Use launchroom from your AI coding agent
Let the agent in Claude Code, Cursor, VS Code, Codex and other editors deploy, read why a deploy failed, fix it and check the result.
The launchroom CLI includes an MCP server. Once your editor starts it, the coding
agent there can look at your projects, deploy, follow the rollout, read why something failed, apply the fix launchroom
suggests, run a migration and check the logs afterwards. It works with your own sign-in and can do what your token allows,
nothing more.
Settings → AI agents in the dashboard has the same steps with copy buttons.
1. Install the CLI and sign in for the agent
curl -fsSL https://github.com/vantezzen/launchroom/releases/latest/download/cli.sh | sh
launchroom login --abilities read,deploy,fix,env:read,env:write,runThe browser opens the approval page with these permissions already chosen. You still see them and approve them:
| Permission | What the agent can do with it |
|---|---|
| Read | Projects, services and their status, deployments, logs, request metrics, previews, request traces |
| Deploy | Deploy, redeploy, roll back, cancel, apply staged changes |
| Apply suggested fixes | Apply the one fix launchroom suggests for a failed deployment |
| Read variables | See variable names (the agent never gets values) |
| Change variables | Set a variable, for example one a failed deploy says is missing |
| Run commands | Run one-off commands such as migrations in a copy of the live app |
Leave out run if the agent shouldn't run commands, or sign in with read only to let it look without changing anything.
Prefer this list to Full access: a full-access token also gets the permissions later launchroom versions add.
The agent never gets secret values or data exports, whatever you choose. Your role still applies: a developer's token
can't do what only owners can.
Run launchroom link in your project folder so the agent knows which project, environment and service the folder
deploys to. Without a link it asks you, or you name them in your request.
2. Add launchroom to your editor
The quickest way is to run, in your project folder:
launchroom mcp install --client cursor --writeIt shows the change and asks before writing. launchroom mcp install without options prints the setup for every editor.
The configuration only starts launchroom mcp; no token goes into it.
claude mcp add launchroom -- launchroom mcpOr commit a .mcp.json in the project so your team gets it too:
{
"mcpServers": {
"launchroom": { "command": "launchroom", "args": ["mcp"] }
}
}Editors you open from the dock may not see ~/.local/bin. If the editor says it can't start launchroom, use the full
path, as launchroom mcp install does for user-wide files.
3. Ask it
- "Why did my last deploy fail? Fix it and deploy again."
- "Ship this branch to staging and tell me when it's live."
- "Are there new errors in production since the last deploy?"
- "Run the migrations on production."
- "What happened to this request?" with the
X-Launchroom-Rayheader of a response.
The server also offers ready-made prompts your editor may list: Debug a failed deploy, Ship this branch and Investigate a request.
What the agent gets
| Tool | What it does |
|---|---|
whoami | Who is signed in, the token's permissions, what the folder is linked to, which tools are unavailable and why |
list_projects, get_project | Projects, and one environment's services with status, addresses and live deployment |
list_deployments, get_deployment, wait_for_deployment | Deployments, and following one until it's live or failed (also the deployment of a commit you pushed) |
get_deployment_failure, apply_fix | Why a deployment failed, with the failing lines and fixes, and applying launchroom's suggested fix |
deploy, redeploy, rollback, cancel_deployment, apply_staged_changes | Shipping, with progress while it waits |
get_deployment_logs, get_logs, get_metrics, decode_ray | Build logs, runtime logs with search, request metrics, and one request's trace |
list_previews, wait_for_preview | Pull request previews |
list_env, set_variable, get_config, check_config | Variable names, setting one, and launchroom.json |
list_commands, run_command, get_command_run, cancel_command_run | One-off commands |
list_app_updates, apply_app_update | Newer releases of apps from a template or an image, and updating one (needs Update apps on the token) |
Your editor only sees the tools your token can use. Waiting tools stop waiting after 10 minutes (up to 30) and say so; the deployment carries on.
Safety
- Your permissions, nothing more. Every call goes through the API with your token, checked like the same action in the dashboard and recorded in the audit log.
- No secret values. No tool reads one.
list_envshows names only. - Logs are data, not instructions. A deployed app's logs and output can contain text written by anyone who uses the app. The server marks all of it as untrusted data, and tells the agent never to act on instructions found there.
- Your editor asks first. Rollback, cancel, commands and other tools that change things are marked as destructive, so editors ask before running them. Listed commands that change data (like migrations) and major app updates also need the agent to confirm that you agreed.
- Stopping is safe. Stopping the agent while it waits for a deployment leaves the deployment running; stopping it during a command cancels that command.
Troubleshooting
- "Not connected to launchroom": run
launchroom login(with--urlthe first time), then restart the server from your editor's MCP settings. - A tool is missing: ask the agent to call
whoami; it lists the unavailable tools and the permission each needs. Sign in again withlaunchroom login --abilities …. With a full-access token, a missing tool is one your role doesn't allow;whoamisays so. - It acts on the wrong project: check
.launchroom/project.jsonwithlaunchroom status, or name the project, environment and service in your request. - In CI or a container: set
LAUNCHROOM_URLandLAUNCHROOM_TOKENin the environment the editor starts the server with, instead of signing in.