MCP Server

Two ways to run it

Hosted — point your client at https://hoststack.dev/api/mcp with a Bearer API key. Nothing to install.

Local stdio — npx -y @hoststack.dev/mcp runs the same server on your machine and talks JSON-RPC over stdin/stdout. Identical tool surface; no telemetry sent back.

Get an API key

Open /dashboard/settings/api-keys and create a key with full_access permission. The key is bound to a single team, so agents never need to pass teamId — every tool call operates on that team automatically.

Wire it into your client

Claude Desktop

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

{
  "mcpServers": {
    "hoststack": {
      "command": "npx",
      "args": ["-y", "@hoststack.dev/mcp"],
      "env": {
        "HOSTSTACK_API_KEY": "hs_live_..."
      }
    }
  }
}

Cursor

Edit ~/.cursor/mcp.json or use Settings → MCP:

{
  "hoststack": {
    "command": "npx",
    "args": ["-y", "@hoststack.dev/mcp"],
    "env": {
      "HOSTSTACK_API_KEY": "hs_live_..."
    }
  }
}

Claude Code

claude mcp add hoststack --env HOSTSTACK_API_KEY=hs_live_... -- npx -y @hoststack.dev/mcp

Hosted (any client)

POST https://hoststack.dev/api/mcp
Authorization: Bearer hs_live_...
Content-Type: application/json

The endpoint speaks the streamable-HTTP MCP transport. Stateless — every request is independent; no session bookkeeping.

Generate config snippets

The CLI knows the shape each client expects, so you don't have to memorise it:

npx @hoststack.dev/mcp --print-config claude-desktop
npx @hoststack.dev/mcp --print-config cursor
npx @hoststack.dev/mcp --print-config claude-code

If HOSTSTACK_API_KEY is set in your shell, it gets baked into the snippet; otherwise you'll see a placeholder.

Tool inventory

Design notes

• Key-based env-var ops. set_env_var and delete_env_var take a key, not an ID — the MCP looks up the existing var by key first, then patches or deletes.
• Secret values stay masked. list_env_vars returns •••••• for any row stored as a secret. Masking happens on the API server, so secrets never reach the agent.
• Logs are snapshots. get_service_logs and get_deploy_logs return a tail (default 200 lines, max 1000); re-call to get newer entries. For live tails, use the dashboard's logs page.
• One deliberate gap: the project cascade. There is no delete_project — deleting a project takes every service, database, domain and volume under it with it, and that is a confirmation a person should give in the dashboard. Everything one level down an agent can do: delete_service, delete_database, delete_volume and delete_environment are all registered, so an agent asked to tear down what it created can.

Telemetry

When you use the hosted transport, we record one row per tool call: tool name, duration, ok/error, and a non-reversible hash of the input args — never the args themselves. Used for the analytics page and on-call alerts on tool error spikes. Retained 30 days, then deleted by a daily cron.

The local stdio install records nothing on our side. Set HOSTSTACK_MCP_TRACE=1 to emit NDJSON traces to your own stderr for local debugging.

Environment variables