Wikipedia & MediaWiki revision-history MCP server

refract mcp starts a stdio Model Context Protocol server that lets AI agents analyze page revision histories from Wikipedia or any MediaWiki API.

The MCP server supports sampling — it can request the host's LLM to interpret events at any BYO-inference boundary without managing API keys.

Quick start

refract mcp

The server starts and waits for MCP client connections on stdio (wikihistory mcp also works). Connect any MCP-compatible client.

Connecting AI coding agents

The server is the mcp command of the @refract-org/cli package, so that is the package a client configuration names. @refract-org/mcp is a library with no executable, and the unscoped npm package refract is an unrelated project — npx refract mcp runs someone else's code.

Claude Code

{
  "mcpServers": {
    "refract": {
      "command": "npx",
      "args": ["-y", "@refract-org/cli", "mcp"]
    }
  }
}

Then prompt Claude Code:

"Analyze the Wikipedia page for Bitcoin and tell me when the most controversial edits happened."

Claude Code calls Refract's MCP tools to fetch the analysis, reads the structured events, and summarizes the findings.

VS Code / Cursor / GitHub Copilot

{
  "mcpServers": {
    "refract": {
      "command": "npx",
      "args": ["-y", "@refract-org/cli", "mcp"]
    }
  }
}

Claude Desktop

{
  "mcpServers": {
    "refract": {
      "command": "npx",
      "args": ["-y", "@refract-org/cli", "mcp"]
    }
  }
}

Tutorial: Ask an AI agent about claim history

1. Start the MCP server in one terminal:

   refract mcp

2. Connect your AI agent using one of the configs above.

3. Ask the agent:

   "Use Refract to analyze the Wikipedia page for ChatGPT. Show me a timeline of when citations were added and removed. Which citations have survived the longest?"

   The agent calls refract analyze via MCP, receives structured events, and interprets the citation lifecycle.

4. Track a specific claim:

   "Using Refract, check the claim 'GPT-4 was trained on 45TB of text data' on the ChatGPT Wikipedia page. When was it first added? Has it been removed or rewritten?"

   The agent calls refract claim via MCP, gets the claim's revision history, and reports.

5. Use model-assisted classification:

   "Classify the most recent edit to the ChatGPT page — is it a revert, a major addition, or a minor edit?"

   The agent calls refract classify heuristic via MCP with the edit data.

Available tools

analyze

Analyze a MediaWiki page's full edit history. Returns structured events with schemaVersion, FactProvenance, and deterministic hashes.

Parameter	Type	Required	Description
page	string	required	Page title
depth	enum		brief, detailed, forensic
api	string		MediaWiki API base URL
from	string		Start revision ID
to	string		End revision ID
since	string		Re-observe from ISO timestamp
config	object		Analyzer overrides: similarityThreshold, spikeFactor, clusterWindowMinutes, talkWindowBeforeDays, talkWindowAfterDays, renameDetection

claim

Track a specific sentence across revisions. Shows first seen, modifications, removals, and reintroductions.

Parameter	Type	Required	Description
page	string	required	Page title
text	string	required	Sentence text to track (partial match)
api	string		MediaWiki API base URL

export

Export page analysis as structured JSON.

Parameter	Type	Required	Description
page	string	required	Page title
depth	enum		brief, detailed, forensic
api	string		MediaWiki API base URL
since	string		Re-observe from ISO timestamp

cron

Re-observe pages for scheduled monitoring.

Parameter	Type	Required	Description
pagesFile	string	required	File path with one page title per line

classify

Ask a model to classify a single observation boundary. See CLI reference for the boundary table and all options.

The tool calls the configured inference provider when an API key is available, and otherwise asks the connected client through MCP sampling. With neither, the result has an empty output and a note saying no provider is configured; refract classify on the command line applies the boundary's default rule instead.

Parameter	Type	Required	Description
boundary	enum	required	revert, sentence_similarity, heuristic, template_signal, activity_spike
input	object	required	Input data (fields vary by boundary)
apiKey	string		Inference provider API key (env: REFRACT_INFERENCE_API_KEY)
endpoint	string		OpenAI-compatible endpoint (DeepSeek, Ollama, etc.)
model	string		Model name

get_statement_history

Track normalized text or partial word matches across revisions. Returns matched revision IDs, timestamps, an overall status, the latest match, and a short change summary. A partial match does not establish equivalent meaning.

Parameter	Type	Required	Description
statement	string	required	Raw text statement to track
page	string	required	MediaWiki page title
context	string		Optional use-context slug
depth	enum		brief, detailed (default), forensic
api	string		MediaWiki API base URL

Protocol

• Transport: stdio (stdin for requests, stdout for responses)
• Format: JSON-RPC 2.0 (protocol version 2025-06-18)
• Capabilities: tools, sampling
• Initialization: Send tools/list to discover tools, tools/call to invoke

Validation

To verify your MCP setup is working:

# Start the server
echo '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | refract mcp | head -1

You should receive a JSON-RPC response listing the available tools. For interactive testing, use the MCP Inspector:

npx @modelcontextprotocol/inspector npx -y @refract-org/cli mcp