AXG — Agent Execution Guard

Deterministic, auditable control over what AI agents are allowed to do.

AI suggests. AXG decides.

An agent proposes an action (write a record, run a shell command, call a tool). Before anything executes, AXG evaluates it against a declarative policy and answers ALLOW, SUGGEST, CONFIRM or BLOCK. An ALLOW comes with a signed Passport that binds the decision to the exact payload, so the system that executes the action can verify it was authorized and was not changed on the way.

• Deterministic: the same request under the same policy always gets the same decision. No model sits in the decision path.
• Framework-agnostic: any agent, orchestrator, MCP client or backend calls one HTTP API, or embeds the engine in process.
• Verifiable: RS256 Passports with payload hashing, JWKS and single-use jti; SDKs for Python and Node.
• Auditable: hash-chained audit log and OpenTelemetry traces for every decision.

Contents

• How it works
• Quickstart
• Integrations
• Decisions
• Security
• Documentation
• Status and roadmap
• Contributing

How it works

flowchart LR
    P["Agent, orchestrator or tool<br/>(any framework, any model)"] -->|DecisionRequest| AXG{{AXG}}
    AXG -->|ALLOW + Passport| X["Executor verifies the Passport,<br/>then acts"]
    AXG -->|SUGGEST / CONFIRM| H[Human in the loop]
    AXG -->|BLOCK| D[Denied]
    AXG -.->|spans + metrics| O[(OpenTelemetry)]
    AXG -.->|hash-chained records| A[(Audit log)]

1. The caller authenticates with an API key and sends a DecisionRequest: who acts (agent and permissions), what it wants to do (action_type, payload) and how sure the proposer is (llm.confidence).
2. AXG loads the policy plugin, evaluates its rules, enforces agent permissions and computes risk, confidence and uncertainty scores. The strictest outcome wins.
3. ALLOW responses carry a Passport. The executor verifies it with an SDK before acting. Everything else carries a human-readable reason and machine-readable audit flags.

AXG never calls a model and never executes the action itself. It is not an LLM wrapper, a prompt framework or an agent framework.

Quickstart

Run the server with an API key for your app:

export AXG_KEY=$(openssl rand -hex 32)
export AXG_CLIENTS="[{\"client_id\":\"quickstart\",\"key_sha256\":\"$(python3 -c "import hashlib,os;print(hashlib.sha256(os.environ['AXG_KEY'].encode()).hexdigest())")\",\"app_ids\":[\"*\"],\"permissions\":[\"*\"]}]"

docker run --rm -p 8090:8090 -e AXG_CLIENTS="$AXG_CLIENTS" ghcr.io/pinheirodps/axg:latest

Ask for decisions (from a clone of this repository, which provides the example requests):

# A coding agent wants to read a file: ALLOW, with a Passport
curl -s localhost:8090/v1/decisions -H "Authorization: Bearer $AXG_KEY" \
  -H "Content-Type: application/json" -d @examples/allow.json

# The same agent wants to run `rm -rf /`: BLOCK
curl -s localhost:8090/v1/decisions -H "Authorization: Bearer $AXG_KEY" \
  -H "Content-Type: application/json" -d @examples/block.json

{
  "decision": "BLOCK",
  "plugin_version": "claude-code@0.1.0",
  "reason": "This command can irreversibly destroy files or devices.",
  "rules_triggered": [{"id": "destructive_shell_command", "decision": "BLOCK", "reason": "..."}],
  "scores": {"risk_score": 1.0, "risk_level": "high", "final_confidence": 0.49},
  "passport": null
}

Or embed the engine in a Python process:

import asyncio
from pathlib import Path

from axg import DecisionEngine, DecisionRequest
from axg.plugin_loader import PluginLoader

engine = DecisionEngine(loader=PluginLoader(Path("plugins")))

async def main():
    decision = await engine.decide(DecisionRequest(
        execution_id="exec-1", tenant_id="acme", app_id="claude-code", plugin_id="claude-code",
        source="coding_agent", action_type="Bash", payload={"command": "git push --force origin main"},
        llm={"confidence": 0.92},
    ))
    print(decision.decision.value, "-", decision.reason)
    # CONFIRM - This Git operation can discard work or rewrite shared history.

asyncio.run(main())

Install the current release with pip install "axg @ git+https://github.com/pinheirodps/axg@v0.4.0" (add [otel] for OpenTelemetry export). PyPI and npm packages are prepared by the release workflow and will be announced in the changelog once uploaded.

Integrations

Example policies live in plugins/: budget_guard (AI spend), claude-code (coding agents), ticket2pr (a headless coding agent driven from Jira tickets), finnorte (personal finance writes) and pocket_lawyer (legal assistant). Write your own with the policy guide.

Decisions

Precedence is BLOCK > CONFIRM > SUGGEST > ALLOW: the strictest applicable outcome wins. CONFIRM and SUGGEST carry a signed approval ticket: the right person approves, and AXG exchanges the ticket for a single-use Passport, with no state kept in AXG (human approvals). Errors never fail open: a policy that cannot be loaded, or a Passport that cannot be signed, results in CONFIRM. See how AXG decides.

Security

• Authenticated callers. API keys, stored as SHA-256, bind each caller to the apps it may request decisions for and cap the permissions it may grant its agents. Anonymous callers are rejected (401) or, in migration mode, can never receive ALLOW.
• Passport v2. A short-lived RS256 JWT with jti, nbf, tenant, caller (azp), policy version and a canonical hash of the whole actionable payload. Keys are published as JWKS, with rotation.
• Hardening. Request size limits, per-caller rate limits, no dynamic code in policies, allow-listed remote policies, a hash-chained audit log, pinned dependencies and actions, SBOM and provenance on images.

Read the security model before exposing AXG outside a private network. Report vulnerabilities privately as described in SECURITY.md.

Documentation

Status and roadmap

AXG is beta (v0.4.0) and runs in production. The API may change before 1.0; breaking changes are versioned in the contracts and listed in the changelog.

Next: the first upload of the packages to PyPI and npm.

Contributing

Contributions are welcome. Start with CONTRIBUTING.md and the code of conduct. CI requires at least 98% test coverage of the core and integrations.

pip install -e ".[otel,test]"
python -m pytest --cov=axg --cov=integrations --cov-fail-under=98
python -m uvicorn axg.api:app --reload --port 8090

License

Apache-2.0