Setup Doctor

What the setup-doctor MCP server does

The setup-doctor MCP server gives an MCP client read-only access to two parts of the setup-doctor command-line application: Doctor and Wrapped. Doctor evaluates how an AI coding agent is configured in a project and in user-level locations. It reports a score from 0 to 100, a quality band, category results, findings, and a specific fix for each finding.

The audit covers instruction files, skills, subagents, MCP servers, plugins, settings, and hooks. It applies 30 rules across six categories, while excluding categories that do not apply to the selected agent. Supported agents include Claude Code, Codex, GitHub Copilot CLI, Cursor, and limited instruction-rule support for other agents using files such as .windsurfrules or .clinerules.

Wrapped takes a different input: local session logs. It summarizes metadata such as session count, active days, token totals, estimated cost, streaks, and a persona label. The resulting shareable card contains aggregate numbers rather than prompts or repository details.

How it works

Run setup-doctor mcp to start the MCP server over standard input and output. An MCP-compatible client can then invoke the exposed Doctor and Wrapped capabilities. The server is designed for local processing and does not make network requests, send telemetry, perform update checks, or require an account, login, or API key.

Session logs are read for metadata only. The application reads timestamps, model names, token counts, and tool names, while message text is discarded line by line. Findings identify rules and locations; secret values are shown as [REDACTED] rather than printed.

Setup and configuration

No package installation is required for the command-line entry point. The documented starting command is npx setup-doctor, and the MCP mode is selected with npx setup-doctor mcp. The project also lists the MCP server in the official MCP Registry under io.github.saxenakapil/setup-doctor.

Doctor, badge, rules, and Wrapped can read a configuration file from .setupdoctorrc in the project. The --config <path> option selects a different file. Configuration supports rule disabling, thresholds, and ignore patterns. Doctor can target project, global, or all locations and can select an agent explicitly with --agent.

Tools and capabilities

• Audit supported agent configuration and return a score, band, category breakdown, and findings.
• Summarize local usage for a selected period, including aggregate session and token information.
• Explain what each rule checks and how to address it through the command-line interface.
• Produce terminal, JSON, or self-contained HTML output for Doctor outside the MCP workflow.
• Apply safe mechanical fixes only through Doctor's explicit fix mode; ordinary audits do not modify configuration.

Limitations and notes

The MCP interface is read-only: Doctor and Wrapped do not edit agent configuration. The separate doctor --fix mode can make safe mechanical changes, but it previews a diff, requires confirmation, and creates a backup first.

Coverage depends on the selected agent. Claude Code receives all 30 Doctor rules; Codex receives instruction and MCP rules; Copilot CLI receives instruction, skills, MCP, settings, and hook checks; Cursor receives instruction, MCP, and project-skill checks. Other agents receive instruction-rule coverage only, and Wrapped is not available for them.

Local HTML and JSON reports may include file paths, so they should be reviewed before sharing. Wrapped cards and badges contain aggregate values only and do not include prompts, file paths, usernames, or repository names. The project is licensed under MIT.