[Feature] Possible to run a PicoClaw instance as a Node of an OpenClaw gateway instance?

[chrisbward]

🎯 The Goal / Use Case

Would be nice if these implementations followed a settled standard so that we could mix/match.

I really want to use PicoClaw, but I'd also like to take advantage of the OpenClaw ecosystem and tooling that is now emerging.

This would be a killer feature I think.

[huaaudio]

Hi, thanks for the feedback! I think even for now, you can already install and configure both instances and instruct OpenClaw to call PicoClaw by picoclaw agent -m "MESSAGES".

[tbaumann]

I would say, let's wait until the dust settles. A lot of how the OpenClaw tooling works is blind cli tool calling with zero respect for secrets protection.
I wouldn't be surprised if OpenClaw started to embrace MCP too.
It's all node.js, but if you have the resources, I don't see much reason why picoclaw couldn't call OpenClaw plugins.

Realistically most 'tools' are just skills. Just use them as they are.

[nikolasdehor]

This is an interesting integration question. As @huaaudio noted, the simplest path today is using PicoClaw as a CLI tool that OpenClaw invokes via picoclaw agent -m. However, there are practical considerations for deeper integration. Session management is duplicated (OpenClaw context plus PicoClaw context means double token usage). CLI invocation loses streaming, so the user sees OpenClaw thinking until PicoClaw finishes its full response. And tool delegation becomes invisible to OpenClaw audit logs. A cleaner long-term integration path would be MCP (Model Context Protocol): PicoClaw could expose its tools as an MCP server, and OpenClaw (Node.js-based) could consume them as an MCP client. This gives tool-level interop without full agent delegation overhead. As @tbaumann mentioned, the ecosystem is still settling. I would suggest starting with the CLI skill approach for now and considering an MCP server mode for PicoClaw if demand grows.

[pormr]

I‘m fascinated by the idea of local agent delegation (preferably in an isolated manner). Is there a lightweight toolkit or plugin enabling direct multi-agent interoperation over IP/sockets, without dependencies on external providers?
Based on my recent investigation, OpenClaw can serve as a model endpoint speaking Open Responses specification¹, which allows encoding the agent id in the model field of request body (e.g., model: "openclaw:<agentId>", model: "agent:<agentId>"). This seems like a promising solution for constructing multi-agent architectures.

Footnotes

1. https://docs.openclaw.ai/gateway/openresponses-http-api ↩

[sipeed-bot]

@chrisbward Hi! This issue has had no activity for over 2 weeks, so I'm closing it for now to keep things tidy. If it's still relevant, feel free to reopen it anytime and we'll pick it back up.