MCP Plugin

A Model Context Protocol server that gives Claude structured skeptical review tools. Confidence history, auto-review mode, and export — all as MCP tools that Claude can call directly.

Features

Structured Review

5 skeptic personas analyze AI output through a structured prompt with confidence ratings and specific concerns.

Confidence History

Scores tracked over time in a local JSONL file. See trends, weakest dimensions, and strongest areas.

Auto-Review Mode

Toggle on to get a quick confidence score appended to every AI response automatically.

Export Reviews

Save any review to your Desktop as markdown or HTML for sharing, archiving, or further analysis.

Installation

Download

curl -LO https://checkie.ai/skill/checkie-mcp.tar.gz
mkdir -p ~/checkie-mcp
tar xzf checkie-mcp.tar.gz -C ~/checkie-mcp
cd ~/checkie-mcp && npm install

Configure for Claude Desktop

Add to your Claude Desktop config file (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "checkie": {
      "command": "node",
      "args": ["/path/to/checkie-mcp/index.js"]
    }
  }
}

Configure for Claude Code

Add to your project's .mcp.json or global MCP config:

{
  "mcpServers": {
    "checkie": {
      "command": "node",
      "args": ["/path/to/checkie-mcp/index.js"]
    }
  }
}

After configuring, restart Claude Desktop or Claude Code. The Checkie tools will be available immediately.

MCP Tools

Tool	Description
`checkie_review`	Run a skeptical review of AI-generated text through 5 personas. Supports full, quick, critical, compare, and ask modes.
`checkie_save_review`	Save review scores to confidence history for trend tracking.
`checkie_get_history`	Retrieve past review scores, averages, weakest/strongest dimensions, and recent review table.
`checkie_auto_toggle`	Enable or disable auto-review mode (quick confidence score after every response).
`checkie_auto_status`	Check whether auto-review mode is currently enabled.
`checkie_export_review`	Export a review to your Desktop as markdown or HTML file.

checkie_review

The main tool. Runs AI text through 5 skeptic personas and returns structured review instructions.

Parameter	Type	Description
`text`	string (required)	The AI response text to review
`mode`	enum (optional)	full, quick, critical, compare, or ask. Default: full
`focus`	string (optional)	Focus area or question (for ask mode or narrowing scope)
`alternative`	string (optional)	Alternative response for compare mode

checkie_save_review

Saves review scores to the local confidence history log.

Parameter	Type	Description
`composite_score`	number	Overall score (1.0-5.0)
`logic`	number	Logic & Reasoning score (1-5)
`evidence`	number	Evidence & Accuracy score (1-5)
`completeness`	number	Completeness score (1-5)
`bias`	number	Bias & Framing score (1-5)
`practicality`	number	Practicality score (1-5)
`topic`	string	Brief topic summary
`mode`	string (optional)	Review mode used

checkie_get_history

Returns review history with averages, trends, and a table of recent reviews.

Parameter	Type	Description
`limit`	number (optional)	Max recent reviews to return. Default: 20

checkie_export_review

Exports a review to a timestamped file on your Desktop.

Parameter	Type	Description
`content`	string	Full review content in markdown format
`format`	enum (optional)	md or html. Default: md

Review Modes

Mode	Description
Full	All personas with detailed analysis, ratings, and actionable suggestions
Quick	Condensed single-table summary with one-line verdicts
Critical	Only shows personas that rated 3 stars or below
Compare	Side-by-side comparison of two responses across all 5 dimensions
Ask	5 personas answer a question independently, then consensus and disagreements

Data Storage

All data is stored locally:

Confidence history — ~/.claude/skills/checkie/.confidence_history.jsonl
Auto-review state — ~/.claude/skills/checkie/.auto_enabled
Exported reviews — ~/Desktop/checkie-review-{timestamp}.md|html

Requirements

Node.js 18+
Claude Desktop or Claude Code (or any MCP-compatible client)
No external API keys needed — the MCP server generates structured prompts that Claude executes