursor, and other MCP clients reach watsonx.ai with the developer's own IBM Cloud credentials.
+
+### Evidence to paste into the activities form after the Champion Program ID arrives
+
+The form says acts of advocacy go on a separate activities form once IBM emails the program ID.
+
+| IBM activity | Evidence (public URL) | Date | Verified 2026-09-16 |
+|---|---|---|---|
+| Speaking at an IBM community event | [Building 3D AI Agents Live](https://community.ibm.com/community/user/events/event-description?CalendarEventKey=2767712c-6efd-47dc-aaeb-019ec4126e27&CommunityKey=f1c2cf2b-28bf-4b68-8570-b239473dcbbc), Global AI and Data Science community (31.2K members), contact "nich, [email protected]" | The IBM page shows Wed 2026-06-24, 6:00 to 6:30 PM PT. [docs/ibm.md](../../docs/ibm.md) records 2026-06-23, 6 to 7 PM MT. Quote IBM's page | 200, contact name read from page |
+| Writing technical articles | [three.ws: The Free AI 3D Model Generator's 10 Best Features](https://community.ibm.com/community/user/blogs/nich8/2026/07/16/threews-the-free-ai-3d-model-generators-10-best), author nich8 | 2026-07-16 | 200 |
+| Writing technical articles | [Type a Sentence, Get a 3D Model: A Hands-On Walkthrough of the three.ws Forge](https://community.ibm.com/community/user/discussion/type-a-sentence-get-a-3d-model-a-hands-on-walkthrough-of-the-threews-forge-user-perspective), author nichxbt (three.ws) | about 2026-07-17 | 200 |
+| Organizing user-group events | [Community Meetup Inside three.ws](https://community.ibm.com/community/user/discussion/community-meetup-inside-threews-join-us-on-the-three-server-at-threewsplay), author nich8; meetup held 2026-08-07 | 2026-08-05 | 200 |
+| Supporting a user group | [Three.ws User Group](https://community.ibm.com/community/user/groups/community-home?communitykey=e71510cc-d953-408f-9a1c-019f5c0a7016): 11 threads, 6 blogs, 2 library entries, 18 members | ongoing | 200, counts read from page |
+| Open-source code | [`@three-ws/ibm-watsonx-mcp`](https://www.npmjs.com/package/@three-ws/ibm-watsonx-mcp) v0.2.2, 494 downloads 2026-08-13 to 2026-09-11; [`@three-ws/ibm-x402-mcp`](https://www.npmjs.com/package/@three-ws/ibm-x402-mcp) v1.1.4, 483 downloads same window | first published 2026-06-02 and 2026-06-13 | npm registry API (the npmjs.com page blocks scripted requests) |
+
+---
+
+## Evidence gaps, stated plainly
+
+- **Granite is not configured on the hosted platform today.** `GET https://three.ws/api/ibm/galaxy`
+ returned `{"available":false,"reason":"watsonx_not_configured"}` on 2026-09-16, and
+ `POST /api/watsonx/embed` answered from `nvidia/nemotron-3-embed-1b`, not Granite. The Cloud Run
+ service carries no `WATSONX_*` variables, and commit `2effca36a` (2026-07-29) records that the
+ credentials "exist nowhere". The pipeline row in
+ [docs/partners/opportunities.md](../../docs/partners/opportunities.md) that says "run Granite in
+ production" is therefore stale. Restoring watsonx credentials would make the product-feedback
+ criterion true again; until then, leave it out.
+- **No IBM certifications are recorded.** IBM's 2025 announcement asks for "clear evidence of
+ technical expertise (such as certifications)". None exist in the repo.
+- **Speaker name.** IBM's event page names the "three.ws team" and lists nich as contact; it does not
+ name an individual presenter. Describe the role as "contact and presenter" only if the nominee
+ presented.
+- **Tenure.** All IBM-property activity falls inside June to September 2026. That is a short record for
+ a program that recognizes sustained advocacy; the monthly group cadence (one post per week allowed,
+ last catalogued post 2026-08-08) is the cheapest way to lengthen it before the window closes.
diff --git a/marketing/partner-packets/ibm-g2-review.md b/marketing/partner-packets/ibm-g2-review.md
new file mode 100644
index 000000000..ce6af778b
--- /dev/null
+++ b/marketing/partner-packets/ibm-g2-review.md
@@ -0,0 +1,106 @@
+# G2 review draft: IBM watsonx.ai (with Granite)
+
+**Status: needs the reviewer's own confirmation.** This is a fact sheet and a first draft for a real
+watsonx.ai user at three.ws to rewrite in their own words. G2's review page says "Be Authentic... allowing
+[generative AI] to write everything does not benefit our community", and asks for a screenshot proving
+use of the software. Do not paste this verbatim.
+
+**Target:** G2 product page for IBM watsonx.ai
+**Verified intake:** https://www.g2.com/products/ibm-watsonx-ai/take_survey (renders "Login or create an
+account to review IBM watsonx.ai", read 2026-09-16). G2 also lists IBM Granite separately at
+https://www.g2.com/products/ibm-granite/take_survey; review one product, not both with the same text.
+**Deadline:** none published. The September 2026 Partner Plus mailer offered a USD 25 gift card for a G2
+review ([docs/partners/ibm-partner-plus.md](../../docs/partners/ibm-partner-plus.md)); the mailer itself
+is not in the repo, so its terms and closing date are unverified.
