# Brutalist MCP [Health: Active]

**Category:** 🤖 Coding Agents  
**Repository:** https://github.com/ejmockler/brutalist-mcp  
**GitHub Stars:** 14  
**npm Downloads (last month):** 740  
**Views:** 1  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/brutalist-mcp

## Description
Deploy Claude Code, Codex & Antigravity (agy) CLI agents to demolish your work before users do. Real file analysis. Brutal honesty. Per-call model pinning, conversation continuation & intelligent pagination.

## Tools
Capabilities this server exposes over MCP:

- **roast** — Review a path and context using a selected domain.
- **roast_cli_debate** — Have two critics defend explicit opposing positions.
- **cli_agent_roster** — Report detected critics and model configuration.
- **brutalist_discover** — Suggest domains from an `intent` string.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "brutalist-mcp": {
    "command": "npx",
    "args": ["-y","@brutalist/mcp@latest"],
    "env": {
      "AGY_BIN": "",
      "BRUTALIST_MCP_SERVERS": "",
      "BRUTALIST_CODEX_ALLOW_MODEL_OVERRIDE": ""
    }
  }
}
```

**Requires environment variables:** `AGY_BIN`, `BRUTALIST_MCP_SERVERS`, `BRUTALIST_CODEX_ALLOW_MODEL_OVERRIDE` — the values above are empty placeholders; fill in real credentials before running (see the repository for what each one is for).

## Documentation

## What Brutalist MCP MCP server does

Brutalist MCP runs review workflows through authenticated CLI agents already installed on the local machine. It supports Claude Code, Codex, and Antigravity (`agy`). A roast asks one or more critics to examine a target using a selected domain and returns their responses together. A debate gives two critics opposing positions and carries the discussion through one to three rounds.

The Brutalist MCP MCP server covers codebases, file structure, dependencies, Git history, test coverage, ideas, architecture, research, security, product decisions, infrastructure, design, and legal material. Reviews can use a repository path plus focused context. For design reviews, a URL can provide a concrete interface to inspect through the registered Playwright MCP server.

## How it works

The MCP process launches the selected local CLI subprocesses. Each CLI uses its own configured service credentials and model settings. If no critic list is supplied, all detected critics run; otherwise, the request can select a subset. Claude and Antigravity accept per-call model choices. Codex uses its own configuration unless the server’s model-override setting is enabled.

Critics run in parallel for a roast. The debate tool requires at least two detected CLIs, assigns the supplied positions, and runs multiple turns. `cli_agent_roster` reports detected command-line agents and model configuration, while `brutalist_discover` suggests review domains from an intent string. Large responses support pagination and configurable character chunk targets.

## Setup and configuration

Use Node.js 24 and authenticate at least one supported CLI before starting the server. Node.js 20 is covered by the project’s CI tests. Install and sign in to Claude Code, Codex, or Antigravity using their official instructions, then register the package as a local stdio MCP server with `npx -y @brutalist/mcp@latest`.

The Brutalist MCP MCP server runs on macOS and Linux. The Antigravity adapter requires `python3` and a POSIX environment; native Windows is unsupported, so Windows users should use WSL. Set `AGY_BIN` when the Antigravity executable is not found at the preferred location. Pin the npm package instead of using `@latest` when repeatable deployment is needed.

Additional MCP servers for a roast can be supplied as JSON through `BRUTALIST_MCP_SERVERS`. The `design` domain automatically requests Playwright. Requested integrations replace Codex’s configured MCP server set for that run, and Antigravity currently ignores this field.

## Tools and capabilities

- `roast`: review a target with a selected domain, context, critics, and optional model settings.
- `roast_cli_debate`: compare explicit pro and con positions across multiple critic rounds.
- `cli_agent_roster`: inspect detected critics and model configuration.
- `brutalist_discover`: suggest domains from an intent string.
- Domain-specific arguments for items such as coverage execution, dependency inclusion, architecture constraints, security threat models, and design URLs.
- Paginated roast and debate responses for large outputs.

## Limitations and notes

Findings are model output and should be checked against cited files, commands, and sources. The target path is included in the critic prompt but does not change the subprocess working directory, so use absolute paths and describe the evidence to inspect in `context`.

The Codex default MCP timeout may be too short for long reviews; configure a larger timeout in its TOML settings. Debates can take longer because they use multiple turns and may retry. A detected CLI is not proof that authentication, quota, or a requested model will work. Legacy names such as `roast_codebase` and `roast_security` are not registered tools; use `roast` with a domain instead.

_Full upstream README: https://allmcps.com/mcp/brutalist-mcp/readme_