Claude · MCP testing
Test how Claude calls your MCP tools.
Claude decides which tool to call and what to do with the result. In Claude, the model sees images returned via direct content and embedded resources, but not via linked resources. MCPJam runs your server through Claude with anthropic/claude-haiku-4.5 so you see the real tool calls instead of guessing from a bug report.
npx @mcpjam/inspector@latestHow Claude handles your tool results
Image results
Images reach the model via direct content and embedded resources. They do not via linked resources. A tool returning those is invisible to the model here.
Emulation model
MCPJam drives Claude with anthropic/claude-haiku-4.5, so tool selection reflects this host, not a generic model.
Extensions
Claude advertises io.modelcontextprotocol/ui.
See the call
Trace shows every step from prompt to response with timings. The exact tool, the exact arguments.
Score it
Evals run realistic requests through Claude and score whether the right tool was picked with the right arguments.
How we know
Assumed: Inferred defaults, not yet verified against the client. Last reviewed 2026-07-08.
Walkthrough: Claude
Run npx @mcpjam/inspector@latest, connect the MCP server config you already use, and select Claude as the host. MCPJam emulates Claude with anthropic/claude-haiku-4.5 on web, so what you see is this host's behavior, not a generic client's.
Ask for something in natural language and watch which tool Claude picks and with what arguments. Open Trace for the timed step-by-step, then turn the recurring cases into an eval suite so a regression fails the gate instead of your users.
Frequently asked questions
Run npx @mcpjam/inspector@latest, connect your server, and select Claude. MCPJam emulates the host with anthropic/claude-haiku-4.5, so you can watch the real behavior locally, free and open source.
In Claude, images reach the model via direct content and embedded resources, but not via linked resources. (Catalog last reviewed 2026-07-08.)
MCPJam's host-compat catalog (the same data behind caniuse.dev). Assumed: Inferred defaults, not yet verified against the client. Last reviewed 2026-07-08.
See it in Claude before your users do.
One command, your existing server config, Claude's real behavior in the loop.
Open source · no credit card · ~2 min