@agent-inspect/mcp

MCP client tool-call tracing for AgentInspect (list/call lifecycle).

Support level: Supported — see SUPPORT-LEVELS.md.

When to use

• Your agent uses an MCP client and you need local tool-call steps

When not to use

• Running an MCP server (use your server stack)
• Gateway/proxy products

Install

npm install agent-inspect @agent-inspect/mcp

Install matching fixed-group versions (same agent-inspect and @agent-inspect/mcp release). The MCP package keeps agent-inspect external so wrapMcpClient shares the application inspectRun context. A version mismatch can still produce a second runtime and lose nested MCP steps.

Example

import { inspectRun } from "agent-inspect";
import { wrapMcpClient } from "@agent-inspect/mcp";

const traced = wrapMcpClient(mcpClient, { serverName: "tools" });

await inspectRun("my-agent", async () => {
  await traced.listTools?.();
  await traced.callTool({ name: "search", arguments: {} });
});

Call wrapped MCP operations inside inspectRun (or an active inspector run). Outside an active run, step() executes without instrumentation.

wrapMcpClient returns a Proxy over the original client so prototype methods (connect, close, listResources, …) and private SDK state remain available. Do not discard the original client reference solely to work around missing methods.

Privacy

• Local JSONL only; no MCP traffic sent to AgentInspect cloud (there is none)

API

| Export | Purpose | | ------ | ------- | | wrapMcpClient | Trace list/call tool operations |

CLI

npx agent-inspect view · search --kind tool

Docs

• Adapters

Troubleshooting

• Missing list_tools steps: Ensure client goes through wrapped instance

Version

Part of the fixed AgentInspect release line. See the npm badge / package manifest for the current version.

License

MIT