Blind Auditor

What sim-xia/blind-auditor MCP server does

sim-xia/blind-auditor MCP server adds a review gate between code generation and final output. An agent submits a draft instead of showing it directly, then receives an instruction to switch into an independent “Ruthless Auditor” role. The agent reviews the draft against a project-specific rulebook and reports its findings back through MCP.

The server is intended to reduce errors and generation bias by making the reviewing step explicit. It does not perform syntax parsing itself; the agent interprets the natural-language rules and evaluates the code. The README states that the approach can be used with any programming language because the audit depends on the agent’s understanding rather than a language parser.

How it works

The workflow has four main stages:

1. The agent calls submit_draft with code content and its programming language.
2. The server locks the session and returns mandatory audit instructions.
3. The agent reviews the draft and calls submit_audit_result with a Boolean verdict, issue list, and score from 0 to 100.
4. The server either unlocks the draft or starts another correction cycle.

A score below 80 is rejected even if the agent reports that the draft passed. Critical findings also cause failure according to the documented workflow. When an audit fails, the retry count increases and the agent must fix and resubmit the code. After the configured retry limit, the workflow can force output with a warning.

The sim-xia/blind-auditor MCP server relies on a prompt-based context switch rather than a second model or separate audit API. It reuses the host IDE’s current model session, so the README states that no additional API key is needed.

Setup and configuration

The project requires Python 3.10 or newer and uv. From the cloned project directory, install dependencies with uv sync, then run the server with uv run blind-auditor. The README also documents uv run python -m src.main as an alternative development launch command.

Create rules.json in the project root to define the audit policy. Its documented settings include a project name, strict-mode flag, maximum retries, and a list of rules. Each rule can specify an identifier, severity, description, and penalty weight. CRITICAL rules are fatal, WARNING rules apply moderate penalties, and PREFERENCE rules represent lower-impact style requirements. The example uses a default passing threshold of 80.

The IDE or MCP client must also receive a system instruction telling the agent not to output code before calling the audit tool. The README provides configuration examples for Claude Desktop, Cursor, and Windsurf. The MCP entry runs uv with run, a project directory, and the blind-auditor command; the directory value must be replaced with the local clone path.

Tools and capabilities

The sim-xia/blind-auditor MCP server exposes three documented tools:

• submit_draft: accepts code and language, locks the session, and returns audit instructions.
• submit_audit_result: accepts passed, issues, and score; it releases or rejects the draft and manages retries.
• reset_session: clears session state and resets the retry count.

Its capabilities include configurable team rules, severity-weighted scoring, a minimum score check, retry-based remediation, and session reset. The project is released under the MIT License. The README describes it as zero-cost and does not document any required credentials or paid service dependency.