Model Context Protocol — Inspector
MCP Server Tester
Connect to any MCP server and see what it's actually telling your agent — every tool, every token, every red flag — before you wire it into production.
What it checks
Five things every MCP integration should know before it ships
Tools, resources, and prompts — as the server actually sends them
Connects over Streamable HTTP or SSE (with automatic fallback), lists every primitive the server advertises, and keeps a full JSON-RPC history — initialize, tools/list, tools/call, and everything between — so you can see exactly what your agent will see.
Token cost and definition quality, before a single query runs
Estimate token usage with a generic heuristic or exact counts from the Claude API and tiktoken (o200k_base / cl100k_base), then grade the tool set A–F on description quality, type coverage, and schema specificity.
Tool poisoning, rug pulls, and prompt-injection attempts
Heuristics scan every name, description, and schema for hidden Unicode, exfiltration hints, and manipulative phrasing. A per-tool hash is pinned per server, so if a tool's definition silently changes after you've approved it, you get a line-by-line diff and have to explicitly re-confirm — nothing updates behind your back. An optional Claude API deep scan catches paraphrased intent heuristics miss.
Every auth flow an MCP server actually uses
None, Bearer, OAuth2 Client Credentials, Custom Header, and full interactive SSO with PKCE — including OAuth endpoint auto-discovery and Dynamic Client Registration (RFC 7591), so most servers need zero manual configuration. The Auth Inspector shows exactly which headers went out and decodes the resulting JWT.
Two servers, side by side — and a report you can hand off
Run the same inspection against two endpoints at once to diff latency, token cost, and quality scores. Export a self-contained Markdown or JSON report with secrets redacted, ready to paste into a PR or a Slack thread.
Exhibit A
A deliberately hostile server, caught on first connect
Grading
Every server gets a letter grade, not a vague warning
The same A–F scale is used for both the LLM Readiness Score (are the tool definitions clear enough for an agent to use correctly?) and the Tool Poisoning Risk score (does anything look designed to manipulate the calling agent?).
- Tool descriptions20%
- Param descriptions25%
- Type definitions25%
- Required annotation15%
- Schema specificity15%
Get started
Running locally takes one command
pip install remote-mcp-server-testermcp-tester# → http://localhost:8080No Claude API key required — Generic and tiktoken token counting, plus the heuristic security scan, work fully offline. Or skip installation entirely and try the hosted demo (first request after idle takes ~50s to wake).