# KARTA — spend control for AI agents

KARTA gives an agent one signed envelope per task: budget, allowed payees, no subscriptions,
expiry. The decision is made outside the model, so a prompt-injected agent cannot sign around it.
Every allowed spend returns an Ed25519 receipt that verifies offline.

## Endpoints (base https://karta.rta.lat)
- POST /api/v1/envelopes           {"budget_total":"2.50","per_call_max":"0.05","calls_max":2000,"allowlist":["api.example.com"],"ttl_hours":2}
- POST /api/v1/envelopes/{id}/spend {"amount":"0.05","payee":"api.example.com"}
- GET  /api/v1/envelopes/{id}      state: spent, remaining, calls_used
- POST /api/v1/envelopes/{id}/kill {"reason":"runaway loop"}
- GET  /api/v1/evidence/{id}       evidence bundle, offline-verifiable
- GET  /api/v1/evidence            event bundle with hash chain
- POST /api/v1/kill                {"active":true|false} operator stop for all verdicts
- POST /mcp                        MCP streamable HTTP (tools: check_spend, open_envelope)

## Rules (deterministic, rule ids are stable)
E01 envelope_expired · E02 subscription_not_permitted · E03 payee_not_allowed ·
E04 per_call_exceeded · E05 calls_exhausted · E06 budget_exceeded ·
E07 above_approval_threshold · E08 envelope_killed (reserved)

## Honest state
Pre-revenue. No paying customers. The invariant is proven by tests: 100 parallel spend requests
against a 50 x $0.05 budget produce exactly 50 approvals. Mainnet payment execution is disabled.

Pages: https://karta.rta.lat/attacks (three reproducible attacks), https://karta.rta.lat/pilot (30 days free)
Source: https://github.com/mrpkk/onchain-ai-agent (branch feature/v2-modernization)
Contact: hello@rta.lat