+**Owner's one step:** the person at three.ws who actually used watsonx.ai signs in to G2 with their work
+email, attaches their own watsonx.ai console screenshot, and submits a review written from the notes
+below.
+
+---
+
+## The honest framing, before any draft
+
+| Fact | Evidence | Captured |
+|---|---|---|
+| three.ws built a real watsonx.ai integration: IAM token exchange, chat, embeddings, TimeSeries forecasting, vision, and Guardian | [`api/_lib/watsonx.js`](../../api/_lib/watsonx.js), [`api/_lib/granite-guardian.js`](../../api/_lib/granite-guardian.js), [`api/_lib/watsonx-forecast.js`](../../api/_lib/watsonx-forecast.js); first commit "feat: watsonx integration for brain/chat + agent wiring" dated 2026-06-02 | repo |
+| Models wired | `ibm/granite-3-8b-instruct`, `ibm/granite-embedding-278m-multilingual`, `ibm/granite-ttm-512-96-r2` (plus 1024 and 1536 context variants), `ibm/granite-vision-3-2-2b`, `ibm/granite-guardian-3-8b` | [docs/ibm.md](../../docs/ibm.md) |
+| **The hosted platform is not calling watsonx.ai today** | `GET https://three.ws/api/ibm/galaxy` returned `watsonx_not_configured`; `/api/watsonx/embed` answered from an NVIDIA model; no `WATSONX_*` variables on the Cloud Run service; commit `2effca36a` on 2026-07-29 records the credentials "exist nowhere" | 2026-09-16 |
+| Open-source connectors that call watsonx.ai with the user's own credentials are published and used | `@three-ws/ibm-watsonx-mcp` 494 downloads and `@three-ws/ibm-x402-mcp` 483 downloads, 2026-08-13 to 2026-09-11 (npm downloads API) | 2026-09-16 |
+
+So the review must be written in the **past tense for hosted production use** (roughly June to July
+2026; the exact end date is not recorded) and must not say "we run Granite in production". If watsonx
+credentials are restored before the review is written, re-run `curl https://three.ws/api/ibm/galaxy` and
+update the tense.
+
+---
+
+## Draft notes for the G2 questions
+
+G2's survey text behind the login was not readable from a logged-out browser, so the prompts below are
+the three G2 open-text questions as they have long appeared on its review form. Confirm the exact
+wording on screen.
+
+**Review title (draft):**
+> Granite on watsonx.ai gave our 3D agents a governance layer, not just another chat model
+
+**What do you like best? (draft notes)**
+- Granite Guardian returns calibrated `Yes`/`No` log-probabilities, so we could score each named risk
+ (jailbreak, harm, social bias, violence, profanity, and more) and collapse them into an
+ allow/review/block decision before an agent took an autonomous action. That was more useful to us
+ than a single moderation flag.
+- The Granite TimeSeries (TinyTimeMixer) models encode context and horizon in the model id
+ (`ttm-512-96`, `-1024-96`, `-1536-96`), which made it trivial to pick the largest model the available
+ history could fill.
+- `granite-embedding-278m-multilingual` was good enough for semantic agent discovery and
+ impersonation checks on names and descriptions.
+- One IBM Cloud API key exchanged for a short-lived IAM token, cached for about an hour, covered chat,
+ embeddings, vision, and forecasting behind one client.
+
+**What do you dislike? (draft notes)**
+- The reviewer should supply real friction from their own experience. Code-level facts that may
+ help them remember: the IAM token exchange is an extra round-trip we had to cache and coalesce
+ across concurrent requests; region-specific hosts and a separate version contract for some
+ endpoints needed per-account overrides; watsonx.ai credentials are project-scoped, so every call
+ needs a project or space id alongside the key.
+- Say plainly if the credentials lapsed for a business reason; that is the most useful thing a reader
+ can learn from this review.
+
+**What problems is watsonx.ai solving and how is that benefiting you? (draft notes)**
+- A governed "brain" option for embodied 3D agents: an agent could think on Granite and have an
+ autonomous money action vetoed by Granite Guardian before it executed.
+- Forecast, narrate, and govern in one vendor stack for the Granite Oracle experiment, which then
+ notarized governed forecasts on Solana.
+- The open-source MCP connector lets other developers reach watsonx.ai from Claude, Cursor, or any MCP
+ client with their own credentials and no intermediary backend.
+
+**Ratings, likelihood to recommend, role, company size, time used:** reviewer only. Do not estimate
+these for them. Company size is not recorded in the repo.
+
+---
+
+## Packet fields
+
+**One-sentence pitch:** a practitioner review of watsonx.ai from a team that wired five Granite model
+families, including Guardian, into an embodied agent platform.
+
+**100-word abstract:** three.ws integrated IBM watsonx.ai in June 2026 and used Granite models for
+chat, multilingual embeddings, time-series forecasting, vision, and governance. Its most distinctive
+use was Granite Guardian as an action veto: log-probability risk scores collapsed into allow, review,
+or block before an AI agent could act. The team also published two open-source MCP connectors that
+reach watsonx.ai with a developer's own credentials, downloaded 977 times in the 30 days to
+2026-09-11. The hosted platform is not currently configured with watsonx credentials, so this review
+describes production use during mid-2026 and continuing developer use through the connectors.
+
+**Working link:** https://three.ws/docs/ibm (200 on 2026-09-16)
+
+**Screenshot:** G2 wants a screenshot of the reviewer using watsonx.ai itself (the IBM console). Only
+the reviewer can capture that; no three.ws page proves watsonx.ai usage today, so none is attached.
