Skip to main content
The Builder MCP Server lets you build and run your own software on Mihu — described in plain language, from your own AI client (Claude Desktop, Claude Code, Cursor, or any MCP client). It is the customer-facing integration and app builder, delivered over the Model Context Protocol. Ask it for a CRM, a connector to another service, a data migration, a webhook handler, or an automation agent, and it generates the code, reviews and tests it, deploys it to an isolated cloud sandbox, and hands you a live HTTPS URL. This is different from the Mihu MCP Server: that one operates your existing agents, contacts, and campaigns; the Builder MCP creates and runs brand-new apps for you.

What the Builder MCP can do

Every app runs in its own isolated sandbox with a public preview URL. Each connection is bound to one tenant and scoped to it — no other tenant is reachable. The table above is a summary, not the full list. The complete, always-current set of tools — with each tool’s parameters, types, and description — is advertised by the server itself.

See the full tool list

You don’t maintain this list by hand — MCP clients discover it automatically on connect:
  • Claude Desktop / Claude Code / Cursor: once the server is registered and connected, open the tools/connectors view (Claude Desktop: the 🔌 / tools icon in the message box; Claude Code: /mcp). Every Builder MCP tool is listed there with its description.
  • Any MCP client: the server answers the standard MCP tools/list request — that response is the authoritative list, including each tool’s JSON input schema.
  • From a prompt: just ask your client “list the tools available from mihu-builder” — it reads them from the connected server.
Each tool’s description includes its required fields, enums, and example shapes, readable by the model at call-time — no external tool reference to keep in sync.

Get your connection details

You get everything you need from the app. Open Builder Mode and go to the Credentials tab — the Builder connection panel there shows your three values:
The token is created automatically the first time you use the builder. You can roll it any time from the Credentials tab (Recreate token), or delete it to revoke access — a new one is minted automatically the next time a builder is used. Keep it secret: it authorizes everything for your tenant.

Before you start

Make sure you have:
  • access to your Mihu tenant,
  • an MCP-compatible client (Claude Desktop, Claude Code, Cursor, or a custom client),
  • Python with the MCP package installed: pip install mcp,
  • and your Server, Tenant, and Token from Builder Mode → Credentials.

Run the server (stdio)

Run mcp_server.py with your three values as environment variables:
For example:

Connect from Claude

Pick whichever fits you. Both end up giving Claude the builder tools.
Don’t want to run a local script? The core builder tools (create_build, answer_build, approve_build, rebuild_build, list_builders, deploy_builder_code, …) are also bridged into the hosted Mihu MCP at mcp.mihu.ai. If you connect that one — see the Mihu MCP Server guide — you get the builder tools alongside everything else, with no local setup. Use the standalone Builder MCP below when you want the full, dedicated builder toolset.

Claude Desktop

  1. Open Settings → Developer → Edit Config.
  2. Add the mihu-builder server below, filling in your Tenant and Token from Builder Mode → Credentials and the absolute path to mcp_server.py.
  3. Save and restart Claude Desktop. The builder tools then appear under the tools/🔌 icon.

Claude Code

Register it from the terminal with the same three values:
Then run /mcp inside Claude Code to confirm mihu-builder is connected and see its tools.

Build flow (asynchronous — poll after each write)

Building is asynchronous, so always poll after a write:
  1. Describe it — call create_agent with a one-sentence prompt.
  2. Poll — call get_agent (~every 2–3s) while status: processing.
  3. Answer questions — while status: gathering, the builder asks what it needs (credentials, size, schedule); call answer_agent and keep polling.
  4. Approve — when status: awaiting_approval, call approve_agent to build, deploy, and run automated checks (AI code review + tests).
  5. Live — once status: deployed, the app is running with a preview URL. Operate it with start_deployment / stop_deployment / get_logs / exec_command, or ask for changes (the builder rebuilds the live app).
You can also skip the conversation and ship your own files directly with deploy_code.

Test the connection

Try these once connected:
  • “List my builders.”
  • “Build a small JSON API that stores and returns leads.”
  • “Show the logs for my last deployment.”

Troubleshooting

Support