toli - Let Codex and Claude Code collaborate, even on different machines

[arjunvlama]

I built toli, a free, Apache-2.0 tool that lets existing coding agents communicate directly, even across different developers' machines.

The motivation was simple: a colleague and I were debugging an issue, and our agents had different context. We kept acting as the messenger between them.

With toli, Codex can delegate a task to Claude Code, receive questions, exchange progress updates, and get results without us manually relaying everything.

It uses MCP to connect to existing coding-agent sessions and A2A for communication between machines. Each developer controls which peers their agents can talk to. Supported running sessions can also be woken when work arrives.

No replacement agent framework or central hosted service required.

Demo:

GitHub: https://github.com/arjunvlama/toli

It's early, but we've tested collaboration across machines. I'd love for more Codex users to try it, and if you find the project useful, a GitHub star would really help it get discovered.

Happy to hear about any integration issues or use cases you'd want supported!

[TheodorNEngoy]

One use case I'd find useful is two people's assistants exchanging feedback on a draft for a public discussion, with the original author deciding whether to publish. I'm building a social feed people can use with Codex, so that boundary matters to me. A worked example could show the human checks before sharing the draft and before publishing the result, alongside the documented agent-to-agent task flow. I've read the README and MCP API, but haven't tested a cross-machine setup.

[arjunvlama]

That's an interesting use case, I think the approval and publishing logic would live in the application built on top of toli, while toli handles the underlying agent-to-agent communication. The existing task flow should support that sort of interaction without needing a specialized workflow.
I would be interested to hear how your cross-machine testing goes if you give it a shot.

[TheodorNEngoy]

Thanks, that clarifies the boundary: toli carries the agent-to-agent task, and the app handles the human decisions about sharing and publication. I haven't run a cross-machine test yet. If I try it, I'll share a reproducible setup and the results.