+
+**Two proposed dates:** not applicable.
+
+**Requested action:** publish the review on G2.
+
+**Approved relationship wording:** three.ws is an **IBM Business Partner**, a designation, not an
+endorsement. G2 asks reviewers to disclose relationships with the vendor: select the partner/reseller
+disclosure if G2 offers it, and do not describe the review as IBM-requested beyond what the mailer
+actually said.
diff --git a/marketing/partner-packets/images/docs-mcp.png b/marketing/partner-packets/images/docs-mcp.png
new file mode 100644
index 000000000..fa1733aa6
Binary files /dev/null and b/marketing/partner-packets/images/docs-mcp.png differ
diff --git a/marketing/partner-packets/images/docs-solana-failover.png b/marketing/partner-packets/images/docs-solana-failover.png
new file mode 100644
index 000000000..66692ca5c
Binary files /dev/null and b/marketing/partner-packets/images/docs-solana-failover.png differ
diff --git a/marketing/partner-packets/images/forge.png b/marketing/partner-packets/images/forge.png
new file mode 100644
index 000000000..8f5d09a4a
Binary files /dev/null and b/marketing/partner-packets/images/forge.png differ
diff --git a/marketing/partner-packets/images/ibm-community-group.png b/marketing/partner-packets/images/ibm-community-group.png
new file mode 100644
index 000000000..a4d2110a1
Binary files /dev/null and b/marketing/partner-packets/images/ibm-community-group.png differ
diff --git a/marketing/partner-packets/images/openai-page.png b/marketing/partner-packets/images/openai-page.png
new file mode 100644
index 000000000..ec36ee7ee
Binary files /dev/null and b/marketing/partner-packets/images/openai-page.png differ
diff --git a/marketing/partner-packets/images/voice-lab-elevenlabs.png b/marketing/partner-packets/images/voice-lab-elevenlabs.png
new file mode 100644
index 000000000..fef933634
Binary files /dev/null and b/marketing/partner-packets/images/voice-lab-elevenlabs.png differ
diff --git a/marketing/partner-packets/openai-share-your-story.md b/marketing/partner-packets/openai-share-your-story.md
new file mode 100644
index 000000000..ec47cd613
--- /dev/null
+++ b/marketing/partner-packets/openai-share-your-story.md
@@ -0,0 +1,152 @@
+# OpenAI "Share your story" submission
+
+**Target:** OpenAI Stories (the form feeds the "Discover other stories" collection on openai.com)
+**Verified intake:** https://openai.com/form/share-your-story/ (rendered in a real browser 2026-09-16;
+plain HTTP clients get a 403, so open it in a normal browser)
+**Deadline:** none on the form. Pipeline target is **2026-10-09**, after the Showcase and App Directory
+submissions ([opportunities.csv](../growth/opportunities.csv)).
+**Owner's one step:** after the ChatGPT App Directory review returns (submitted 2026-09-12, version
+1.0.0, status Review per [TRACKER.md](../../prompts/store-submissions/_generated/TRACKER.md)), fill in
+first name, last name, email, city and state, paste the four answers below, and submit.
+
+**Why wait for the review outcome:** the form asks how OpenAI products "helped". The strongest honest
+answer is a live directory listing plus usage that came through it, and neither exists yet (see
+"What cannot be measured yet").
+
+---
+
+## What the form asks (read live 2026-09-16)
+
+Intro: "We're collecting real stories from people using OpenAI products - what you're building,
+learning, or figuring out along the way."
+
+| Field | Required | Limit |
+|---|---|---|
+| First name, Last name, Email, City, State | Yes | |
+| Is this story about: You / Someone you know | Yes | |
+| Tell us about yourself or the person this is about | Yes | 1000 characters |
+| How do you/they use OpenAI products? | Yes | 1000 characters |
+| How has using OpenAI products helped you/them? | Yes | 1000 characters |
+| What makes your/their story special? | Yes | 1000 characters |
+| Any links you'd like to share? | No | |
+| Consent: contacted by OpenAI, submission used under its Privacy Policy | By submitting | |
+
+This is a **personal** story form, not a company case-study intake. Write it as the founder's story.
+The City and State fields assume a US location; the founder's location is not recorded in the repo.
+
+---
+
+## The answers (each checked under 1000 characters)
+
+**Is this story about:** You
+
+**Tell us about yourself or the person this is about** (about 610 characters)
+
+> I build three.ws, an open-source platform that gives AI a body. You type a sentence and get a
+> textured, rigged 3D character you can animate, place in your room in AR, or embed on any web page
+> with one tag. Making a 3D asset has traditionally meant specialist software and skills, while AI
+> assistants mostly answer in flat text. I want anyone who can describe a thing to be able to hold it,
+> move it, and share it. The platform is free to try with no account or API key, and the viewer,
+> runtime, and web component are open source.
+
+**How do you/they use OpenAI products?** (about 820 characters)
+
+> Two ways. First, as a distribution surface: we built a free 3D Studio connector for ChatGPT on the
+> Apps SDK and MCP. It gives ChatGPT eleven keyless tools that generate a 3D model from a prompt,
+> inspect it, rig and animate a character, and hand it off to AR, with an interactive viewer rendered
+> inline in the conversation. The same capability ships as a custom GPT in the GPT Store. We are an
+> OpenAI Select Partner in the OpenAI Partner Network, and the connector is in App Directory review.
