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.
Tripo x three.ws: partnership brief
Drafted 2026-10-09, after Tripo reached out. A working brief for the first call and the reply, not a status report. Every number about three.ws was read from production on the drafting date; every number about Tripo is labelled with where it came from, and the ones Tripo publishes about itself are marked as such.
TL;DR. Tripo is the only 3D generation company whose work already runs in three separate places inside three.ws: a paid API lane, a self-hosted open-weight worker, and a hosted failover rung. The open-weight lane is busy and the paid lane is almost unused, for one reason: the paid lane needs the user's own Tripo key. A partner key closes that gap, puts Tripo in front of every Forge user instead of the few who bring a key, and gives Tripo a production surface for the endpoints we have not wired yet (rigging, retargeting, low-poly, splats, segmentation).
Who Tripo is
| Fact | Value | Source |
|---|---|---|
| Company | VAST, the team behind Tripo, TripoSR and TripoSG | tripo3d.ai |
| Latest round | Series B/B+, about RMB 3B (about USD 446M), 2026-09-01, led by Matrix Partners China with game publishers as strategic investors | Press (36kr, DealStreetAsia), not a filing |
| Earlier 2026 rounds | USD 50M Series A (March), about USD 200M A+/A++ (June), USD 150M A3 (July) | Press |
| Users | 6.5M creators, about 90K developers, about 100M models | Tripo's own figures, not independently verified |
| API | v3 at openapi.tripo3d.com/v3, with a published v2-to-v3 migration guide |
developers.tripo3d.com |
| API pricing | 1 credit = USD 0.01. Text to 3D 10 (untextured) or 20 (textured); image or multiview 20 or 30; HD texture +10; quad +5; auto-rig 25; retarget 10 per clip; rig check free | developers.tripo3d.com, checked 2026-10-09 |
| Enterprise route | "custom volumes, dedicated support, or SLA guarantees" via [email protected] | Tripo's official pricing page |
| Affiliate program | 30% recurring for 12 months on Tripo's own terms page; a third-party affiliate network lists different terms | Confirm on the call |
Their open source, which we already depend on
Star counts from the GitHub API on 2026-10-09:
| Repo | Stars | License | Used by three.ws |
|---|---|---|---|
| TripoSR | 7,035 | MIT | Yes: hosted failover rung (Hugging Face Space, Replicate) |
| TripoSG | 1,825 | MIT | Yes: self-hosted worker workers/model-triposg |
| UniRig | 1,804 | MIT | Planned: non-humanoid rigging, order 069-best3d-03 |
| TripoSplat | 1,365 | MIT | No |
| TripoSF, HoloPart, AniGen, SkinTokens, MIDI-3D | 463 to 952 each | check each LICENSE before use | AniGen and SkinTokens are planned in 069-best3d-03 |
What three.ws already runs on Tripo
Verified in this repository and in production on 2026-10-09.
| Surface | Where | What it does |
|---|---|---|
| Tripo v3.1 API lane, bring-your-own-key or house key | api/_providers/tripo.js, tier entry in api/_lib/forge-tiers.js |
Text and image to 3D on model v3.1-20260211, on the v3 endpoint openapi.tripo3d.com/v3. Runs on a pasted user key, or on TRIPO_API_KEY when the platform sets one (a High-tier default and an explicit "Tripo" engine choice, billed to $THREE credits); the BYOK key link in the Forge points at platform.tripo3d.ai/api-keys |
| Signed webhook receiver | api/forge/webhook/tripo.js |
Verifies Tripo-Webhook-Signature, claims Tripo-Webhook-Delivery for idempotency, and completes a job the same way polling does; registering the endpoint URL in Tripo's dashboard and setting TRIPO_WEBHOOK_SECRET are the owner's last step |
| TripoSG, self-hosted | workers/model-triposg |
The sketch-to-3D lane, and the last automatic failover rung for photo jobs (forge failover) |
| TripoSR, self-hosted and hosted | workers/model-triposr (the fast lane in the avatar pipeline controller), plus a rung in the free Hugging Face Spaces chain (api/_providers/huggingface.js) |
The fast path, at about 3 seconds of GPU time per mesh, and part of the free photo-to-3D failover chain |
The number that shapes the pitch
Production, last 30 days, internal jobs excluded (forge_creations, grouped by backend):
| Lane | Jobs | Finished |
|---|---|---|
| TripoSG, self-hosted (open weights) | 197 | 124 |
| Tripo API, user's own key | 2 | 2 |
Tripo-family jobs have grown every month since July (32, 81, 101, then 114 in the first nine days of October), almost all of it on the open-weight worker. The paid lane is nearly idle because almost nobody arrives at a 3D tool already holding a Tripo key. That is the gap a partnership closes: Tripo's best model is one setting away from a Forge that took 12,373 generation and rigging requests from real makers in September (the audited figure in the proof brief, which excludes catalog seeders and benchmarks), and the setting is currently "go and sign up somewhere else first".
