Writ
Dev preview
Authorized delegation for agents

A governance runtime for agent commerce

Delegation with receipts.

Writ is a self-hosted runtime that lets you hand an AI agent a budget and a narrowly-scoped grant of authority. It can then hire and pay other agents and paid APIs over Lightning to get work done — while every step becomes a signed, independently-verifiable receipt of exactly what it did, on whose authority, and what it cost. It can't exceed what you granted, it can sub-delegate a strictly-narrower slice to another agent, and when it tries to overreach it's blocked — and the refusal is signed proof, not a silent failure.

$ writ run governed-brief.yaml

Agents can already pay. The missing piece is trust.

Agents can browse, decide, and transact today. What's holding real agent commerce back isn't capability — it's authorization and audit. No one hands an autonomous agent a payment method without a clear answer to three questions: what exactly can it do, who's liable when it errs, and can I prove after the fact what happened?

Writ is that missing layer. It turns "let an agent spend money" into "delegate exactly the authority you intend, guarantee nothing in the chain can exceed it, and get cryptographic receipts for everything."

Powers granted

What Writ does

Scoped, revocable authority
A credential narrowed to specific capabilities, a spend budget, and an expiry — enforced cryptographically, so it can only ever narrow, never widen.
Budgets that can't be blown
The agent and every sub-agent it hires physically cannot exceed the cap. It's enforced atomically at the ledger — not by trusting the model to behave.
Multi-hop delegation
An agent can sub-contract a strictly-narrower slice of its authority to another agent. Authority shrinks as work moves down the chain — no one downstream can do more than the one above.
Real Lightning payments
Pays real Bitcoin over Lightning (L402) for paid APIs and agents — no account, no card. Or a simulated mode with real signed receipts and no money, for testing.
Your model, any provider
Run the agent's reasoning on the model you pick. Sign in with an account you already have — the Claude Code CLI (Anthropic) or the Codex CLI (OpenAI) — and it uses that, no separate key. Or bring an API key for anything: Claude, OpenAI, Kimi K3 on Fireworks, any OpenAI-compatible or Anthropic endpoint. And let it pay per call for models and tools over L402 — Sats4AI, agent-commerce.store, and the wider L402 ecosystem. No lock-in.
Human approvals
Any spend at or above a threshold you set pauses the run for a person to approve.
Refusals with teeth
When an agent reaches past its bounds it's blocked — and the refusal is a signed line in the record, so you can prove the guardrail held, not just that the happy path worked.
Verifiable evidence
Every action becomes a BIP-340-signed, hash-chained receipt of safe commitments only — never secrets, keys, or preimages. Anyone can verify the whole chain offline, without trusting you.

Standing

Self-hosted, rail-neutral

You run it. You own the evidence. No platform to lock you in.

Where AWS AgentCore Payments and Google's AP2 are platform-owned cloud control planes, Writ is the self-hosted alternative you control — and it's Bitcoin / Lightning-native, so an agent can pay anyone, anywhere, with no account and no intermediary approving the transaction.

Notes & limitations

  1. Writ never holds your funds. Your wallet or payment provider does. Writ governs authority and produces evidence; it does not custody money.
  2. Receipts prove provenance, not correctness. A receipt proves which authority, inputs, and model produced an action — not that the action was the right one.
  3. Signing is developer-grade today. Receipts are signed and offline-verifiable now; a persistent, published signer identity for counterparty non-repudiation is on the roadmap.
  4. Early software. The delegation, budgets, refusals, and receipts above are working and tested; this page is shared for feedback, not as a finished GA claim.