Paranoia

Paranoia

Get a cold, adversarial review of your code, plans, and technical decisions from the other frontier coding agent.

Paranoia Local is an MCP server that connects Claude Code and Codex CLI. Install it in one agent and it runs the other as a read-only reviewer, using that reviewer's existing subscription and full repository context.

Claude Code ──MCP──> Paranoia Local ──read-only──> Codex
Codex       ──MCP──> Paranoia Local ──read-only──> Claude Code

Use it to:

• review a branch, diff, or dirty working tree;
• challenge a plan against the code it describes;
• track findings and defect classes until they are closed;
• verify load-bearing plan claims against captured authoritative sources;
• ask a focused repository question;
• rebut a finding in the same reviewer session; or
• arbitrate a decision with both vendors independently.

[Get started](#quickstart) · [Choose a tool](#choose-a-tool) · [Understand tracked reviews](#tracked-reviews) · [Configure](#configuration) · [Full tool reference](docs/tool-reference.md) · [LLM and agent entry point](llms.txt)

Quickstart

1. Install the prerequisites

You need:

• Python 3.11 or later;
• Git 2.36 or later on PATH; and
• the reviewing agent's stable CLI, installed and signed in: Codex CLI 0.144.6 or later, or Claude Code 2.1.197 or later.

Most tools need only the reviewer CLI. `arbitrate` uses both vendors and needs both CLIs.

2. Install Paranoia Local

git clone https://github.com/subvertnormality/paranoia-local
cd paranoia-local
pip install -e .

3. Add it to your coding agent

`--engine` names the agent that performs the review, so it is the opposite of the agent you are configuring.

Claude Code: reviews performed by Codex

claude mcp add paranoia -- paranoia-local --engine codex

Codex: reviews performed by Claude Code

codex mcp add paranoia -- paranoia-local --eng