For the wider evidence (completion rate, makers, MCP traffic), use the 30-day proof brief and quote its numbers the way it says to.
The proposal
What we ask Tripo for
- A partner-priced platform key, so Tripo becomes a house lane in the Forge rather than a bring-your-own-key option. The natural placement is the High tier, where Tripo competes on quality, and an explicit "Tripo" engine choice for users who want it by name.
- Raised concurrency above the published defaults (H-series 10, P-series 5, rig 10), sized to the house-lane volume we agree.
- Early access to new H-series and P-series models and to the splat endpoint, so the Forge can ship them on Tripo's launch day.
- Co-marketing: a joint launch post when the house lane goes live, and a listing on Tripo's integrations or plugins page.
- Written confirmation that we may keep self-hosting TripoSG and TripoSR in production with attribution, and say so publicly. Both are MIT, so this is courtesy, not licensing, but it makes the joint story clean.
- The affiliate or referral terms that apply to the BYOK key link we already render.
What we offer Tripo
- Migration to the v3 API with signed webhooks, replacing polling on the legacy v2 endpoint.
- The endpoints we do not use yet, wired as Forge features: auto-rig and retarget on the avatar path (they slot into our universal retargeter, which already drives any humanoid rig), P1/P2 low-poly and quad as a game-ready preset, the image-to-splat endpoint on the splat stage, and Smart Segmentation on the parts editor.
- Agent distribution: Tripo exposed through our MCP servers (13,898 tool calls in
September at 99.4% handler success, from 13 distinct MCP clients including ChatGPT,
Claude and Cursor), through the
<agent-3d>embed with engine attribution, and through our paid per-call agent lane. - Benchmarks: anonymised per-engine completion, latency and quality-gate verdicts from our failover grid, where Tripo runs side by side with TRELLIS and Hunyuan3D on the same prompts. No other customer can give them that comparison from live traffic.
- Upstream work on the open repos we run: fixes to TripoSG, UniRig and ComfyUI-Tripo as we hit them in production.
Deal shapes, from smallest to largest
| Shape | What changes | Who pays |
|---|---|---|
| A. Credits pilot | Tripo grants credits for a 60-day house-lane pilot; we ship v3 plus webhooks and report the numbers | Tripo (credits) |
| B. Partner pricing | A discounted per-credit rate on a committed monthly volume; Tripo becomes a permanent High-tier lane | three.ws, at a partner rate |
| C. Revenue share | Tripo jobs sold through our paid lanes carry a margin both sides agree; Tripo lists three.ws as an integration | Shared |
Start the conversation at A. It costs Tripo little, proves volume with real data, and every engineering item we ship for it is kept whichever way the deal goes.
Questions for the first call
- Who on their side owns platform partnerships, and is there a partner tier separate from enterprise and affiliate?
- Are credits, a partner rate, or a revenue share the shape they want?
- Concurrency ceiling for a partner key, and whether webhooks count against it.
- Their roadmap for the splat and P-series endpoints, and whether early access is possible.
- Attribution rules: how they want Tripo named in the Forge engine picker and in exports.
- Whether their own users can generate in Tripo and land straight in a three.ws agent (a "send to three.ws" export), which is the reciprocal integration.
Engineering, shipped
The integration work closed as work order 084; see the Tripo house lane doc for the live, user-facing description of what shipped:
- v3 migration: done. The provider speaks the v3 REST API end to end (dedicated
endpoints per capability, the unified
input/task-id model, the unified error envelope). - House key: done.
TRIPO_API_KEYmakes Tripo a High-tier default and an explicit engine choice, billed to $THREE credits like the other house lanes; with the variable unset, behavior is unchanged. It does not exist on the production service yet (checked 2026-10-09), so flipping it on is the owner's decision: it onboards a paid API. - Signed webhooks: done. Polling stays the fallback.
