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.
Announcement pack: the $THREE token MCP server
Surface: @three-ws/three-token-mcp ·
Ledger key: @three-ws/three-token-mcp · Stage: drafted · Published to npm: 1.1.3 is latest
· Announced externally: never
Listed as never announced in the coverage audit, which flags it as the $THREE-native post that "writes itself". Token topic is the account's highest measured lift (13.3x), and this is a feature that genuinely touches the token rather than mentioning it. Written against the announcement voice.
The claim, and where it is checked
An MCP server that lets an AI assistant burn $THREE on Solana. Its burn tool is declared irreversible to the client, and by default the server refuses to sign until the call carries
confirm: true.
| Part of the claim | Where it is real |
|---|---|
| It is a published MCP server anyone can run | npm view @three-ws/three-token-mcp returns 1.1.3 as latest (4 versions, Apache-2.0, maintainer three-ws). Running npx -y @three-ws/[email protected] over stdio on 2026-09-17 answered initialize as three-token-mcp 1.1.3 and tools/list with exactly three_price, three_balance, three_burn. |
| It burns $THREE on Solana | packages/three-token-mcp/src/lib/token.js burnThree() builds one Solana mainnet transaction (SPL transfer per leg plus a memo), signs, sends, and confirms. The live catalog detail reads "signs and broadcasts on Solana mainnet". |
| It can only ever burn $THREE | assertCanonicalThree() in the same file refuses to sign unless the mint served by /api/token/config equals THREE_MINT in src/config.js (FeMbDoX7R1Psc4GEcvJdsbNbZA3bfztcyDCatJVJpump). Not in the post, but it is why the post is safe to make. |
| Flagged irreversible | annotations: { readOnlyHint: false, destructiveHint: true, ... } in src/tools/three-burn.js. The public MCP Tool Catalog classifies it as Irreversible ("Cannot be undone once it runs.") and the two reads as Read-only; that page is generated from tool source, so the label cannot drift. |
By default it will not sign without confirm: true |
The handler in three-burn.js returns confirmation_required before burnThree() is ever called when REQUIRE_CONFIRM && confirm !== true, and REQUIRE_CONFIRM in config.js is true unless the operator sets it to 0/false. Confirmed live: a three_burn { usd: 1 } call against the published 1.1.3 package came back "error": "confirmation_required" with nothing signed. |
| Where the burned tokens go | The burn leg routes to the Solana incinerator (1nc1nerator11111111111111111111111111111111). computeSplit() sends the rest to the treasury only when the live config has one; /api/token/config currently serves "treasury": null, so today a burn sends 100% to the incinerator. This is why the post says "burn" and does not promise a treasury split. |
Full mechanism: packages/three-token-mcp/README.md;
install lines for every three.ws server: docs/mcp.md.
Media
Captured from the live route by npm run announce:media against the recipe below. Provenance
(route, commit, time, sha256) lands in
public/announce/media-manifest.json.
| Shot | File | Notes |
|---|---|---|
three-token-mcp-hero |
/announce/img/three-token-mcp-hero.webp |
1800x1013, signed out. The MCP Tool Catalog filtered to this server: the "Let your assistant run the safe ones by itself" policy with its Strict, Balanced and Open profiles, then the server's three tools with three_burn badged Irreversible and the two reads badged Read-only. |
Alt text, required on the post:
The three.ws MCP Tool Catalog filtered to the @three-ws/three-token-mcp server: 3 of 321 tools, with three_balance and three_price labelled read-only and three_burn labelled irreversible, all free
What the recipe frames out, and why. The catalog's tool subtitles and descriptions name the
third-party price source the tools quote from, which would put another crypto project into a
committed frame; they are hidden (.mc-title, .mc-desc), leaving each tool's name and its safety
and price badges, which are the subject. The policy snippet block (.mc-policy-out) is hidden so the
profiles and the three tools fit one 16:9 frame. Nothing is restaged: every badge, count and label is
what the live page rendered.
The two existing three-token-hero and three-token-trades shots already carry this package as
their surface, and were checked before this one was captured. They were not reused: they frame
/three-token, show no part of the MCP server, were captured 2026-09-04 with a price that no longer
matches the live page, and carry third-party venue buttons in the frame.
The post
Pattern: mechanism lead. 179 weighted characters by post-tweet.mjs (which counts the bare
link at its real 41 characters; X itself counts it as 23, so it lands near 161 on the platform),
inside the 100 to 179 band. Tags @solana because the burn is a Solana mainnet transaction and
$THREE is an SPL token.
Postable file: three-token-mcp.post.txt.
An MCP server that lets an AI burn $THREE on @solana. three_burn is flagged irreversible and by default won't sign without confirm: true. three.ws/mcp-tools?server=three-token-mcp
Ship it with three-token-mcp-hero attached and the alt text above.
Why it is written that way
A token burn tool invites two readings: a gimmick, or a foot-gun that lets a model spend your wallet. The second is the objection a developer actually raises, so the post answers it inside the mechanism instead of hyping the first. "Flagged irreversible" is the MCP annotation clients use to decide whether to prompt; "won't sign without confirm: true" is the server-side gate that holds even against a client that ignores the annotation. Both are checkable in the linked page.
@AnthropicAI was considered (MCP is their protocol, and the voice guide allows the tag for an
MCP server) and left off: the post already carries the tag the mechanism depends on, and a second
tag spends characters the band does not have. The npx line is not in the post on purpose: X
would parse @three-ws in @three-ws/three-token-mcp as a mention of an account we do not own.
The linked catalog entry carries the exact install command.
Cut to hold the band, all true and all in the Telegram variant: the canonical-mint assertion, the default $100 per-burn cap, and the two read-only tools.
Telegram variant
The $THREE token now has its own MCP server, and it has never been posted about.
npx -y @three-ws/three-token-mcp
Any MCP client (Claude Code, Claude Desktop, Cursor) gets three tools:
- three_price: the live USD price of $THREE, and how much $THREE a USD amount buys. Read-only.
- three_balance: the $THREE and SOL balance of any Solana wallet. Read-only.
- three_burn: burns a USD amount of $THREE your wallet already holds, in one Solana transaction.
The burn is built to be hard to misuse:
- It is flagged irreversible, so clients that honour MCP annotations ask you before it runs.
- The server itself refuses until the call carries confirm: true, whatever the client does.
- Each burn is capped at $100 by default (MAX_BURN_USD).
- Before signing, it checks the token is the canonical $THREE mint, so a bad config can never
point it at another token.
See all three tools and their safety labels: three.ws/mcp-tools?server=three-token-mcp
Changelog
No new entry. The package shipped with the 2026-06-13 entry ("All 14 three.ws MCP servers are now live on npm and the official MCP registry"). This pack announces an existing package externally for the first time; it ships nothing new.
Posting
Publishing to an external channel is owner-gated under the operating rules, every time. Nothing in
this pipeline posts. The queue item is three-token-mcp (status review, lane token, pattern
mechanism); preview with:
node scripts/post-tweet.mjs --file docs/announcements/three-token-mcp.post.txt --dry-run
npm run x:content -- run --dry-run --id three-token-mcp