MIT License · Built in Rust

Cryptographic governance for AI agents

Every tool call signed. Every action auditable. Every mandate revocable. Zero LLM calls. Deterministic.

View on GitHub Quick Start
# Create governance identity
a2g sovereign
a2g init --name my-agent

# Sign the mandate
a2g sign --mandate my-agent.mandate.toml --key sovereign.secret.key --ttl 24

# Govern a tool call
a2g enforce --mandate my-agent.mandate.toml --tool read_file --params '{"path":"data.csv"}'
# => ALLOW | receipt: abc123 | hash-chained | signed

8-Step Enforcement Pipeline

Every decision is deterministic. No LLM calls. No probabilistic filtering.

STEP 0

Input Validation

Reject path traversal, oversized payloads, null bytes

STEP 1

Revocation

Check if mandate has been revoked before TTL

STEP 2

Signature

ed25519 verification with domain separation

STEP 3

TTL Check

Reject expired mandates (max 1 year)

STEP 4

Tool + Boundary

Allow/deny list, filesystem and network globs

STEP 5

Jurisdiction

Region, environment, operating hours

STEP 6

Escalation

Pattern-matched human approval triggers

STEP 7

Rate Limit

Per-minute call caps from mandate

Guardrails

  • Check output after execution
  • Probabilistic (LLM-based)
  • No audit trail
  • No revocation mechanism
  • No authorization proof

A2G Governance

  • Check authorization before execution
  • Deterministic (zero LLM calls)
  • Hash-chained cryptographic receipts
  • Instant mandate revocation
  • ed25519 signed proof of every action

Works with everything

Drop-in governance for your existing agent stack. 3 lines of code.

LangChain
CrewAI
OpenAI SDK
Claude SDK
MCP
8
Enforcement steps
0
LLM calls per decision
71
Tests (unit + battle)
<5MB
Single binary

Your agents need governance.
Not guardrails.

MIT License. Self-hosted. No vendor lock-in. Works with any LLM, any framework.

Get Started on GitHub