authgraph

Show me what will happen before the AI executes anything.

Python execution planning + TrigGuard authorization simulation. No permit, no run.

pip install authgraph

from authgraph import Plan, simulate

plan = Plan()
plan.add("file.read", read_file, title="read_file")
plan.add("git.push", push_branch, title="push_branch")

result = simulate(plan)
print(result)          # predicted PERMIT / ESCALATE / BLOCKED per step
print(plan.visualize())  # execution graph

Before anything runs, authgraph asks TrigGuard. Simulation never executes your callables.

Your code → authgraph → TrigGuard → Gateway

Not a CrewAI/LangGraph/Quanta product. Optional integration hooks live under authgraph.integrations and do not bundle external runtimes.

Requires TRIGGUARD_API_KEY, TRIGGUARD_ORG_ID, and gateway https://api.trigguardai.com (default).

Full guide: docs/QUICKSTART.md

APIs (0.3)

from authgraph import Plan, PlanBuilder, simulate, explain, guard

CLI:

authgraph simulate examples/deployment.py
authgraph explain receipt_123 --action deploy.release
authgraph graph examples/deployment.py --dot

Demos: examples/demos/ (agents, GitHub Actions, DB migrate, Terraform, K8s, robotics)

Quick start (@guard)

import os
from authgraph import AuthgraphRuntime, run_step

os.environ["TRIGGUARD_API_KEY"] = "tg_live_..."
os.environ["TRIGGUARD_ORG_ID"] = "org_..."

runtime = AuthgraphRuntime()
graph = runtime.create_graph(title="Deploy", description="Plan then ship")
repo = os.environ.get("TRIGGUARD_REPOSITORY", "your-org/your-repo")
plan = runtime.add_node(
    graph, title="Plan", description="Define scope",
    authority_surface="deploy.release", metadata={"repository": repo},
)
runtime.start(graph)

result = run_step(
    runtime, graph, plan, agent_id="my-agent",
    fn=lambda: {"spec": "ship it"},
)
print(result.receipt, result.execution_time)

More examples: examples/

What is shipped (0.3)

• Execution plans — Plan, PlanBuilder, sequential steps with metadata
• Simulation — simulate() / plan.simulate() never runs callables; queries TrigGuard
• Graph execution — plan.execute() with downstream BLOCKED cascade
• Explainability — explain() for receipts, decisions, gateway JSON
• Risk breakdown — result.risk_score, result.risk_factors (local heuristic, not policy)
• Visualization — console tree + Graphviz DOT export
• CLI — authgraph simulate|explain|graph
• Universal execution engine (authgraph.execution) — one claim/execute/complete path
• @guard, authorization(), sync + async + batch
• Optional integrations (CrewAI, LangGraph, …) — thin hooks only
• Fail-closed TrigGuard gate

Not shipped: hosted service, bundled agent frameworks, policy duplication, GitHub Apps, K8s controllers, Terraform providers (build when design partners ask).

Environment

License

See LICENSE.