MCP Inspector as MCP Server

A lean MCP server that enables LLMs to inspect and test other MCP servers. This is a self-contained implementation that uses the MCP SDK directly, without shelling out to external CLIs.

Features

• Direct SDK integration: Uses @modelcontextprotocol/sdk directly for both server and client operations
• All transport types: Supports stdio, SSE, and HTTP (streamable) transports
• Minimal footprint: Single dependency (@modelcontextprotocol/sdk)
• Full MCP inspection: List tools, call tools, list resources, read resources, list prompts, get prompts
• Session management: Persistent connections with automatic garbage collection
• Event buffering: Capture notifications, traffic, and errors for debugging

Installation

npm install
npm run build

Usage

As an MCP Server

Add to your MCP config. While there are slight variances between different harnesses, the general format is the same:

{
  "mcpServers": {
    "mcp-inspector": {
      "command": "node",
      "args": ["/path/to/mcp-inspector-as-mcp-server/dist/server.js"]
    }
  }
}

Available Tools

Session Management (NEW in v2.0)

Inspection Tools

Connection Parameters

All tools accept the following connection parameters:

For stdio transport (local commands):

• command: Command to run (e.g., "node", "python")
• args: Array of arguments (e.g., ["path/to/server.js"])

For SSE/HTTP transport (remote servers):

• url: Server URL (e.g., "http://localhost:3000/sse")
• headers: Optional HTTP headers object

Common:

• transport: Force transport type ("stdio", "sse", or "http"). Auto-detected if not specified.
• session_id: (Optional) Use an existing persistent session instead of creating an ephemeral connection.

Session Workflow

For debugging stateful server behavior, use persistent sessions:

1. insp_connect → returns session_id
2. insp_tools_list (with session_id) → uses persistent connection
3. insp_tools_call (with session_id) → state is preserved
4. insp_read_events (with session_id) → see notifications
5. insp_disconnect (with session_id) → cleanup

Sessions auto-close after 30 minutes of inactivity.

Human Steering & Observability

The inspector enables human-in-the-loop workflows where you can observe and guide LLM-driven MCP testing in real-time.

How It Works

┌─────────────┐     MCP calls      ┌─────────────────┐     forwards     ┌─────────────┐
│   LLM Agent │ ◄────────────────► │  MCP Inspector  │ ◄──────────────► │  Target MCP │
│  (Antigravity)                   │    (v2.0)       │                  │   Server    │
└─────────────┘                    └────────┬────────┘                  └─────────────┘
                                            │
                                   Events logged to
                                   session EventBuffer
                                            │
                    ┌───────────────────────┼───────────────────────┐
                    │                       │                       │
                    ▼                       ▼                       ▼
            insp_read_events         HTTP :9847/api          mcp-steer CLI
            (LLM reads events)       (external access)       (human injection)

Viewing Activity

Via LLM: The agent can call insp_read_events to see what's happening:

{
  "session_id": "sess_abc123",
  "types": ["traffic_in", "traffic_out"],
  "limit": 20
}

Via HTTP: Query the steering API directly:

curl http://127.0.0.1:9847/api/sessions

Steering the Agent

Inject guidance messages that appear in the LLM's next tool response.

Using the CLI:

./bin/mcp-steer.mjs "Focus on testing the error handling paths"
./bin/mcp-steer.mjs --session sess_abc123 "Try calling with invalid params"

Using HTTP:

curl -X POST http://127.0.0.1:9847/api/steer \
  -H "Content-Type: application/json" \
  -d '{"message": "Check the authentication flow next"}'

Using the MCP tool:

{
  "tool": "insp_inject_steering",
  "arguments": {
    "session_id": "sess_abc123",
    "message": "Great progress! Now test edge cases."
  }
}

Event Types

Typical Workflow

1. LLM creates session: insp_connect → gets sess_abc123
2. LLM starts testing: insp_tools_call with session_id
3. Human observes: curl http://127.0.0.1:9847/api/sessions
4. Human steers: ./bin/mcp-steer.mjs "Also test the batch endpoint"
5. LLM receives steering: Next tool response includes ⚡ STEERING from human: ...
6. LLM adapts: Takes the human guidance into account

Examples

List tools from a local MCP server (ephemeral):

{
  "command": "node",
  "args": ["/path/to/some-mcp-server/dist/server.js"]
}

Create a persistent session:

{
  "command": "node",
  "args": ["/path/to/some-mcp-server/dist/server.js"]
}
// Returns: { "session_id": "sess_abc123", "server_info": {...} }

Call a tool using a session:

{
  "session_id": "sess_abc123",
  "tool_name": "search",
  "tool_args": {"query": "hello"}
}

Architecture

├── src/
│   ├── server.ts     # MCP server exposing inspector tools
│   ├── client.ts     # Client wrapper (hybrid stateless/session mode)
│   ├── transport.ts  # Transport factory (stdio, SSE, HTTP) + TracingWrapper
│   ├── session.ts    # SessionRegistry with GC (30-min TTL)
│   └── events.ts     # EventBuffer (ring buffer for notifications)
├── bin/
│   └── mcp-steer.mjs # CLI tool for human steering
└── tests/            # Ad-hoc test scripts (run with npx tsx)

Why This Exists

The original MCP Inspector is a web-based UI + CLI combo spread across multiple projects. This consolidates the core functionality into a single, lean MCP server that an LLM can use to:

1. Develop and debug MCP servers iteratively
2. Test MCP server functionality without leaving the conversation
3. Explore what tools/resources/prompts an MCP server exposes
4. Debug stateful behavior with persistent sessions

Changelog

v2.1.0

• Added human steering (insp_inject_steering) for human-in-the-loop workflows
• Added HTTP API on port 9847 for external steering/observability
• Added mcp-steer.mjs CLI tool for easy human interaction
• Fixed critical bug in TracingTransportWrapper where handler capture timing caused message loss

v2.0.0

• Added session management (insp_connect, insp_disconnect, insp_list_sessions)
• Added event buffering (insp_read_events)
• All inspection tools now support optional session_id for persistent connections
• Added automatic garbage collection (30-minute TTL for idle sessions)
• Backward compatible: omit session_id for original ephemeral behavior

v1.0.0

• Initial release with ephemeral connections

License

MIT