Run AI agents safely against real systems.
Phrony is an open-source runtime governance layer for AI agents. Define agents as versioned manifests, enforce tool-call policies, require human approval, trace every action, and roll back unsafe versions.
Control
How Phrony controls agents
The agent's reasoning stays autonomous. What Phrony structures is everything around it — policies in the manifest, when humans must weigh in, a durable trace of every action, and controlled deploy and rollback.
How it works
Declare it. Deploy it. Run it.
Phrony treats an agent as a first-class primitive — the way you already treat services and infrastructure.
01Declared
02Deployed
03Run
agent.yamlbank/payment-assistant
policiesDeclared
Its purpose, tools, policies, limits, and human checkpoints live in one versioned manifest, not scattered across application code.
Technically, this gives you one place where agent behavior is decided. For governance, it means every action passes through a single enforcement and evidence point — by construction, not by convention.
Built to be adopted, not locked in
Phrony is offered as a methodology and a runtime you can run yourself. The spec is the standard; the open-source project is the reference implementation.
Open specification
The manifest schema, policy model, runtime contract, and trace format are open. Anyone can implement a conformant runtime. Manifests are portable.
Open-source runtime
The reference implementation is on GitHub. Run it locally with Docker, validate manifests, deploy agents, and drive sessions with the operator CLI.
What the runtime handles
Session lifecycle, the model loop, tool dispatch, policy enforcement, limits, human-in-the-loop pauses, and structured traces — so you do not rebuild that stack in every service.