# falsepass

> falsepass is an open-source MCP server for ai models and llm tooling. Test RL environment graders for wrong answers they pay and right ones they reject. It runs locally on your machine and is started by your AI client.

- **Kind:** MCP server
- **Pricing:** Open source
- **Platforms:** CLI, self-hosted
- **Website:** https://super-eric.github.io/falsepass/
- **Repository:** https://github.com/super-eric/falsepass
- **MCP Registry name:** io.github.Super-Eric/falsepass
- **Categories:** AI models and LLM tooling, developer tools
- **Listing updated:** 2026-11-14

## About

A grader in an RL environment is effectively an assertion, and falsepass tests it like one: it calls the environment's own reward code on completions whose correct verdict is already known and reports every disagreement. Wrong answers that get paid are false passes, and right answers that are rejected are false fails. No model, rollouts or GPU are involved.

Probes include the reference answer with harmless variations; empty, echoed, every-option, hedged and wrong answers; malformed output; and repeats, plus a mutation search and a static scan of the reward source. Each failing probe becomes a stand-alone repro. The README describes one public environment where echoing the question back earned full reward because the grader only checked that the right answer appeared somewhere in the reply.

It works with verifiers (Environments Hub), OpenEnv, TRL reward functions, Inspect AI and HUD, and every audit runs in its own throw-away virtualenv (uv and Python 3.11+ needed). A GitHub Action can fail a build when a known-bad answer earns reward, and the MCP server exposes scan, init, audit, results, repro and discover.

## Features

- **False pass and fail probes:** Feeds graders completions with known verdicts and lists every mismatch.
- **Mutation and static scan:** Searches variations and points at suspicious reward-code lines.
- **Stand-alone repros:** Turns each failing probe into a reproducible script.
- **Framework coverage:** Handles verifiers, OpenEnv, TRL, Inspect AI and HUD environments.
- **CI gate:** A GitHub Action fails builds when a bad answer earns reward.
- **Agent-ready output:** Non-interactive commands with JSON results and exit codes.

## Good for

- Audit an RL environment before training on it
- Check that a reward function does not pay empty or echoed answers
- Gate pull requests that change a grader
- Have a coding agent produce reproduced leads for grader defects

## Install

### Claude Code (Terminal)

```sh
claude mcp add falsepass -- uvx falsepass
```

### Claude Desktop (claude_desktop_config.json)

```json
{
  "mcpServers": {
    "falsepass": {
      "command": "uvx",
      "args": [
        "falsepass"
      ]
    }
  }
}
```

### Cursor (~/.cursor/mcp.json)

```json
{
  "mcpServers": {
    "falsepass": {
      "command": "uvx",
      "args": [
        "falsepass"
      ]
    }
  }
}
```

### VS Code (.vscode/mcp.json)

```json
{
  "servers": {
    "falsepass": {
      "type": "stdio",
      "command": "uvx",
      "args": [
        "falsepass"
      ]
    }
  }
}
```

### Windsurf (~/.codeium/windsurf/mcp_config.json)

```json
{
  "mcpServers": {
    "falsepass": {
      "command": "uvx",
      "args": [
        "falsepass"
      ]
    }
  }
}
```

### Other clients (your client's MCP settings (most use this shape))

```json
{
  "mcpServers": {
    "falsepass": {
      "command": "uvx",
      "args": [
        "falsepass"
      ]
    }
  }
}
```

## Tools

- `scan`: Static scan of reward code.
- `init`: Sets up an environment for auditing.
- `audit`: Probes the grader.
- `results`: Returns headline numbers and failing probes.
- `repro`: Creates a stand-alone repro for a probe.
- `discover`: Finds environments to audit.

## FAQ

### Does it need a model or GPU?

No. It runs the environment's own reward code on known completions.

### What can it test?

verifiers, OpenEnv, TRL reward functions, Inspect AI and HUD graders.

### How do I try it?

Run uvx falsepass demo, which audits a bundled toy environment.

### What are the requirements?

uv on the PATH and Python 3.11 or newer.

### Is falsepass free?

Yes. falsepass is open source, so you can use and self-host it for free.

### How do I add falsepass to Claude Code?

Run this command in your terminal: claude mcp add falsepass -- uvx falsepass

### What are the best alternatives to falsepass?

The closest alternatives to falsepass are REA, Codebase Memory, GitHub, Serena MCP: the IDE for your agent and XERJ.

## Alternatives

- [REA](https://ai-powered.com/mcp/rea): REA is an open-source MCP server for developer tools.
- [Codebase Memory](https://ai-powered.com/mcp/codebase-memory): Codebase Memory is an open-source MCP server for knowledge and memory.
- [GitHub](https://ai-powered.com/mcp/github-2): GitHub is an open-source MCP server for developer tools.
- [Serena MCP: the IDE for your agent](https://ai-powered.com/mcp/serena-mcp-the-ide-for-your-agent): Serena MCP: the IDE for your agent is an open-source MCP server for developer tools.
- [XERJ](https://ai-powered.com/mcp/xerj): XERJ is an open-source MCP server for knowledge and memory.
- [Ouroboros](https://ai-powered.com/mcp/ouroboros): Ouroboros is an open-source MCP server for developer tools.
- [Claude Code Ultimate Guide](https://ai-powered.com/mcp/claude-code-ultimate-guide): Claude Code Ultimate Guide is an open-source MCP server for ai models and llm tooling.
- [Concord MCP](https://ai-powered.com/mcp/concord-mcp): Concord MCP is a freemium MCP server for developer tools.

---
Source: https://ai-powered.com/mcp/falsepass
