Every tool call. Scanned, scoped, signed, before it executes.

Your agents make tool calls you can't see, can't control, and can't prove. Two lines of Python fixes that.

184 tests · 2 dependencies · MIT licensed · works offline

pip install agentmint ⎘
View on GitHub See it work ↓
RUNTIME ENFORCEMENT
agentmint demo

github.com/aniketh-maddipati/agentmint-python

SIX ENFORCEMENT LAYERS

Scan

23 patterns catch injection, PII, credentials, encoding evasion.

Scope

Per-agent, per-action permissions with delegation and narrowing.

Rate limit

Circuit breaker kills runaway agents before they do damage.

Checkpoint

Sensitive actions block until a human approves.

Sign

Ed25519 + SHA-256 hash chain on every allow and deny.

Prove

Export evidence, verify with openssl alone, no vendor.

No platform. No sidecar. No policy language. Two dependencies, works offline, lives in your code.

Compliance mapping: SOC 2 · HIPAA · EU AI Act · AIUC-1 — details in COMPLIANCE.md

PROOF

What your stack produces today

2026-02-15 03:42:17 INFO  Tool called: delete_user
2026-02-15 03:42:17 INFO  Args: {"user_id": "usr_8291"}
2026-02-15 03:42:18 INFO  Result: success

// No proof of what happened
// No verification possible
// "The logs say it worked"

What AgentMint adds

{
  "receipt_id": "7d92b1a4",
  "agent": "sre-bot",
  "action": "delete_database",
  "args": {"target": "production"},
  "decision": "BLOCKED",
  "reason": "destructive action not in scope: ops-readonly",
  "checks": {
    "classification": "DANGER",
    "scope": "FAIL",
    "injection": "pass"
  },
  "signature": "Ed25519:a3f9...",
  "timestamp_rfc3161": "2026-02-15T03:42:17Z",
  "prev_hash": "a1f3c8e2..."
}

$ bash VERIFY.sh receipt.json
  Ed25519 signature:  ✓ verified
  RFC 3161 timestamp: ✓ verified
  Hash chain:         ✓ intact
VERIFIED — OpenSSL only, no vendor required
PROVE IT ON YOUR STACK

First one's free.

I'll run AgentMint against your agents this week. You get a working enforcement layer and compliance evidence your auditor can verify independently. You keep everything — the code is MIT, the receipts never expire. If it's useful, we talk about production. If not, we part ways.

Production integration: 2-3 weeks. You own everything when I leave.

Let's do it →

Aniketh Maddipati

Engineering Manager at Capital One. Built the ML fraud inference layer protecting 100M+ customers at 10k+ TPS. Led API security for third-party integrations across the enterprise. Now building runtime enforcement for AI agents.

Book 15 minutes

NYC, down to grab coffee.