How to inspect an MCP server locally without running its tools

Why inspect an MCP server before trusting it?

MCP makes it easy to connect tools to AI-assisted workflows. That is useful, but it also changes the trust boundary. A server may advertise tools that call internal APIs, read files, list resources or expose prompts that affect how a client behaves.

The safe first step is inspection: look at the protocol messages, negotiated capabilities and advertised catalog before any tool is called. MCP Inspector is built for that step.

Start with an offline transcript

The most conservative path is an offline transcript. Paste or open JSON, JSONL or concatenated JSON-RPC messages and MCP Inspector validates the message sequence against the MCP revision you select or the revision negotiated in the transcript.

The raw transcript stays visible, while the inspection panel turns protocol details into developer-facing sections: initialization, capabilities, advertised tools, resources, prompts and errors. Invalid or suspicious messages are tied back to the line that caused the problem.

Use live discovery only when you ask for it

Live Connection is explicit and read-only. It contacts the endpoint you type, initializes the connection and lists only what the server advertises through its negotiated capabilities.

It does not invoke tools/call, read a resource, get a prompt or start a local process. The point is discovery, not execution. The UI names this path directly and includes the guarantee: no tools will be executed.

What to review in an MCP inspection

For a new or unfamiliar server, review the advertised surface before connecting it to a client that can act on your behalf.

• Capabilities: confirm what the server says it supports and whether that matches your expectations.
• Tools: read names, descriptions and input schemas before a client is allowed to call them.
• Resources: check what content could be listed or read by a capable client.
• Prompts: inspect prompt names and descriptions that may shape model behavior.
• Protocol errors: identify invalid messages, incompatible revisions and malformed catalogs early.

Export sanitized evidence

When you need to share an inspection result with a teammate or vendor, MCP Inspector exports a sanitized transcript with likely credentials redacted. The goal is to keep the raw protocol sequence useful without leaking tokens, authorization values or similar secret-looking content.

That makes it practical to discuss why a server failed validation, which tools it advertised or what changed between revisions without pasting sensitive runtime material into a browser-based utility.