+> Second, inside the product: OpenAI models are lanes in our model router for agent chat and
+> embeddings, and OpenAI voices are one of the text-to-speech options an avatar can speak with.
+
+**How has using OpenAI products helped you/them?** (about 700 characters)
+
+> It changed who can reach 3D. Before the connector, making a model meant finding our site and
+> learning an interface. Inside ChatGPT, someone describes an object in the conversation they are
+> already having, turns it in their hands a couple of minutes later, and can place it on their desk in
+> AR from their phone. Across all our surfaces, people have started 39,675 generations since June
+> 2026 and 33,941 of them finished as downloadable 3D models. Building to the Apps SDK review bar also
+> made the product better everywhere: we wrote honest loading, error, and timeout states, removed
+> anything that needed a key or a wallet from the free path, and documented every tool.
+
+**What makes your/their story special?** (about 640 characters)
+
+> 3D is usually the last thing people expect a chat assistant to do, and the hardest to do well:
+> meshes, textures, skeletons, and animation all have to work, not just look right in a thumbnail. We
+> made it free and keyless on purpose so that a student, a teacher, or a small shop owner can try it
+> without a card or an account. We also published the open response shape the connector uses, Spatial
+> MCP, under CC0 so any assistant can render 3D results the same way. The goal is not a demo: it is
+> making "describe it, then hold it" an ordinary thing to do in a conversation.
+
+**Any links you'd like to share?**
+
+> https://three.ws/openai
+> https://chatgpt.com/g/g-6a563a3b49a88191abf346245491a444-three-ws-3d-studio
+> https://three.ws/docs/mcp-studio
+> https://github.com/nirholas/three.ws
+
+All four returned 200 on 2026-09-16.
+
+**Before submitting, re-check two sentences:** "the connector is in App Directory review" (change to
+"listed in the ChatGPT app directory" with the listing URL once approved, or delete if rejected), and
+the generation counts (re-run the query below on submission day).
+
+---
+
+## Metrics, with source and capture date
+
+| Metric | Value | Source | Captured |
+|---|---|---|---|
+| Generations started, all surfaces | 39,675 since 2026-06-11 | Production database, `select count(*) from forge_creations` (not a public URL) | 2026-09-16 |
+| Generations finished as a model | 33,941 (5,732 failed, 2 in progress) | Same table, grouped by `status` | 2026-09-16 |
+| Generations in the last 30 days | 22,638 | Same table, `created_at > now() - interval '30 days'` | 2026-09-16 |
+| Avatars, agents, embedded widgets | 74,846 avatars, 3,817 agents, 632 widgets | `GET https://three.ws/api/platform/stats` (public) | 2026-09-16T16:02Z |
+| Tools in the keyless ChatGPT connector | 11 | [TRACKER.md](../../prompts/store-submissions/_generated/TRACKER.md), live `tools/list` | 2026-09-09 |
+| Keyless end-to-end generation through the connector | 5.40 MB GLB in 137 s | Same tracker row | 2026-09-09 |
+| Custom GPT publicly reachable | Yes, logged out | `curl` returned 200 | 2026-09-16 |
+
+## What cannot be measured yet
+
+- **ChatGPT-attributed usage.** The connector does not tag creations it starts
+ (`api/_mcp-studio/forge-client.js` supports an `x-forge-client` key, but no caller passes it), and a
+ Cloud Run log sample of 20,000 `POST /api/mcp-studio` requests between 2026-09-15T23:51Z and
+ 2026-09-16T16:02Z contained no user agent naming OpenAI or ChatGPT. The 39,675 figure is
+ platform-wide and the answer says so. Do not rewrite it as "through ChatGPT".
+- **Directory listing outcome.** In review since 2026-09-12.
+- Founder name, city, and state are not in the repo.
+- **Founder voice.** The answers are written in the first person from repo facts only. The founder
+ should add one personal sentence of their own (why they started, or a real user moment) rather than
+ accept a motive written for them. If they add a personal use of ChatGPT or Codex, it must be true.
+
+---
+
+## Packet fields
+
+**One-sentence pitch:** a founder who made 3D creation free and keyless inside ChatGPT, so anyone who
+can describe an object can hold it, animate it, and place it in their room.
+
+**100-word abstract:** three.ws is an open-source platform that turns a sentence into a textured,
+rigged 3D character. Its founder built a free 3D Studio connector for ChatGPT on the Apps SDK and MCP:
+eleven keyless tools that generate, inspect, rig, animate, and place models in AR, rendered inline in
+the conversation. The same capability runs as a public custom GPT. Across all surfaces, 39,675
+generations have started since June 2026 and 33,941 finished as downloadable models. three.ws is an
+OpenAI Select Partner and published Spatial MCP, the CC0 response shape the connector uses, so any
+assistant can render 3D results.
+
+**Working link:** https://three.ws/openai
+
+**Screenshot:** [images/openai-page.png](images/openai-page.png), captured 2026-09-16. Alt text: "The
+three.ws OpenAI page headed '3D, natively, inside ChatGPT' with the OpenAI Select Partner badge and a
+'Connect it to ChatGPT' button." The form has no upload field; keep it for the follow-up if OpenAI's
+editorial team replies.
+
+**Speaker/founder bio:** the first answer above doubles as the bio.
+
+**Two proposed dates:** not applicable (no meeting requested).
+
+**Requested action:** consider the story for OpenAI's stories collection.
