Talk to Launchmatic from inside your editor

We just shipped the biggest single integration of the platform with the way most of our users actually write code in 2026 — through their AI assistant.

There are now three new ways to operate Launchmatic without leaving the editor.

1. Remote MCP at mcp.launchmatic.io

Connect Claude Code, Cursor, Codex, OpenCode, GitHub Copilot — anything that speaks the Model Context Protocol — directly to your Launchmatic account. No API keys to copy. No tokens on disk. OAuth in your browser, scopes you actually picked, revocable per-app from the dashboard.

claude mcp add launchmatic --transport http https://mcp.launchmatic.io

That's it. First call opens a browser, you approve which workspaces the client can touch, and you're connected. Your AI assistant can now see your projects, trigger deploys, set environment variables, fetch logs from a crashing pod, register custom domains — without you context-switching to a terminal.

We expose 11 concrete tools plus one meta-tool. The concrete tools — whoami, list-services, redeploy, set-env-vars, add-domain, get-logs, etc. — are the deterministic things you might want to do. The meta-tool, launchmatic-agent, is for fuzzy multi-step requests:

"Why did my last deploy fail?"

"Add a Postgres database and wire it up to the api service."

"Show me the last five deploys, tell me which broke things."

We picked this design deliberately. Every MCP tool definition burns context in your AI client every turn it's connected. A 25-tool surface gets expensive fast. So we ship 11 concrete tools that are cheap-and-deterministic, plus one meta-tool that delegates the multi-step thinking to a server-side agent loop, returning one consolidated answer instead of orchestrating six round-trips client-side.

2. lm agent in the CLI

If you live in your terminal, the same agent is also reachable as a CLI command:

$ lm agent
Launchmatic Agent — type your request, /help for commands, /quit to exit

> deploy this and wait for it to come up

agent: Deploying api (current branch: main, commit a3f1e9). Build queued
as deployment dpl_4kf9. I'll watch for status changes...

  → list-services ✓ array(1)
  → redeploy ✓ object{deploymentId,status,branch}

agent: Build started. Watching for ACTIVE status...

Streams in real time over SSE. -p "..." for one-shot. --json for piping into jq. --thread-id to resume a conversation across CLI sessions. All authenticated against your existing lm login session — no extra setup.

3. Connected Apps + AI Usage dashboards

Two new pages under Settings:

Connected Apps — every client you've ever authorized over OAuth, grouped by app, with the scopes it has, when it last called a tool, and a one-click revoke. If a client misbehaves or you forget which laptop you connected from, you can sever access immediately — every grant gets a fresh 401 on the next request.

AI Usage — tokens consumed this billing period across lm agent, Lightspeed (AI code generation), Auto-fix (post-deploy repair), and image generation. Per-provider breakdown, percent-used bars, daily burn-down chart for the agent. You always know what you're spending.

Under the hood

This shipped as a real OAuth 2.1 Authorization Server that any MCP client can discover via standard .well-known endpoints (RFC 8414 + RFC 9728). PKCE-only public clients are first-class; dynamic client registration (RFC 7591) means a fresh client can self-register on first connect with no admin steps.

The agent loop runs server-side against Anthropic's Claude Sonnet 4.5 with a tool surface that's bounded by the OAuth scopes you granted at consent — write capabilities are opt-in per-scope (services:write, envvars:write, domains:write, databases:write, deployments:write, deployments:rollback). New clients default to read-only.

Per-user daily soft caps and per-team aggregate budgets stop a single runaway session from burning your monthly quota. Every consent grant, token issuance, refresh, and revocation lands in a tamper-evident audit log. Old auth codes and revoked tokens are swept hourly.

What's next

This is foundation. We're already exploring:

• Per-tool granular scopes — say yes to redeploy while saying no to add-domain, even within the same app.
• Workflow-as-tool — let the launchmatic-agent meta-tool run your saved Stirrup AI workflows so an MCP client can trigger them with natural language.
• Read-only audit feed for clients that want to subscribe to your account events.

If you've been waiting to try Claude Code or Cursor with real infrastructure access — start at mcp.launchmatic.io or read the docs.

— The Launchmatic team