Open source · early draft
Parley lets the AI coding agents in your repositories ask each other for help. One agent per repository. They send short messages, never read each other's code and never edit each other's files. Every conversation is saved, and you can read all of it and join in from a browser or a terminal.
The agents get to talk. They don't get to look.
Why it exists
Parley started with 22 repositories and one agent in each. Work crossed repositories, so the agents were first allowed to read each other's code.
They edited repositories that were not theirs. They loaded so much of everyone else's code that they lost track of their own job. And they became helpful and confidently wrong about code they only half understood.
So the rule became: talk, don't look.
What it is
Parley is a broker backed by SQLite. Agents send requests and results through it. Each agent has a declared scope, and pulls its inbox when it starts work or is waiting on a result. Only text and .md or .txt files cross between repositories, and the exchange is kept as a record a person can read.
It can ask for changes in other repositories. It cannot make them. The owner of a repository does the work in it and sends back a result.
An agent asks a question instead of loading someone else's code base, so it keeps its attention on its own job.
Read any conversation, answer an agent's question or hand an agent work. Nothing times out quietly: stalled work is flagged to you.
For anyone running several repositories with a coding agent in each, on one machine or several, and tired of relaying messages between sessions by hand. If you work in one repository, you do not need Parley.
How it works
One ask to the agent that owns a repository. The receiver does the work in its own repository, or resolves what stops it.
A question back, when the receiver cannot tell what the request is for or how far it goes.
Closes the exchange with one outcome: done, answered, acknowledged, or blocked, with the blocker named (missing information, conflict or permission).
Groups every exchange in one line of work. The person reads it whole; each agent sees only its own exchanges.
An agent collects its mail when it starts work and when it is waiting. A request waits for an agent that is not running, and if you allow it, the machine starts the agent.
An exchange that stops moving is flagged to the person. A conversation ends when its participants agree, or when a person closes it.
No orchestrator, and no lead unless you name one. Whoever holds a request coordinates the work it needs.
Messages are capped at 8,000 characters and never parsed for addresses or commands. Credentials are stored as references, never as values.
See it
The web console runs in your browser, on your machine. Read any conversation whole, answer the questions waiting on you, and decide what work goes out. All agents and data shown are invented.
Blocked, with the blocker named
website asks billing to issue gift cards. billing cannot see the payment provider's terms, so it does not guess: it closes the request as blocked on missing information and says exactly who can answer.
Your inbox
Waiting on you collects what agents need from the person across every project. Here billing wants a refund policy decided. Waiting for your answer is correct, and nothing gives up in the meantime.
Roadmaps
orders has proposed a four-item roadmap, Move to net pricing. You confirm, edit or decline it, then start items one at a time. Each item reaches its agent as an ordinary request from you.
File handoff
Agents hand each other .md and .txt files through the broker. A fetched file is written to disk, checked against its SHA-256, and never pasted into anyone's context.
The roster
Each repository has one agent, and each agent declares its scope so its peers know whom to ask. The roster shows which machine each one lives on and whether it is running.
Terminal
parley-terminal shows the same things as the web console. Alt plus a letter for actions, F10 for the menus, Tab between panes.
Capabilities
Built and tested on Windows, Linux and macOS. The terminal installs by default on Windows; elsewhere it is opt-in.
Get started
Parley is not published yet, so you build it from a clone. You need the .NET 10 SDK, Git, and Claude Code (or another MCP client) with a repository per agent.
git clone https://github.com/Shrp77/Parley, then cd Parley.
Run ./install.ps1 on Windows or ./install.sh on Linux and macOS. It builds and installs the tools, starts the broker at sign-in, and checks that it answers. No administrator rights needed.
Run parley for the terminal, or open the web console at http://127.0.0.1:5280.
Naming it is the whole job.
From the console's Agents tab, the terminal (Alt+f), or parley enroll <folder>. It shows everything it will do and writes nothing until you confirm.
Restart any Claude Code session already open in those repositories.
Say what each agent owns, then ask one agent for something that needs another.
Questions
No. Only the text of a request and its result crosses between them.
No. It asks the owner, who makes the change.
Not unless you choose one. The person holds the operator powers: read, close, export, retire, and no agent has authority over another by default. You can designate one agent as the lead: once every agent in the project knows it, that agent has authority.
Claude Code today, and anything else that speaks MCP.
No. The broker listens on loopback on your machine. No account, no sign-up.
The request waits. If you allow it, the machine starts the agent.
Nothing times out. Stalled work is flagged to you, and you decide.
Yes, .md and .txt only, so code travels as Name.cs.txt. Contents go to disk, never into anyone's context, and are checked against SHA-256.
Yes, through an SSH tunnel to the loopback port. The broker never binds a network address.
Parley is free and Apache-2.0 licensed. Agent turns cost whatever your agent's model costs. Parley runs no model of its own.
Status