Keep OpenHands.
Give it your machines.
AI Commander is not an alternative to OpenHands, Claude Code, Codex, ChatGPT, Cursor, GitHub Copilot, Sourcegraph Cody, Grok, Gemini, Windsurf, Continue, Cline, Roo Code, Zed, OpenCode, Qwen Code, goose, Amp, or Kilo Code. It is the remote-machine harness those tools plug into over MCP: outbound-only shell, detached jobs that outlive the chat, and file transfer on hardware you already own.
- Install the host agent on the machine you want your AI to control. It shows a private machine code.
- Connect your AI: paste this prompt, then replace
AIC-…with your code —Help me connect this AI client to AI Commander using https://aicommander.dev/start.md. Then check the status of my machine: AIC-…— or add the connector URLhttps://aicommander.dev/mcpand sign in with an email code (per-client steps).
Sibling harness pages: Cursor · GitHub Copilot · Cody · Aider · Cline · Roo Code · Zed · OpenCode · Qwen Code · goose · Amp · Kilo Code · Continue · Windsurf · Claude Code · Grok · Codex · ChatGPT · Amazon Q · Augment · Devin · Gemini.
Connect OpenHands (MCP HTTP)
OpenHands (CLI MCP docs) connects to remote MCP servers over HTTP (and stdio / SSE). Point it at AI Commander’s hosted MCP endpoint:
The legacy CLI keeps servers in ~/.openhands/mcp.json under mcpServers
(restart the conversation after edits; use /mcp in-chat to check status):
{
"mcpServers": {
"aicommander": {
"url": "https://aicommander.dev/mcp",
"auth": "oauth"
}
}
}
Agent Canvas (local or Cloud): Customize → MCP Servers → Add custom server, type SHTTP, same URL. Prefer a local stdio bridge?
Full client matrix (OpenHands, Claude Code, Codex, ChatGPT, Cursor, Copilot, Cody, Grok, Gemini, OpenCode, Qwen Code, goose, Amp, Kilo Code, Continue, Windsurf, Cline, Roo Code, Zed, REST): Connect your AI client. OpenHands MCP docs: CLI → MCP Servers.
Over the remote URL, first connect may open a one-time browser sign-in (OAuth Dynamic Client Registration).
The npx bridge does not sign in through a browser: without AICOMMANDER_TOKEN in its env it connects anonymously with a code-only first hour; set it to your API key to bind it to your account.
On the remote URL, append ?anonymous=1 to skip the sign-in, for a code-only first hour.
For optional account features (saved aliases), pass a Bearer API key header where your OpenHands build supports it.
What you get that the chat alone cannot
- Remote shell on Linux, macOS, and Windows — by session code or saved alias.
- Detached jobs for builds, training, and headless agents on the box that has the repo.
- Fleet list + GPU facts so OpenHands picks the right machine before it burns an hour.
- File courier (Pro) for checkpoints and diffs up to 100 MiB.
Not a replacement — a multiplier
| You already use | AI Commander adds |
|---|---|
| OpenHands | openhands mcp add aicommander --transport http https://aicommander.dev/mcp |
| Amp | amp mcp add aicommander https://aicommander.dev/mcp |
| goose | goose session --with-streamable-http-extension "https://aicommander.dev/mcp" |
| Kilo Code | mcp.aicommander.url → https://aicommander.dev/mcp |
| Claude Code | claude mcp add --transport http … → remote exec + jobs |
| Codex / ChatGPT | Same MCP endpoint or REST /api/v1/exec |
| Cursor / Continue / Windsurf | npx @aicommander/mcp stdio bridge |
| Copilot / Cody | HTTP MCP / cody.mcpServers + bridge |
| Cline / Roo Code / Zed | MCP settings / context_servers + bridge or remote URL |
How OpenHands handles MCP servers
- OpenHands supports SSE, Streamable HTTP (SHTTP) and stdio servers. The docs advise against stdio in production and suggest a proxy such as supergateway. For AI Commander, the HTTP endpoint is the simpler choice.
- In Agent Canvas, go to Customize → MCP Servers → Add custom server. Test connection reports how many tools the server exposes.
- Auth options are none, a bearer token, a custom header, or OAuth. OAuth tokens are cached on disk by the FastMCP client under
~/.fastmcp/oauth-mcp-client-cache/. - Agent Canvas on a local backend saves servers in
~/.openhands/settings.json, with credentials encrypted, and has no per-repository MCP file. The legacy CLI uses~/.openhands/mcp.jsoninstead. - Server changes take effect in new conversations. One that is already running keeps the old tool list.
- Each server has a timeout in seconds.
Checked against OpenHands docs: MCP settings in November 2026. Client behaviour changes between releases; the official docs win if they disagree.
A workflow that suits OpenHands
OpenHands runs its agent in a sandbox, usually a container. AI Commander lets that sandboxed agent run commands on real machines outside it (as root by default), without opening SSH into the container.
- Add AI Commander as an SHTTP server with a bearer API key, so the connection doesn't depend on a browser sign-in.
- Put standing rules in the repo's
AGENTS.md, for example: "Run integration tests on the staging box viaremote_job_start, never on production; pollremote_job_statusevery few minutes." - Start a new conversation, have OpenHands fix the code in its sandbox, then run the tests on the real staging machine and report the result.
Limits and gotchas in OpenHands
- The OAuth option starts with an interactive browser sign-in. For unattended runs, a bearer API key avoids that step.
- If a new tool doesn't appear, you are probably still in the conversation that was open before you changed the settings.
FAQ
openhands mcp add aicommander --transport http https://aicommander.dev/mcp. Or edit ~/.openhands/mcp.json under mcpServers with "url": "https://aicommander.dev/mcp" (and "auth": "oauth" when you want the browser sign-in). Or Customize → MCP Servers in Agent Canvas. First connect may open a one-time browser sign-in. On the remote URL, use ?anonymous=1 for a code-only first hour. See /howto/ for the full matrix.session_status, list_machines and remote_exec before you start a conversation.Next steps
Install the agent on a machine you control, connect OpenHands via openhands mcp add … --transport http, then ask for a real task on a remote box.