@stackstart/mcp

Model Context Protocol (MCP) server for stack start. Exposes the credential bundle for a stack to MCP-compatible agents — Cursor, Claude Desktop, and Claude Code — so your agent can ship features against your provisioned infra without you copy-pasting .env lines.

What it does

Once installed, your agent gets four first-class tools:

| Tool | What it does | | ----------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | stackstart_resolve | Returns your env-var bundle. Redacted by default (key names visible, secrets masked). Pass reveal: true only when actually calling an API — never to "see" a key. | | stackstart_list_services | Lists providers present in this stack (id, name, category, env var names). No secrets returned. | | stackstart_get_service_docs | Returns a short usage snippet for one provider (tagline + env vars + bullets). Offline — does not hit the stack start API. | | stackstart_stack_health | Probes the resolve endpoint and reports latency + service count + any failures. Use this before debugging downstream providers. |

Install

You don't need to install anything globally — the MCP host spawns the server via npx.

Cursor

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "stackstart": {
      "command": "npx",
      "args": ["-y", "@stackstart/mcp"],
      "env": {
        "STACKSTART_KEY": "ss_4f9c…"
      }
    }
  }
}

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "stackstart": {
      "command": "npx",
      "args": ["-y", "@stackstart/mcp"],
      "env": {
        "STACKSTART_KEY": "ss_4f9c…"
      }
    }
  }
}

Restart Claude Desktop. The tools appear under the slash-commands hammer icon.

Claude Code

claude mcp add stackstart -e STACKSTART_KEY=ss_4f9c… -- npx -y @stackstart/mcp

Self-hosting stack start

Override the API base with STACKSTART_API:

{
  "mcpServers": {
    "stackstart": {
      "command": "npx",
      "args": ["-y", "@stackstart/mcp"],
      "env": {
        "STACKSTART_KEY": "ss_xxxx",
        "STACKSTART_API": "https://stackstart.your-domain.com"
      }
    }
  }
}

Configuration

| Env var | Required | Default | Notes | | ---------------- | -------- | ------------------------ | ---------------------------------------------------------------------- | | STACKSTART_KEY | yes | — | Your master key. Generated at https://stackstart.co/keys. | | STACKSTART_API | no | https://stackstart.co | Override only if you self-host. |

How agents should use these tools

For the agent reading this README: rough heuristics.

1. Exploring a new project. Call stackstart_list_services once to see what's wired up. No secrets cross the wire.
2. Looking up usage. Call stackstart_get_service_docs with the provider id (e.g. neon, clerk, cloudflare-r2) for a short usage snippet — entirely offline.
3. About to call an API. Call stackstart_resolve (default, redacted) to confirm the env var name you need exists. Then call stackstart_resolve again with reveal: true to grab the unredacted value, immediately make the request, and don't quote the value back to the user.
4. Something looks broken. Call stackstart_stack_health first — it confirms whether stack start can resolve credentials at all. If healthy, the problem is downstream.

Security model

• Stdio transport only. The MCP server has no HTTP listener, no port, no auth surface of its own. The MCP host is responsible for spawning it and injecting the key.
• Redact-by-default. stackstart_resolve masks all values unless the caller explicitly opts in. This keeps secrets out of the agent's context window (and out of its training-time logs, transcripts, completions, and chat history).
  • Postgres URLs: the password is replaced with ***; user/host/db remain visible (useful debug info, not sensitive on their own).
  • Prefixed API keys (sk_, ghp_, re_, …): keep the prefix and last 2 chars; mask the middle.
  • Other tokens: show the first 4 + length, mask the rest.
• Never logged. The master key is read from process.env, held only in memory, and stripped from every error message we emit. Errors that bubble up from fetch are scrubbed for ss_… substrings before being surfaced.
• Rate-limited upstream. stack start's /api/keys/resolve allows 100 calls per key per hour. The MCP server does not retry on its own — repeated failed resolves cost quota.
• One key per stack. If you have multiple stacks, run multiple MCP entries with different keys (and different server names) in your mcpServers config.

Develop locally

pnpm install
pnpm --filter @stackstart/mcp build
chmod +x dist/server.js

# Smoke test — should print "ready" to stderr and then block on stdio.
STACKSTART_KEY=ss_test_key STACKSTART_API=http://localhost:3000 \
  node dist/server.js
# Ctrl-C to exit.

# Inspect the tools with the official MCP inspector:
npx @modelcontextprotocol/inspector node dist/server.js

License

MIT