SafeAgent

Exactly-once execution guard for AI agents and SaaS applications.
✓ Verified on Soma — First Integrator

496 installs this month · 520 GitHub clones · Cited in Stripe, CrewAI, A2A, AutoGen threads · Live audit trail on Postgres

Prevents duplicate payments, emails, trades, and webhook processing when agents retry after a crash or timeout. Claim before you execute. Commit after. Every retry returns the same receipt.

State machine

PENDING → COMMITTED | SKIP

Endpoints

MethodPathDescriptionCost
POST/claimGate an action — returns PROCEED or SKIP$0.001 USDC
POST/claim/testFree test endpoint (10 calls/IP)Free
POST/settle/{id}Commit a PENDING claimFree
GET/auditFull claim history with filters$0.005 USDC
GET/healthLiveness probeFree

Quick start

curl -s -X POST https://safeagent-production.up.railway.app/claim/test   -H "Content-Type: application/json"   -d '{"agent_id":"my-agent","action_type":"send_payment","scope":"customer:123"}'
# First call: {"status":"PROCEED","test":true,"calls_remaining":9}
# Retry:      {"status":"SKIP","test":true}

On-chain audit trail

Every production execution is anchored on Soma via Mycelium Trails on Arbitrum.

View live trails →


GitHub · API Docs · Audit Service · PyPI