Writes adversarial test suites for AI-built code. Your agent's test engineer.
One address for any MCP client. You sign in at gauntlet.mcpize.run once, after adding it.
At most: only reads or looks up information.
Odel charges nothing per call. The publisher may charge for its own service.
| Access to | Read | Change |
|---|---|---|
| Mcp | Not stated: mcp:tools, mcp:resources, mcp:prompts | |
From the sign-in at gauntlet.mcpize.run, as the endpoint publishes it. You approve it there.
Odel could not list its tools without signing in.
An MCP server that writes test suites designed to break your code.
This repo is the public front door — install guide, examples, and issue tracking. The generation engine runs as a hosted service, so there is nothing here to build or self-host.
Point your AI agent at a file, and Edgehound writes a complete, runnable test suite that attacks it: boundary values, unicode and emoji, timezone and month-end math, float money rounding, double submits, injection strings, pagination edges. It also predicts — before you run anything — which bugs the suite will expose, with the exact code evidence for each.
Works with Claude Code, Cursor, opencode, VS Code, and any MCP client.
Add the server to your MCP config:
{
"mcpServers": {
"faultline": {
"url": "https://gauntlet.mcpize.run/mcp"
}
}
}
On first use your client will ask you to sign in to MCPize. After that:
"write tests for src/pricing.ts"
and your agent does the rest.
Paste this into your project's rules file (CLAUDE.md, AGENTS.md, or .cursor/rules), and your agent will reach for Edgehound every time it writes or changes input-handling code — no prompting needed:
Whenever you create or modify code that accepts input, touches money, handles
dates, or manages state — call the Edgehound tool (generate_test_suite) before
marking the task complete. Save the returned suite next to the file, run it,
and fix any failing test by fixing the code, not by weakening the test.
Skip only for pure UI markup, copy changes, and configuration.
examples/user-utils.js has four innocent-looking helpers. A real generation produced 60 tests — 55 passed, 5 failed. Every failure was a real bug in the sample:
parseDuration('1h30m') returns 3600 — it silently drops the minutesmonthlyPrice leaks float drift on non-divisible centstruncate blows up when handed a number instead of a stringTry it:
node --test examples/user-utils.test.js
Languages today: JavaScript/TypeScript (Jest, Vitest, node:test), Python (pytest), Go, PHP (PHPUnit), Ruby (RSpec), Rust (cargo test).
Bugs, weird output, missing languages — open an issue here. We read everything and ship fixes fast. Want Go/Rust/PHP improvements or a language we don't cover yet (Java, C#, Swift)? Say so in an issue; demand decides what gets added next.
v1.1 — predicted-bug reports, syntax-validated output with auto-repair, multi-file context, 8 test frameworks.