+
+**Approved relationship wording** (from [docs/press-kit.md](../../docs/press-kit.md)): write the status
+as "OpenAI Select Partner", never "partnered with OpenAI on" a product. Independence line: three.ws is
+an independent member of the OpenAI Partner Network at the Select tier: not an OpenAI product, and not
+endorsed by OpenAI beyond the partner designation shown.
diff --git a/marketing/partner-packets/quicknode-rpc-failover-field-note.md b/marketing/partner-packets/quicknode-rpc-failover-field-note.md
new file mode 100644
index 000000000..24fbd3557
--- /dev/null
+++ b/marketing/partner-packets/quicknode-rpc-failover-field-note.md
@@ -0,0 +1,79 @@
+# Field note: running a Solana RPC failover chain in production
+
+*three.ws, September 2026. One page, written for the Quicknode Startup Program team and for any
+builder running more than one RPC provider. Every sentence below is checked against
+[`api/_lib/solana/connection.js`](../../api/_lib/solana/connection.js),
+[`api/_lib/solana/rpc-fallback.js`](../../api/_lib/solana/rpc-fallback.js),
+[docs/solana.md](../../docs/solana.md), and the production service on 2026-09-16.*
+
+---
+
+three.ws gives AI agents Solana wallets, reads balances for them, and verifies agent-to-agent payments.
+All of that is JSON-RPC, and none of it may depend on a single endpoint. Server-side calls build one
+priority-ordered chain and rotate past any lane that is rate-limited or cooling down.
+
+**The chain, in order.** An explicit endpoint a call site passes; the operator's `SOLANA_RPC_URL`; a
+dedicated Quicknode endpoint when `QUICKNODE_RPC_URL` is set; other keyed providers; operator-supplied
+free fallbacks; a curated keyless public set ending with the most-throttled public cluster; and last,
+`SOLANA_RPC_LAST_RESORT_URLS`, the metered reserve that serves only when every free lane is down. In
+production today the Quicknode endpoint sits in that reserve slot on purpose.
+
+**Lesson 1: dedupe order is policy.** The chain removes duplicate URLs and keeps the first occurrence.
+Until 2026-07-28 production named the metered Quicknode endpoint as both `SOLANA_RPC_URL` and the
+reserve. It was therefore the primary, took all traffic, and hit its daily cap
+(`-32003 daily request limit reached`) while the free lanes sat idle. The fix was configuration: a
+reserve URL appears in the reserve list and nowhere else. A second guard stops a bare public-cluster
+default (roughly 35 call sites wrote one) from jumping ahead of every paid lane.
+
+**Lesson 2: size the bench to the failure.** `cooldownMsFor()` parks a lane for 6 hours on quota
+exhaustion, 30 minutes on a bad key or dead URL, 10 minutes on a plain 429, 2 minutes on a provider
+5xx, and 30 seconds on a network error. Park too long and you waste paid capacity; too short and every
+caller rediscovers the outage.
+
+**Lesson 3: a 403 is not always a bad key.** Some providers answer HTTP 403 "Request blocked" for one
+call shape, such as `getTokenAccountsByOwner` filtered by program, while serving everything else. We
+used to bench the whole lane for 30 minutes on it, so routine balance reads evicted a healthy primary.
+Now that call fails over alone and only that method is demoted on that endpoint, for 15 minutes.
+Deterministic errors such as invalid params never rotate, because every lane would fail them the same
+way.
+
+**Lesson 4: breakers must be fleet-wide.** Cooldowns are mirrored through a shared cache, so a new
+Cloud Run instance inherits "this lane is out of quota" instead of spending a request to learn it. A
+lane-health sensor now reports whether any paid lane is still serving. Before 2026-07-29 the only
+sensor watched one provider from per-instance memory, and it missed every paid lane being exhausted at
+once. When every lane fails, the last good answer for the same read is served with an
+`x-solana-rpc-stale` age header rather than an error.
+
+**What is still hard.** Failover moves load; it does not create capacity. At 16:01Z on 2026-09-16 the
+production health check reported all three paid lanes and all nine lanes cooling. A change committed
+2026-09-15, not yet deployed, makes the outer wrapper probe once through cooldowns when everything is
+parked, after warm instances returned fast 502s while fresh ones reached a healthy lane. The
+remaining lever is plan sizing, which is where a provider conversation helps more than code.
+
+*Code: [github.com/nirholas/three.ws](https://github.com/nirholas/three.ws), Apache-2.0. Docs:
+[three.ws/docs/solana](https://three.ws/docs/solana).*
+
+---
+
+## Source check for every claim
+
+| Claim | Source | Checked |
+|---|---|---|
+| Chain order, Quicknode slot after `SOLANA_RPC_URL`, reserve last | `solanaRpcEndpoints()` in `connection.js` | 2026-09-16 |
+| Quicknode endpoint is in the production reserve list; `QUICKNODE_RPC_URL` and `SOLANA_RPC_URL` unset | `node scripts/read-service-env.mjs` (host name only inspected) | 2026-09-16 |
+| The 2026-07-28 misconfiguration and `-32003` | [docs/solana.md](../../docs/solana.md) | 2026-09-16 |
+| About 35 call sites with a public-cluster default | docs/solana.md | 2026-09-16 |
+| Cooldown windows | Constants at `connection.js` lines 199 to 203; docs/solana.md table | 2026-09-16 |
+| Method demotion, 15 minutes | `METHOD_DEMOTION_MS` in `connection.js` | 2026-09-16 |
+| Deterministic errors do not rotate | docs/solana.md | 2026-09-16 |
+| Fleet-wide cooldowns, lane-health sensor, 2026-07-29 blind spot | `rpcLaneHealth()` comment in `connection.js`; docs/solana.md | 2026-09-16 |
+| Last-good serve with `x-solana-rpc-stale` | `serveLastGood()` in `connection.js` | 2026-09-16 |
+| All paid lanes and all nine lanes cooling | `GET https://three.ws/api/healthz`, `rpc_lanes` subsystem, `checkedAt` 2026-09-16T16:01Z | 2026-09-16 |
+| Recovery-probe change committed, not deployed | Commit `ac1ba86ad` (2026-09-15) is not an ancestor of live commit `58224bd69` from `https://three.ws/api/version` | 2026-09-16 |
+
+**Before sending:** re-run `curl -s https://three.ws/api/healthz` and `curl -s https://three.ws/api/version`.
