Put your AI agents
under control.
Phrony is the open-source runtime for governing AI agents in production. Define each agent in one file, set the rules it can never break, require a human for risky actions, and keep a full record of everything it does.
Four controls that keep agents in check
One agent, one file you control
Everything about an agent lives in a file you can read, review, and version — just like code. Eight simple building blocks cover it all, from a single agent to a whole fleet.
Agent
Who the agent is and how it should behave — its job, its limits, and the shape of what it returns.
Secrets
Point to the credentials the agent needs by name only. The real values are pulled in safely when it runs.
Tool
What a tool takes, what it returns, and how risky it is — the contract, kept separate from the code.
Policy
Simple if-this-then-that rules, checked every time the agent acts. When rules disagree, deny always wins.
Tools
Decide exactly which tools an agent is allowed to use — and add extra rules just for that agent.
Bundle
When one agent delegates to others, package the root and every delegate together — publish, deploy, and run the whole system as one unit.
Agent delegation
Let one agent call another to get a job done — with the same rules and full record applied throughout.
MCP servers
Bring in remote MCP tools like any other tool — governed by the exact same rules as the rest.
Govern your first agent in minutes
Run Phrony on your own machine, write one file, and watch it enforce your rules on a live agent — free and open source.
Why teams pick Phrony
Frameworks help you build an agent. Phrony is what keeps it safe once it touches real systems — with control, proof, and a way back that you own end to end.
- 01
One file per agent
Policies, tool access, limits, and human checkpoints live in one versioned manifest — not scattered across application code. Your team reviews the agent file in git like any other config.
- Change what the agent is allowed to do without redeploying the app that calls it.
- Publish a new locked version when the rules change, then deploy when you are ready.
- 02
Instant rollbacks
Every publish creates an immutable version with a content digest. Deploy promotes a version to active; rollback restores the previous deployment in one command.
- Version history stays in the ledger — you always know what was live and when.
- If a new deployment misbehaves, roll back before it affects more sessions.
- 03
Every action checked
The runtime sits between the model and your systems. On every tool call it evaluates your policies, enforces limits, and pauses for human approval when required.
- Nothing reaches production APIs unless the manifest allows it.
- Every policy decision is written to the session trace for audit and replay.
- 04
Open spec and runtime
The Agent Spec defines the manifest schema; the reference runtime is Apache-2.0 on GitHub. Run it locally, validate manifests, and deploy agents on your own infrastructure.
- No vendor lock-in — conformant implementations can run the same agent files.
- The same CLI you use locally governs agents in production.
Everything you need to get started
- Quick start
Run Phrony locally, write your first agent file, and put it under control in a few minutes.
- Documentation
Everything you need to run governed agents in production — the ideas, the agent file, and the runtime.
- Blog
Notes and guidance from the Phrony team — starting with why production agents need a runtime, not just a framework.
Start with the open-source runtime
It is free, it runs on your own machine, and it is the same runtime that governs agents in production.
Try the quick startStay connected
Open-source runtime
Phrony is free and open source. Run it locally with Docker, check your agent files, deploy agents, and drive them from the command line.