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(onchain-agent-wallets): draft the launch post copy

nirholas590ece34d268
diff --git a/docs/onchain-agent-wallets-x-post.md b/docs/onchain-agent-wallets-x-post.md
new file mode 100644
index 000000000..ea44d0d42
--- /dev/null
+++ b/docs/onchain-agent-wallets-x-post.md
@@ -0,0 +1,175 @@
+# X post: onchain-agent-wallets
+
+Announcement copy for [github.com/nirholas/onchain-agent-wallets](https://github.com/nirholas/onchain-agent-wallets),
+the MCP server that gives an AI agent a Solana spending allowance instead of a private key.
+
+**Thesis of this announcement:** handing an agent a keypair is the wrong primitive, and the chain
+already has the right one. Nothing else. No token talk, no roadmap, no "the future of agentic
+commerce."
+
+**Every claim below is true today**, verified against the shipped 0.1.0 code: the vault token
+account is owned by the human and derived deterministically from their address plus the agent id;
+the agent holds only an SPL Token delegation capped by `approve`; the token program decrements
+`delegated_amount` on every spend and refuses anything above it; one owner-signed `revoke` sets it
+to zero; the owner can withdraw without the agent's cooperation; 14 MCP tools; 22 tests; the
+delegation test runs the real SPL Token program in process and asserts that spending 61 against a
+remaining allowance of 60 fails on-chain while 460 sits in the vault; guardrails (per-transaction
+cap, rolling 24h cap, recipient allowlist, x402 host allowlist, expiry, pause, confirm-above
+threshold) are checked before signing; `pay_x402` reads the price from the unpaid 402 before any
+money moves; owner actions return an unsigned transaction for Phantom, Solflare, Backpack, or a
+Ledger, so no secret key is needed on the machine.
+
+**Things to NOT claim yet:**
+
+1. **Do not put an `npx` or `npm install` line in the post.** `@three-ws/onchain-agent-wallets` is
+ not on npm yet: publishing is blocked on npm's 2FA policy for the current token. Every draft
+ below links the repo instead, on purpose. The moment the package is live, the builder-facing
+ alternate below becomes the better post.
+2. **Do not say it is in the MCP registry.** That entry is gated on the npm publish above.
+3. **Do not call it audited.** It is tested against the real token program, which is a different and
+ weaker claim. "The token program enforces it" is true and strong enough.
+
+---
+
+## 1. Main post (recommended)
+
+267 characters with X's 23-character URL count, so it posts from any account, Premium or not.
+
+> Everyone hands their AI agent a private key.
+> Then keeps $20 in the wallet, because that's the blast radius.
+>
+> Give it an allowance instead.
+>
+> Solana's token program enforces the cap, decrements it every spend, and you revoke in one
+> instruction.
+>
+> github.com/nirholas/onchain-agent-wallets
+
+Why this one: the first two lines name a thing every builder has actually done and quietly felt bad
+about, which is what earns the repost. "Blast radius" is the phrase doing the work. The fix arrives
+as one short line, and the mechanism follows so nobody has to take it on faith. It never says "MCP",
+because the people who need MCP will click, and the people who do not would have bounced on the
+acronym.
+
+## 2. Alternate: the proof
+
+For a technical timeline. Leads with the test rather than the pitch. 259 characters.
+
+> We wrote a test that gives an agent a 100 USDC allowance, spends 40, then tries to spend 61 with
+> 460 USDC still sitting in the vault.
+>
+> It fails on-chain. Not in our code. In Solana's token program.
+>
+> That refusal is the entire product.
+>
+> github.com/nirholas/onchain-agent-wallets
+
+Why this one: engineers trust a failing transaction more than any amount of copy. It also
+pre-empts the first reply you will get, which is "so what stops the agent ignoring your limits."
+
+## 3. Alternate: x402 / agent-payments timeline
+
+269 characters.
+
+> Your agent can pay for its own APIs now, without you handing it a wallet.
+>
+> It gets an allowance. It pays the x402 invoice out of that allowance. The price is read before any
+> money moves, and the chain caps the total.
+>
+> Revoke in one instruction.
+>
+> github.com/nirholas/onchain-agent-wallets
+
+## 4. Alternate: builder-facing (HOLD until npm is live)
+
+Do not post this one until `@three-ws/onchain-agent-wallets` is published. Then it is probably the
+best of the four.
+
+> ```
+> npx -y @three-ws/onchain-agent-wallets
+> ```
+>
+> 14 MCP tools that give an agent a Solana wallet it cannot drain: an on-chain spending cap, per-tx
+> and daily limits, recipient and host allowlists, an audit log of every refusal, and x402.
+>
+> You keep custody. Revoke any time.
+>
+> github.com/nirholas/onchain-agent-wallets
+
+---
+
+## Thread version
+
+Post 1 is the main post above. Then:
+
+**2/**
+
+> The problem with agent wallets: a keypair is all-or-nothing. Whoever holds it can move everything
+> in the account. There is no ceiling, no allowlist, no expiry, and no way to take it back without
+> racing the agent.
+
+**3/**
+
+> So the money never goes to the agent.
+>
+> It sits in a vault account you own. The agent gets an SPL Token delegation over it, capped. Every
+> spend decrements the cap. Spend past it and the token program rejects the transaction.
+
+**4/**
+
+> On top of that: per-transaction cap, rolling 24h cap, recipient allowlist, x402 host allowlist,
+> expiry, pause, and a threshold above which a human has to say yes.
+>
+> Refusals are logged with the exact rule that fired, so you can see what your agent tried to do.
+
+**5/**
+
+> If the machine running the agent is fully compromised, the local rules can be bypassed and the
+> on-chain cap still cannot.
+>
+> That is why they are two layers and not one.
+
+**6/**
+
+> No secret key has to touch the machine either. Owner actions come back as an unsigned transaction
+> for Phantom, Solflare, Backpack, or a Ledger to sign.
+>
+> Free rehearsal on devnet.
+>
+> github.com/nirholas/onchain-agent-wallets
+
+---
+
+## Replies worth pre-writing
+
+**"Why not a multisig / smart wallet?"**
+
+> Because this needs no new program and no new trust assumption. It is `approve` and `revoke` from
+> the SPL Token program, which has been on mainnet for years and is already audited harder than
+> anything we would ship. A multisig is a fine answer when you want co-signing. This is the answer
+> when you want a ceiling.
+
+**"What stops the agent from just asking you to raise the limit?"**
+
+> Nothing, and that is correct. It should be able to ask. Raising the ceiling is an owner-signed
+> `approve`, so a human is in the loop for exactly the decision that deserves one.
+
+**"Solana only?"**
+
+> Today, yes. The custody model is a token-program delegation, so the EVM equivalent is ERC-20
+> `approve` with the same shape. Solana first because that is where the agent economy we are
+> building actually lives.
+
+---
+
+## Notes on framing
+
+- **Do not lead with MCP.** It narrows the audience to people who already know the acronym, and the
+ idea is bigger than the transport.
+- **Do not mention $THREE in this post.** There is no fee in this package and no buyback lane, so
+ attaching the coin to it would be the kind of claim someone checks and finds hollow. The
+ metaplex-agent-mcp post is where the $THREE story is true and load-bearing.
+- **"Allowance" is the word.** It is the one term that makes a non-crypto reader understand the
+ whole design instantly. Keep it in every rewrite.
+- The Pages overview at nirholas.github.io/onchain-agent-wallets is a good second link for a reply,
+ not for the main post: one link per post converts better.