+If the recovery-probe change has shipped, change "not yet deployed" to "deployed on [date]". If the lanes
+are healthy, keep the incident sentence but say it was a point-in-time reading.
+
+**Naming rule used here:** other RPC providers are described generically on purpose. Naming them in a
+note sent to one provider is unnecessary and would put other companies' outages in writing.
diff --git a/marketing/partner-packets/quicknode-spotlight-ask.md b/marketing/partner-packets/quicknode-spotlight-ask.md
new file mode 100644
index 000000000..e8e1ee5b7
--- /dev/null
+++ b/marketing/partner-packets/quicknode-spotlight-ask.md
@@ -0,0 +1,87 @@
+# Quicknode startup spotlight ask
+
+**Target:** Quicknode Startup Program
+**Program page:** https://www.quicknode.com/startup (the pipeline's `/startup-program` URL also returns
+200; the program text quoted below was read from `/startup` on 2026-09-16)
+**Verified intake:** the program contact from the 2026-07 acceptance. The repo records acceptance and
+approved credits ([docs/listings.md](../../docs/listings.md)) but **no contact name or email**. If none
+exists in the owner's inbox, the public routes are https://www.quicknode.com/contact-us and the
+Quicknode Discord.
+**What the program publicly offers that fits:** "Join Feature Fridays to showcase your solutions",
+"Present your vision at our events to an influential audience", and "partnership stages, from
+Marketplace to Developer Relations" (program page, 2026-09-16). There is no published spotlight
+application form.
+**Deadline:** none published. Pipeline target **2026-10-09**.
+**Owner's one step:** reply in the existing Quicknode Startup Program thread with the email below and
+attach the field note.
+
+---
+
+## The email
+
+**Subject:** Feature Friday idea: what running Quicknode as a Solana failover reserve taught us
+
+> Hi [program contact first name],
+>
+> three.ws joined the Quicknode Startup Program in July. We run a multi-provider Solana RPC failover
+> chain in production for AI agent wallets, balance reads, and payment verification, and Quicknode is
+> the metered reserve at the end of it.
+>
+> We wrote up what we learned in one page, attached: why dedupe order silently turned our reserve into
+> our primary until 2026-07-28, how long to bench a lane for each failure class, why a 403 is not always
+> a bad key, and why breakers have to be fleet-wide. It includes what is still hard, not just what
+> worked.
+>
+> Would the team consider it for a Feature Friday or a founder spotlight? We can present live, turn it
+> into a co-authored post for your blog, or both. Two dates that work for us: Friday 2026-10-16 or
+> Friday 2026-10-23. Code and docs are public: https://github.com/nirholas/three.ws and
+> https://three.ws/docs/solana.
+>
+> Thanks,
+> [Owner name]
+> three.ws
+
+Fill the two bracketed names from the real thread before sending; neither is recorded in the repo.
+
+---
+
+## Packet fields
+
+**One-sentence pitch:** a startup-program member with a candid production write-up of running Quicknode
+inside a multi-provider Solana failover chain.
+
+**100-word abstract:** three.ws runs AI agent wallets, balance reads, and payment verification on
+Solana, and routes every server-side RPC call through one priority-ordered failover chain with a
+Quicknode endpoint as the metered reserve. The field note covers four production lessons: dedupe order
+turned the reserve into the primary until July 28 and exhausted its daily cap; lanes are benched from
+30 seconds to 6 hours by failure class; a call-shape 403 now demotes one method instead of a whole lane;
+and cooldowns are shared across the Cloud Run fleet. It closes with what failover cannot fix: capacity,
+the conversation this spotlight would start.
+
+**Working link:** https://three.ws/docs/solana (200 on 2026-09-16)
+
+**Screenshot:** [images/docs-solana-failover.png](images/docs-solana-failover.png), captured 2026-09-16.
+Alt text: "three.ws Solana docs explaining never to name the same endpoint as both primary and reserve,
+followed by the cooldown table for failing RPC endpoints."
+
+**Founder bio:** not recorded in the repo; add two lines under the signature if Quicknode asks.
+
+**Two proposed dates:** Friday **2026-10-16** or Friday **2026-10-23** (proposals, nothing booked).
+
+**Requested action:** consider three.ws for a Feature Friday session or a founder spotlight, with the
+field note as the technical basis.
+
+**Approved relationship wording** (from [docs/partners.md](../../docs/partners.md) and
+[docs/listings.md](../../docs/listings.md)): three.ws is a **member of the Quicknode Startup Program**,
+accepted 2026-07, with approved infrastructure credits. Quicknode is **a rung in the Solana RPC
+failover chain, not the whole chain**. The `/partners` "Infrastructure" chip is a grouping three.ws
+chose, not a Quicknode tier. Quicknode styles its own name "Quicknode" on its site.
