Deploy Apps From Your AI Coding Agent via MCP (Cursor, VS Code, Claude Code)

·4 min read·Updated
mcpai-agentsgitautomationhow-to

JustRunMy.App runs its own MCP server at /api/mcp, so any MCP-capable AI agent — Cursor, VS Code, Claude Code, or a custom one — can operate your apps directly: list them, read logs and status, change env vars, ports, and volumes, start/stop/restart, run one-off commands in a live container, and get a ready-to-run git command that auto-deploys a new or existing app. You authorize the agent with a short-lived, revocable token generated from the dashboard — not your account password or a permanent API key.

What your agent can actually do

The MCP server exposes the same operations as the dashboard, as callable tools:

Category Tools
App lifecycle list apps, rename, delete, start/restart, stop
Runtime config get/set env vars, get/set ports, get/add/remove volumes, set CPU/RAM quotas, set startup command
Image point the app at a registry image, or at an image already built on the platform
Operate get status (including live endpoints), get logs, run a command inside the running container
Git deploy get a git push URL + ready-to-run push command for a new app, or for an existing one

One current gap worth knowing about: there's no single "create app" tool call today (it's implemented but disabled behind a known bug). In practice that doesn't block agent-driven deploys — the git deploy tool is the actual path an agent uses to stand up a brand-new app, covered below.

Why a first-party MCP server beats a generic GitHub MCP + shell scripts

A generic Git-hosting MCP server gives an agent files and commits — it doesn't know what a "port," a "quota," or a "volume" means on your specific platform. JustRunMy.App's MCP server talks in the platform's own domain model, so an agent can reason about — and change — the actual running configuration, not just the source.

It's also scoped auth, not a static credential: the token an agent uses expires on its own (1 day to 1 month, your choice) instead of living forever in a script or .env file until someone remembers to revoke it.

Step 1: Get your MCP config

Open MCP Server Config in the dashboard sidebar (/panel/mcp-tool-config). It generates an mcp.json pointing at your instance's /api/mcp endpoint with a fresh identity token attached, and gives you:

  • A one-click Add to Cursor button.
  • A one-click Add to VS Code button.
  • An "Advanced" panel to pick the token's lifetime (1 day, 1 week, or 1 month) and copy or download the raw config for any other MCP client.

Step 2: Wire it into your agent

For Cursor or VS Code, the one-click buttons handle it. For Claude Code or any other client that speaks the MCP streamable-HTTP transport, add the downloaded config's entry to your .mcp.json:

{
  "mcpServers": {
    "justrunmy.app": {
      "url": "https://<your-instance>/api/mcp",
      "headers": {
        "X-User-Identity": "<token from the dashboard>"
      }
    }
  }
}

Treat that token as a credential — it decrypts to your account identity. It's short-lived by design, but don't commit it to a repo.

Step 3: Deploy something with a prompt

With the server connected, plain-language requests map onto the tools above:

  • "Push this repo to JustRunMy.App as a new app." The agent asks the git deploy tool for credentials, gets back a push URL and an exact shell command, and runs it. The platform builds the image (using a Dockerfile at your repo root, if present) and creates the app from that first push — the same outcome as using Deploy from Git in the dashboard, just without leaving your editor.
  • "Set DATABASE_URL on app 42 and restart it." The agent updates the environment variable and restarts the app.
  • "Why did my last deploy fail?" The agent pulls logs and status to help you debug without you opening the dashboard.

What it can't do yet

  • One-call app creation. There's no jrnm_create_app-equivalent tool live right now — use the git-push path above, which creates the app as a side effect of the first deploy.
  • Anything outside this tool surface. Billing, plan changes, and anything not listed above still go through the dashboard.

FAQ

What is MCP?

The Model Context Protocol is an open standard for connecting AI agents to external tools and systems over a consistent API, so an agent can call real operations instead of just reading text.

Do I need to install anything to use this?

No — it's a remote HTTP MCP server. Add the URL and header from your dashboard's MCP Server Config page to any MCP-capable client.

Is the token a permanent API key?

No. It's short-lived (1 day to 1 month, your choice) and simply expires — generate a fresh one from the dashboard whenever you need to.

Can my agent create a brand-new app in a single step?

Not via a dedicated "create app" tool today. Ask it to push your repository instead — the platform creates the app automatically from that first push.

Which AI agents work with this?

Any MCP client that supports the streamable HTTP transport with custom headers. Cursor and VS Code have one-click install buttons in the dashboard; others, including Claude Code, just need the same URL and header pasted into their MCP config.

Wrapping up

The difference between "an AI agent that can edit files and run git" and "an AI agent that can actually operate your hosting platform" is whether it understands your platform's own model — ports, volumes, quotas, live container state — not just your source tree. Point an MCP-capable agent at /api/mcp with a scoped token, and prompts like "deploy this" or "why is this app down" turn into real tool calls instead of instructions you have to carry out yourself in a dashboard.