Keep Roo Code.
Give it your machines.
AI Commander is not an alternative to Roo Code, Cline, Cursor, Claude Code, Codex, ChatGPT, Gemini, Grok, Windsurf, or Continue. 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 · Aider · Cline · Roo Code · Zed · OpenCode · Qwen Code · Continue · Windsurf · Claude Code · Grok · Codex · ChatGPT · Gemini.
Connect Roo Code (MCP settings)
Roo Code is the VS Code AI agent extension (fork lineage of Cline, with its own MCP settings).
It talks to AI Commander through the same stdio MCP bridge as Cursor, Cline, and Windsurf.
Add this under mcpServers in Roo’s MCP settings
(open Roo Code → gear → MCP Servers, then Edit Global MCP / Edit Project MCP,
or edit mcp_settings.json / project .roo/mcp.json):
{
"mcpServers": {
"aicommander": {
"command": "npx",
"args": ["-y", "@aicommander/mcp"],
"env": {
"AICOMMANDER_SERVER": "https://aicommander.dev"
}
}
}
}
The config shape matches Cursor and other IDE MCP clients; only the settings path differs. Full client matrix (Roo Code, Cline, Cursor, Continue, Windsurf, Claude Code, Codex, ChatGPT, Gemini, Grok, REST): Connect your AI client.
Without AICOMMANDER_TOKEN in its env, the bridge connects anonymously with a code-only first hour. Set it to your API key to bind the bridge to your account.
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 Roo 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 |
|---|---|
| Roo Code | Roo MCP settings / mcp_settings.json / .roo/mcp.json + npx @aicommander/mcp |
| Cline | Cline MCP settings / cline_mcp_settings.json + npx @aicommander/mcp |
| Continue | ~/.continue/config.yaml / .continue/mcpServers/ + stdio bridge |
| Cursor | npx @aicommander/mcp stdio bridge → remote exec + jobs |
| Windsurf (Devin Desktop) | ~/.config/devin/mcp_config.json + stdio bridge |
| Claude Code | MCP remote exec + jobs on your fleet |
| Codex / ChatGPT | Same MCP endpoint or REST /api/v1/exec |
| Gemini / Grok | gemini mcp add / grok mcp add --transport http … |
How Roo Code handles MCP servers
- Roo reads two files: a global
mcp_settings.json(opened from the MCP settings view) and a per-project.roo/mcp.json. If both define a server with the same name, the project entry wins, so a repo can pin its ownaicommanderentry. - Remote servers need an explicit
"type":"streamable-http"forhttps://aicommander.dev/mcp, or"sse"for legacy servers. A URL entry withouttypeis rejected straight away instead of being guessed. - Every server has its own timeout, 1 to 3600 seconds, with 60 seconds as the default.
- Per-tool switches live in the same JSON:
alwaysAllowlists tools that run without a prompt,disabledToolshides tools from the model, and${env:VAR}pulls values from your environment intoargs. alwaysAllowonly takes effect when the global auto-approve option for MCP tools is switched on in Roo's settings.- On Windows, stdio servers started through npx need
cmd /c npx …as the command.
Checked against Roo Code docs: Using MCP in Roo Code in November 2026. Client behaviour changes between releases; the official docs win if they disagree.
A workflow that suits Roo Code
Roo is good at long, multi-step tasks inside one repo. A pattern that works: keep the code in your editor and let Roo drive a separate build or test box.
- Add
aicommanderto your globalmcp_settings.jsonas a"streamable-http"entry forhttps://aicommander.dev/mcp, with your ownAuthorization: Bearer <API key>header. Keep keys out of any committed.roo/mcp.json. - Put the read-only tools in
alwaysAllow:list_machines,session_status,remote_job_statusandremote_job_logs. Leaveremote_execandremote_job_starton manual approval. - Ask Roo to start the full test suite on the build box with
remote_job_start, keep refactoring locally, then have it readremote_job_logsand fix whatever failed.
Limits and gotchas in Roo Code
remote_execreturns one buffered result, so anything longer than Roo's 60-second default timeout fails on the client side. Raise the server timeout, or better, useremote_job_startfor builds and training.- A project
.roo/mcp.jsonis usually committed, so never paste an API key into it. - Turning off MCP in Roo's settings removes the MCP instructions from the system prompt, which saves tokens. While it is off, Roo will not see AI Commander at all.
FAQ
mcp_settings.json / .roo/mcp.json) with npx -y @aicommander/mcp and AICOMMANDER_SERVER=https://aicommander.dev. Without AICOMMANDER_TOKEN in env the bridge connects anonymously, code-only; set it to your API key to bind it to your account. See /howto/ for the full matrix.remote_exec runs past it. Raise timeout for the server (up to 3600 seconds), or ask Roo to use remote_job_start and poll remote_job_status..roo/mcp.json at the repo root overrides a global entry with the same name, and it is usually committed, so it must not hold a key. Let each person add the server to their own global settings with their own API key. The npx stdio bridge runs anonymously (code-only) unless AICOMMANDER_TOKEN is set.Next steps
Install the agent on a machine you control, connect Roo Code via MCP settings, then ask for a real task on a remote box.