+
+## Metrics used
+
+| Metric | Value | Source | Captured |
+|---|---|---|---|
+| Program status | Accepted 2026-07, credits approved | docs/listings.md | repo |
+| Quicknode position in production chain | Reserve (`SOLANA_RPC_LAST_RESORT_URLS`) | Cloud Run service env, host name only | 2026-09-16 |
+| RPC lane health at capture | 3 of 3 paid lanes and 9 of 9 lanes cooling | `https://three.ws/api/healthz` | 2026-09-16T16:01Z |
+| RPC request volume through Quicknode | **Not measured here.** Quicknode's own dashboard is the source; the repo keeps no per-lane request count | n/a | n/a |
diff --git a/marketing/three-utility-season/capture.mjs b/marketing/three-utility-season/capture.mjs
new file mode 100644
index 000000000..a18ec129c
--- /dev/null
+++ b/marketing/three-utility-season/capture.mjs
@@ -0,0 +1,108 @@
+#!/usr/bin/env node
+/**
+ * Captures the media for the "What $THREE does this week" utility season from the
+ * live site. Re-run on posting day so the image matches the figures in the post.
+ *
+ * node marketing/three-utility-season/capture.mjs # every edition
+ * node marketing/three-utility-season/capture.mjs --only 01,09 # selected editions
+ * node marketing/three-utility-season/capture.mjs --base http://localhost:3000
+ *
+ * Each recipe names the live route, an optional element to frame, and the output
+ * file. Floating site chrome (the getting-started stack, the walking companion,
+ * tip popovers) is hidden so the product surface itself fills the frame; nothing
+ * on the surface is edited.
+ */
+import { chromium } from 'playwright';
+import path from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { mkdirSync } from 'node:fs';
+
+const HERE = path.dirname(fileURLToPath(import.meta.url));
+const OUT = path.join(HERE, 'images');
+const args = process.argv.slice(2);
+const argValue = (flag) => {
+ const i = args.indexOf(flag);
+ return i >= 0 ? args[i + 1] : null;
+};
+const BASE = argValue('--base') || 'https://three.ws';
+const ONLY = (argValue('--only') || '').split(',').filter(Boolean);
+
+const HIDE_CHROME = `
+ #tws-corner-stack, .tws-corner-item, .twx-i18n-fab, .tws-atlas-hint,
+ .walk-companion, .walk-trail-layer, .walk-c2w-fx, .pricing-avatar-canvas,
+ a[href="#main-content"], a[href="#main"], .skip-link { display: none !important; }
+`;
+
+// Wallet-free JSON read of the public tier ladder (what an agent sees).
+const TIER_API = '/api/three/tier';
+const GH = 'https://github.com/nirholas/three.ws/blob/main';
+
+const RECIPES = [
+ { id: '01', file: '01-access-tiers.png', route: '/three', size: [1600, 900] },
+ { id: '02', file: '02-holder-free-quota.png', route: '/forge', size: [1600, 900] },
+ { id: '03', file: '03-gated-embed-mcp-tool.png', route: `${GH}/api/_mcp/tools/embed.js#L227`, scrollText: "name: 'create_gated_embed'", offset: 260, size: [1600, 900] },
+ { id: '04', file: '04-forge-max-holder.png', route: '/forge-max', size: [1600, 900] },
+ { id: '05', file: '05-gated-embed-docs.png', route: '/docs/token-gated-3d-embeds', size: [1600, 900] },
+ { id: '06', file: '06-deploy-fee-waiver.png', route: `${GH}/packages/metaplex-agent-mcp/src/config.js#L87`, scrollText: 'the deploy fee and the holder waiver', offset: 40, size: [1600, 900] },
+ { id: '07', file: '07-holder-scene-unlock.png', route: '/embed/v1/gated.html?asset=avatar%3Abf2d5c4b-2536-4593-bf15-ec934a6c9f48', size: [1600, 900] },
+ { id: '08', file: '08-tier-api.png', route: TIER_API, prettyPrint: true, size: [1600, 900] },
+ { id: '09', file: '09-buyback-ledger.png', route: '/three-token', scrollText: 'in platform revenue earned so far', offset: 420, size: [1600, 900] },
+ { id: '10', file: '10-gameready-holder.png', route: '/three', scrollY: 470, size: [1600, 900] },
+ { id: '11', file: '11-built-in-public.png', route: `${GH}/api/_lib/three-access.js#L30`, scrollText: 'enforced ', offset: 160, size: [1600, 900] },
+ { id: '12', file: '12-quarter-recap.png', route: '/three', size: [1080, 1350] },
+];
+
+async function frame(page, recipe) {
+ if (recipe.prettyPrint) {
+ // Chromium's JSON viewer: tick its own Pretty-print toggle so the payload reads.
