Platforms get API traffic from AI agents trying to do real economic actions — refill an account, top up credits, place an order — and block it as fraud, because nothing answers "who committed this action?" for an agent. Writ resolves any action to its chain of agency: action ← agent (Ed25519 keypair + runtime metadata) ← mandate (scope, spend caps, expiry — revocable live) ← KYC'd principal ← liability. Platforms integrate one middleware (requireKYA) and get allow/deny decisions plus signed, tamper-evident records; agents carry the passport via SDK or MCP server. Records, not rails: no custody, no tokens, rails-neutral (x402 / AP2 / agentic-checkout interop). Live now: a demo brokerage where an anonymous agent gets 403'd into onboarding, a passported agent refills under a $2,500/week mandate, the cap denies at the boundary, and revocation lands on the next call — 11/11 acceptance checks green against production. Launch plan: design-partner pitches to usage-billed AI platforms are going out now, the Y Combinator application is in flight (late/rolling), and the public launch lands on writhq.com once the domain is wired.
Palette
Milestones
- Jul 30Renamed to Writ — npm packages (@writ/sdk, verify, mcp, demo) and the public repo ship under the final brand
- Jul 30Shipped hard: secrets hashed at rest, rate limiting, nonce sweeper — then rebranded to Onbehalf; production re-verified 11/11
- Jul 30Production verified end to end (11/11 acceptance) — YC Fall 2026 application drafted, design-partner outreach begun
- Jul 26Marketing site live, pointed at the running sandbox
- Jul 24MVP built: verify core, mandates, middleware, agent SDK + MCP + CLI, Northbank demo brokerage — 47 tests
- Jul 23Protocol spec v0.1 written and the build greenlit — chain-of-agency model, mandate schema, verify API