- Auto-rig: done.
rig({spec:'mixamo'})output canonicalizes throughglb-canonicalize.js's existing Mixamo mapping and drives the universal retargeter, the same as any other humanoid rig. - Low-poly/quad game-ready preset and Smart Segmentation: the provider methods exist
(
decimate(),segment()inapi/_providers/tripo.js) but are not dispatched from any endpoint. Both capabilities already run on self-hosted GCP workers in production (workers/remeshfor game-ready, with FBX export and PBR re-bake Tripo's endpoint does not do;workers/segment+ PartCrafter/HoloPart for parts, order070-best3d-04), which the owner's standing GCP-first directive prefers over an equivalent paid third-party lane. Wiring Tripo here would add a strictly narrower alternative, not new capability; revisit if the partnership itself makes Tripo's version the one worth shipping. - Splat: no live Tripo endpoint exists to wire (checked again 2026-10-12). The splat
stage (
docs/splat.md) has no Tripo-backed path; this is the "early access" item in the proposal above, not something we can build today.
Refresh, 2026-10-23
What was re-checked on this date. Anything not listed here is unchanged from the 2026-10-09 draft above.
| Item | 2026-10-23 reading | How |
|---|---|---|
| Tripo API pricing | Unchanged: text to 3D 10 to 20 credits, image to 3D 20 to 30, auto-rig 25, retarget 10 per animation, 1 credit = USD 0.01 | developers.tripo3d.ai fetched today |
| API surface | v3 endpoint (/v3/generation/text-to-model), model v3.1-20260211. Listed models: H3.1, P2.0 (native quad), AI Texture, Segmentation V2 |
Same page. It has no changelog or dated entries, so nothing newer than the model list can be reported |
| Announcements since 2026-10-09 | None found. A web search surfaced no October 2026 Tripo announcement; the latest items were July 2026 funding coverage and the March 2026 H3.1 and P1.0 launch. Third-party sources disagree on some details, so those press items stay labelled as press | Web search, not an official Tripo source |
| House key on the service | Still absent. node scripts/read-service-env.mjs '^TRIPO' --names finds no TRIPO* variable on three-ws-api |
Cloud Run service env |
| Order 084 | Closed (no 084 file in prompts/finish) |
ls prompts/finish |
Tripo-family jobs, forge_creations, internal jobs excluded
Read with a read-only SELECT against production on 2026-10-23.
Last 30 days, by backend and status:
| Lane | Finished | Failed |
|---|---|---|
| TripoSG, self-hosted (open weights) | 134 | 92 |
| Tripo API, user's own key | 2 | 0 |
Per calendar month, all Tripo-family backends:
| Month | Jobs |
|---|---|
| 2026-07 | 32 |
| 2026-08 | 81 |
| 2026-09 | 99 |
| 2026-10 (through the 23rd) | 190 |
July and August match the 2026-10-09 table. September reads 99 here against 101 earlier; the
cause was not investigated, so treat the earlier figure as the one that went into the draft.
This query excludes internal rows but not the forge_seed_jobs catalog seeders, so for the
wider usage claim use the audited 12,373 September figure in the
proof brief. The conclusion is unchanged: the open-weight lane
grows each month and the paid lane is nearly unused (2 jobs).
Reply draft
Send from the owner's account to the person who reached out, copying
[email protected] if the first contact did not come from there.
Subject: three.ws x Tripo: a house lane for Tripo in the Forge
Hi, thanks for reaching out. Good timing: Tripo already runs in three places inside three.ws. Your v3.1 model is a lane in our Forge, TripoSG powers our sketch-to-3D tool and sits in our photo failover chain, and TripoSR is a rung in the free tier.
The catch is that the v3.1 lane only works for users who bring their own Tripo key, so almost nobody uses it, while the Forge took over 12,000 generation and rigging requests from real makers last month. We would like to change that: make Tripo a built-in lane on our top quality tier, move to your v3 API with webhooks, and add your rigging, low-poly, splat and segmentation endpoints as Forge features. Those would also reach the agents that use three.ws through MCP (about 14,000 tool calls last month, from ChatGPT, Claude, Cursor and others).
A simple first step would be a 60-day pilot on partner credits, with us reporting completion, latency and quality numbers side by side with the other engines we run.
Are you free for a call next week? Happy to show the Forge live.
Best, [owner name], three.ws
Related
- Partner prospects: the full list of companies worth approaching, of which Tripo is the first.
- Partnership outreach plan: how the list is worked, wave by wave.
- Forge and How the Forge works: the engine grid Tripo plugs into.
- 3D landscape, October 2026: the open-source sweep that put UniRig and the rigging stack on the roadmap.