+ await page.getByLabel('Pretty-print').check().catch(() => page.locator('input[type=checkbox]').first().check());
+ await page.waitForTimeout(800);
+ }
+ if (recipe.scrollY) {
+ await page.evaluate((y) => window.scrollTo(0, y), recipe.scrollY);
+ await page.waitForTimeout(1200);
+ return page.screenshot();
+ }
+ if (recipe.scrollText) {
+ const loc = page.getByText(recipe.scrollText, { exact: false }).first();
+ await loc.scrollIntoViewIfNeeded({ timeout: 15000 });
+ await page.evaluate((dy) => window.scrollBy(0, -dy), recipe.offset ?? 140);
+ await page.waitForTimeout(1200);
+ return page.screenshot();
+ }
+ if (recipe.element) {
+ const el = page.locator(recipe.element).first();
+ await el.scrollIntoViewIfNeeded({ timeout: 15000 });
+ await page.evaluate(() => window.scrollBy(0, -200));
+ await page.waitForTimeout(1200);
+ return page.screenshot();
+ }
+ return page.screenshot();
+}
+
+mkdirSync(OUT, { recursive: true });
+const browser = await chromium.launch();
+let failed = 0;
+for (const recipe of RECIPES) {
+ if (ONLY.length && !ONLY.includes(recipe.id)) continue;
+ const [width, height] = recipe.size;
+ const ctx = await browser.newContext({ viewport: { width, height }, colorScheme: 'dark' });
+ const page = await ctx.newPage();
+ const url = recipe.route.startsWith('http') ? recipe.route : BASE + recipe.route;
+ try {
+ await page.goto(url, { waitUntil: 'domcontentloaded', timeout: 60000 });
+ await page.waitForLoadState('networkidle', { timeout: 20000 }).catch(() => {});
+ await page.addStyleTag({ content: HIDE_CHROME });
+ await page.waitForTimeout(recipe.settle ?? 6000);
+ const png = await frame(page, recipe);
+ const { writeFileSync } = await import('node:fs');
+ writeFileSync(path.join(OUT, recipe.file), png);
+ console.log(`captured ${recipe.id} ${url} -> images/${recipe.file} (${width}x${height}) at ${new Date().toISOString()}`);
+ } catch (err) {
+ failed += 1;
+ console.error(`FAILED ${recipe.id} ${url}: ${err.message.split('\n')[0]}`);
+ }
+ await ctx.close();
+}
+await browser.close();
+process.exit(failed ? 1 : 0);
diff --git a/marketing/three-utility-season/images/01-access-tiers.png b/marketing/three-utility-season/images/01-access-tiers.png
new file mode 100644
index 000000000..8bf4980bf
Binary files /dev/null and b/marketing/three-utility-season/images/01-access-tiers.png differ
diff --git a/marketing/three-utility-season/images/02-holder-free-quota.png b/marketing/three-utility-season/images/02-holder-free-quota.png
new file mode 100644
index 000000000..8c751baca
Binary files /dev/null and b/marketing/three-utility-season/images/02-holder-free-quota.png differ
diff --git a/marketing/three-utility-season/images/03-gated-embed-mcp-tool.png b/marketing/three-utility-season/images/03-gated-embed-mcp-tool.png
new file mode 100644
index 000000000..d97f19112
Binary files /dev/null and b/marketing/three-utility-season/images/03-gated-embed-mcp-tool.png differ
diff --git a/marketing/three-utility-season/images/04-forge-max-holder.png b/marketing/three-utility-season/images/04-forge-max-holder.png
new file mode 100644
index 000000000..189347d1f
Binary files /dev/null and b/marketing/three-utility-season/images/04-forge-max-holder.png differ
diff --git a/marketing/three-utility-season/images/05-gated-embed-docs.png b/marketing/three-utility-season/images/05-gated-embed-docs.png
new file mode 100644
index 000000000..4bed3be92
Binary files /dev/null and b/marketing/three-utility-season/images/05-gated-embed-docs.png differ
diff --git a/marketing/three-utility-season/images/06-deploy-fee-waiver.png b/marketing/three-utility-season/images/06-deploy-fee-waiver.png
new file mode 100644
index 000000000..899f3e57c
Binary files /dev/null and b/marketing/three-utility-season/images/06-deploy-fee-waiver.png differ
diff --git a/marketing/three-utility-season/images/07-holder-scene-unlock.png b/marketing/three-utility-season/images/07-holder-scene-unlock.png
new file mode 100644
index 000000000..54ca1ec43
Binary files /dev/null and b/marketing/three-utility-season/images/07-holder-scene-unlock.png differ
diff --git a/marketing/three-utility-season/images/08-tier-api.png b/marketing/three-utility-season/images/08-tier-api.png
new file mode 100644
index 000000000..db0cfa7d1
Binary files /dev/null and b/marketing/three-utility-season/images/08-tier-api.png differ
diff --git a/marketing/three-utility-season/images/09-buyback-ledger.png b/marketing/three-utility-season/images/09-buyback-ledger.png
new file mode 100644
index 000000000..712499ac8
Binary files /dev/null and b/marketing/three-utility-season/images/09-buyback-ledger.png differ
diff --git a/marketing/three-utility-season/images/10-gameready-holder.png b/marketing/three-utility-season/images/10-gameready-holder.png
new file mode 100644
index 000000000..613969615
Binary files /dev/null and b/marketing/three-utility-season/images/10-gameready-holder.png differ
diff --git a/marketing/three-utility-season/images/11-built-in-public.png b/marketing/three-utility-season/images/11-built-in-public.png
new file mode 100644
index 000000000..98ce2af60
Binary files /dev/null and b/marketing/three-utility-season/images/11-built-in-public.png differ
diff --git a/marketing/three-utility-season/images/12-quarter-recap.png b/marketing/three-utility-season/images/12-quarter-recap.png
new file mode 100644
index 000000000..7b3f326c5
Binary files /dev/null and b/marketing/three-utility-season/images/12-quarter-recap.png differ