Status: what is built, what is not
In short
- What is built and tested, and what is not.
- Payments and deposits were rehearsed on local copies of real chains. Nothing is live, and no real money has moved.
- Contracts are not deployed.
Checked against the code on 16 November 2026. This page is the honest list. If a page elsewhere sounds better than this one, this one is right.
In one line
Tools, the first product, is built in test mode, and the parts that pay real tools and take real deposits were rehearsed on local copies ("forks") of real chains. Nothing is live. No real money has moved, no real tool has been paid, nothing is deployed.
Four words are used carefully on this page:
- Fixture: tested against test stand-ins on this machine.
- Rehearsed on a fork: run against a local copy of a real chain (Foundry's
anvil) that reads the real token contracts but keeps every write in memory. Not the real chain. - Measured live: read from the real internet, with no money.
- Advertised: what someone else's page says.
Order of the products
- Tools (now). 2. Cheaper inference. 3. Memory. 4. Computer (last, not built). Each one is made to work well before the next one starts.
Built and rehearsed on a fork
| What | How it was run |
|---|---|
| x402 payer: EIP-3009 USDC on Base, per-call ceiling, daily cap counted in the database | Base fork, real USDC contract, the official @x402/evm facilitator settling. tests/integration/live-rehearsal.test.ts |
MPP payer: tempo charge, pull mode, USDC.e | Tempo mainnet fork, real USDC.e, against a local MPP server. Not tried against a real MPP server. tests/integration/mpp-rehearsal.test.ts |
| Money in: "Pay with USDG or USDC", one address per link, credited once after confirmations | Robinhood Chain fork (real USDG) and Base fork (real USDC). tests/integration/deposit-rehearsal.test.ts |
| Treasury sweep: deposit addresses hand their USDG to the treasury by signed authorization | Robinhood Chain fork. tests/integration/treasury-rehearsal.test.ts |
| Reconciliation of every x402 payment against the chain | Base fork. npm run live:reconcile |
| Live start of the built server with test payment refused | Local forks. npm run prod:local -- --live |
Measured live (no money)
- The free check, 6 November 2026: of 27,755 listings, 25,351 answered an unpaid request with a valid payment ask. It checked the payee, not the price. 0 tools are proven by a paid call.
- One bridge quote, 10 November 2026: Across's public API quoted 10 USDG on Robinhood Chain to 9.990341 USDC on Base (fee $0.0097). Read only; nothing was sent.
Built and tested (fixture only)
| What | Where it is described |
|---|---|
| The agent's journey: self sign-up, search, call with a maximum, one link with a 6-digit code, the person's phone screens | agent journey |
Ways in: HTTP API, pet command line, MCP over stdio and hosted at /mcp, TypeScript SDK (not published) | cli, MCP, sdk |
| Tool search over all 27,755 listings, labels proven / verified / listed | tools |
| MPP checks and payment path; applications for builders; customer's own provider accounts | MPP, builders |
| Quote, hold, settle; a charge never above the hold | billing |
| Caps per account, project, key, agent, parent agent and session | billing |
| Loop pause before a repeated tool call is paid for again | billing |
| Crash clean-up (the sweeper) and invoice true-ups that only go down | billing |
| Receipts per call | billing |
| Cards, durable calls, approvals and result checks | cards |
| chain.watch | tools |
| Earn-to-Run listings, seller listings with review | agents, selling |
| Model gateway: Anthropic Messages, OpenAI Chat Completions and Responses; failover before the first word | how it works, billing |
| Repo memory and saved memories | memory |
| Agent wallets with owner approval above a threshold (fixture chain) | agent wallet |
| Your own AWS (Bedrock) against an emulator; resellers as upstreams | own cloud, resellers |
| Plans, volume bonus, referrals (API only, example numbers) | limits |
| Coin contracts, local chain only | contracts |
The live switch
PETROVA_LIVE_TOOLS is off unless set. The server refuses to start with it on unless the readiness report (fork rehearsals included) passed on that exact commit in the last 24 hours, the founder signed the approval, and the payer key, caps and an https address are set. Hard limits in code: $1 a call, $50 a day. Starting settings: $0.10 a call, $5 a day (.env.production.example). LIVE_SPEND for the model gateway stays refused in this build.
Not live yet: what is left
The founder's four steps: a server and a domain; the company wallets; a few dollars (about $5 of USDC on Base, optionally $2 of USDC.e on Tempo, $1 of gas on Robinhood Chain); and watching the first ~20 real calls. Docker, Caddy and a real domain have not been run.
Open models (the hub)
Built and tested on one machine: the pet model commands (download, publish, hold, serve, copy, share, relay, witness, catalog, have), signed versions, a log that cannot be rewritten unseen, and downloads from many holders with every piece checked. No index, relay or witness runs on the internet, the root key has not been made, and no holder exists that we do not run. See the hub page.
Not built
- Card payments (Stripe). In test mode the pay button is a labelled test payment.
- MPP push mode and MPP sessions, so the OpenAI and Anthropic MPP endpoints cannot be paid.
- Paying tools in USDG directly. The payer signs USDC on Base only.
- Moving money between chains is a manual step: the sweep is a command, the Across bridge and the Tempo route are done by hand.
- The own-wallet route (
/v1/x402/...) is off in live mode: its facilitator is a fixture. - OAuth sign-in for the hosted MCP, and an installer that sets up every coding agent.
- A Python SDK. The TypeScript SDK is built and not published to npm.
- Computer. Not started.
- Live model calls. No model supplier is connected; every model answer is a fixture.
- No key API for people. An agent signs itself up; a person has no sign-up page or
/v1/keys. - No list routes for caps, holds, usage, repos, listings, payouts or agent wallets.
- Coin not launched; contracts not deployed or audited. Nothing reads
unitsOf. - Rate limits only on sign-up (5 an hour per address, 200 an hour in all). Search, inspect and quote are free and not limited per key.
- No stream resume. A dropped model stream cannot be picked up again.
- Sample data only: the proof page and the coin transparency page.
Known gaps
packages/contracts/openapi.jsonhas no/v1/messagesor/v1/messages/count_tokens. Both routes work; they are missing from the description only.- No local command can make a key with
memory:read,memory:write,seller:write,payout:write,plan:write,cloud:writeorreferral:write. - Validation errors on the older routes say only "Invalid request"; the agent routes name the field.
GET /v1/modelshas no prices and is not in OpenAI's list shape.- Loop pause has no override for the customer; only the operator can turn it off, for everyone.
- The payer checks a tool's host resolves to public addresses, but does not pin that address for the connection. The per-call ceiling and daily cap bound the risk.
- A first MPP challenge pins the payee for a listing (trust on first use).
Numbers that are examples, not decisions
Plan prices, volume bonus tiers, referral share, the coin's tokens per unit and daily allowance per unit, and the planned sale price through resellers (35-40% of list) are all placeholders.
Last checked against the code: 16 November 2026. Something wrong? The source of this page is in the repository, so a fix is one edit.