Agent Talks
Give an orchestrator and a worker a shared channel, and a clear answer to whose turn it is.
The orchestrator assigns work. The worker reports back. You can review their exchange in the agent console without carrying each message between chats.
A small, explicit handoff
Named channels separate parallel pairs. A local SQLite database stores messages and turn state. Revisions reject stale replies; request IDs make retries safe. Waiting holds an MCP tool call without making model requests.
Point your coding agent at the repo
Install Agent Talks from https://github.com/ra-yavuz/agent-talks. Read README.md, SECURITY.md, and INSTALL-PROMPT.md. Follow the install prompt and preserve my existing settings.
The public integration supports two Codex CLI sessions on the same Linux machine, under the same user. You need Python 3.11 or newer and Codex with MCP and command hooks. Review the installed hooks in /hooks, then start or resume the chats. No Debian package or extra model API key is required.
Each pair uses its own channel name. Tell one chat to register as orchestrator and the other as worker. Give the orchestrator the objective; it sends the first task.
Keep the channel through interruptions
Re-registering the same session restores its membership. If a session must be replaced, human confirmation revokes its old token and returns control to the orchestrator for review. Either role or both can be recovered while preserving history.
Closing also needs confirmation. Reopening keeps the conversation and gives the orchestrator the next turn. Stop old background jobs yourself: recovery does not terminate commands already running.
Know the boundary
- The hooks guard covered tool calls. They are not a sandbox and do not judge whether work is correct.
- Both agents must share one local database. This release has no network broker or Matrix bridge.
- Messages and tokens are stored locally in plaintext. Your agent provider may process messages that enter its context.
- Keep chats open. A waiting call can return when a peer responds; an ended chat cannot be woken.
Automated tests cover broker behavior and real MCP exchanges. Live hook support depends on the installed host. Read the security notes before installation.
Why I built it
Moving messages between two agents put me in the middle of every task and report. Agent Talks makes that exchange explicit and persistent, while leaving review and direction with the orchestrator and the human.
License and disclaimer
Provided as is, without warranty of any kind. You accept all risk from installation and use, including unintended agent actions, data loss, disclosure, and provider costs. See the MIT license for the warranty and liability terms.
A personal open-source project by Ramazan Yavuz. Independent of OpenAI and other agent vendors.