Blog / Test MCP servers with agents

See exactly what your agent sends to the model: wire-capturing Claude Code, Codex and Grok

Most of what we know about how agent apps load MCP tools came from looking at the actual requests. Here's the method, minus the scripts, and the traps that change what you see.

Leonid Bugaev · Published 21 Oct 2026

Why capture the wire

Agent apps change fast and their docs lag. The only reliable way to know what reaches the model is to look at the requests. That's how we found that Claude Code sends no MCP definitions up front, that Codex writes JavaScript to find tools, and that a lab flag had silently turned tool search off. See Claude Code tool search, seen from the wire.

The method

1. Start an HTTPS proxy (we use mitmproxy) on a local port, with a fresh CA in a temporary config folder.
2. Trust the CA only in the child process: set HTTPS_PROXY plus the app's CA variable (NODE_EXTRA_CA_CERTS for Node apps, SSL_CERT_FILE or the app's own setting for others) in that process's environment only.
3. Run the app normally, with its usual login and endpoint, against your MCP server.
4. Delete the CA when the process exits (a shell trap), so nothing stays trusted.
5. Tokenize the saved bodies with a fixed tokenizer (we use o200k_base) and keep each provider's own usage for per-request input.

Redaction

• Strip Authorization, x-api-key and cookie headers.
• Mask token patterns in bodies: API-key prefixes, JWTs, Bearer strings.
• Grep excerpts again before sharing. Keep raw dumps local and out of git.
• For MCP traffic, record JSON-RPC bodies only, never headers.

What the first requests contained

Same prompt and 99-tool server; all three called the right tool. Totals for the task: about 76k (Claude Code), 137k (Codex), 158k (Grok Build) input tokens.

Traps

• Base-URL proxies change Claude Code. With a custom ANTHROPIC_BASE_URL, first-request input jumped to 143,357 tokens with tool search off; setting ENABLE_TOOL_SEARCH=true brought it back to 22,446. See Claude Code can't see your MCP tools?
• Lab flags change defaults. --tools "" also removed ToolSearch: 110,998 tokens, and the env var couldn't restore it. Check your lab command against a default run.
• Untrusted folders. Grok Build dropped a project MCP server silently in headless mode until the folder was trusted.
• Async connections. Grok connects servers asynchronously; in about 1 of 8 preflight runs the first request didn't yet list the server.

Go deeper: How Codex finds MCP tools, How agents load MCP tools, Test against production read-only.

Back to testing MCP servers with agents.

How we measured

• Captures: one prompt per app against a local 99-tool stub, 16 Oct 2026, normal logins.
• Config matrix: Claude Code 2.1.301, “say hi” with the 99-tool mock, first-request input tokens, single runs per config.
• We don't publish our capture scripts; this page describes the method only.