MCP server

Connect in three steps

# Claude Code
claude mcp add --transport http mushi "https://kensaur.us/mushi-mushi/hosted-mcp/"
 
# Cursor (writes .cursor/mcp.json)
npx mushi-mushi setup --ide cursor

The rest of this page is detail: how sign-in works, manual config files, the hosted transport for headless agents, and the tool list.

How sign-in works (OAuth, no keys to copy)

The hosted MCP endpoint speaks the standard MCP OAuth flow (authorization code + PKCE), so clients that support MCP login connect with zero manual credentials:

# Claude Code
claude mcp add --transport http mushi "https://kensaur.us/mushi-mushi/hosted-mcp/"
claude mcp login mushi   # or run /mcp inside a session and pick "mushi"

Your browser opens the console consent page. Sign in, choose what the connection reaches, and click Approve:

• An account key does in each project only what your role there allows. The agent calls list_projects and passes projectId to project tools.
• To keep one connection on a single project, add ?project=<project-id> to the server URL, for example https://kensaur.us/mushi-mushi/hosted-mcp/?project=<project-id>.
• A pinned connection, or one limited to picked projects, leaves out the organization-wide tools (portfolio, connectors, spend, accounts register).
• Cursor and VS Code follow the same flow when you add the hosted URL as a remote MCP server. Do not set an Authorization header if you want the login flow: a static header tells the client that OAuth isn’t needed.

Use the kensaur.us/mushi-mushi/hosted-mcp/ URL for the login flow. Clients look for the OAuth metadata at the root of the server’s host (/.well-known/oauth-protected-resource/…), and only that host serves it; on the supabase.co URL the lookup fails with an HTTP 404. With an API key header (below), either URL works.

Self-hosted on your own <ref>.supabase.co? Point Claude Code at the function’s own metadata document with its authServerMetadataUrl override (npx mushi-mushi setup --ide claude writes this for you):

{
  "mcpServers": {
    "mushi": {
      "type": "http",
      "url": "https://<ref>.supabase.co/functions/v1/mcp",
      "oauth": {
        "authServerMetadataUrl": "https://<ref>.supabase.co/functions/v1/mcp/.well-known/oauth-authorization-server"
      }
    }
  }
}

One-liner setup (recommended)

npx mushi-mushi setup --ide cursor
# or: npx mushi-mushi setup --ide claude

For Cursor and Claude Code this writes the hosted MCP URL with OAuth login by default. No API key lands in your repo, the entry is safe to commit and share with your team, and each teammate signs in from the IDE on first use. Pass --stdio (or --ci in headless environments) for the previous local subprocess entry with a key.

It reads CLI config at ~/.config/mushi/config.json, or %APPDATA%\mushi\config.json on Windows (full order; written by mushi login; legacy ~/.mushirc auto-migrates on first load), detects the IDE config path, and writes the mcpServers block for you. Restart the IDE and ask the agent: “list mushi tools”.

Claude Code plugin

One command installs the hosted MCP hookup plus a mushi-debugger subagent (report triage → fix context → implement → close the loop):

/plugin marketplace add kensaurus/mushi-mushi
/plugin install mushi-debugger@mushi-mushi

Then: /mcp → pick mushi → sign in via the browser. Ask Claude to “use mushi-debugger to fix the top report”.

Hosted MCP

The hosted server speaks Streamable HTTP at https://kensaur.us/mushi-mushi/hosted-mcp/. The server card at https://kensaur.us/mushi-mushi/hosted-mcp/.well-known/mcp/server-card.json lists every tool for directory scanners.

Manual setup: Cursor

Add to .cursor/mcp.json in your repo root:

{
  "mcpServers": {
    "mushi": {
      "command": "npx",
      "args": ["-y", "@mushi-mushi/mcp@0.30.7"],
      "env": {
        "MUSHI_PROJECT_ID": "YOUR_PROJECT_ID",
        "MUSHI_API_KEY": "mushi_..."
      }
    }
  }
}

Self-hosting? Add one more line to env, pointing at your own server: "MUSHI_API_ENDPOINT": "https://<your-ref>.supabase.co/functions/v1/api". On Mushi Mushi Cloud, leave it out.

Manual setup: Claude Desktop

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

{
  "mcpServers": {
    "mushi": {
      "command": "npx",
      "args": ["-y", "@mushi-mushi/mcp@0.30.7"],
      "env": {
        "MUSHI_PROJECT_ID": "YOUR_PROJECT_ID",
        "MUSHI_API_KEY": "mushi_..."
      }
    }
  }
}

As above, add MUSHI_API_ENDPOINT only when you self-host.

Hosted Streamable HTTP transport (no subprocess)

Skip the local subprocess: use the hosted MCP endpoint when you’re running OpenAI Agents SDK, ChatGPT Agent, CrewAI, or any remote orchestrator. Headless orchestrators can’t drive the browser OAuth flow, so this is the one place a static API key header is the right tool:

{
  "mcpServers": {
    "mushi-hosted": {
      "url": "https://kensaur.us/mushi-mushi/hosted-mcp/",
      "headers": {
        "X-Mushi-Api-Key": "mushi_…",
        "X-Mushi-Project-Id": "YOUR_PROJECT_ID"
      }
    }
  }
}

Available tools

Cursor rules hook (optional)

After setup, run:

npx mushi-mushi setup --ide cursor --with-rules

This writes a .cursorrules file that tells Cursor to call get_fix_context before writing a fix and submit_fix_result after, closing the loop from inside the editor. Download the template: cursor.cursorrules

Troubleshooting