Second Read v0.1.3

Second Read is now listed in the official MCP Registry as io.github.sbanthia92/second-read.

• Adds server.json for MCP directories and clients that read the registry.
• No change to how reviews work.

Second Read v0.1.2

Fixes:

• Argument sentences that state checkable facts (dates, scores, figures) are now fact-checked, not only sentences typed as facts.
• The copy editor no longer edits quotations, block quotes or sentences the originality check flags as borrowed.
• The README shown on PyPI now uses the short install (uvx second-read).

Second Read v0.1.1

Second Read is now on PyPI.

claude mcp add second-read \
  -e ANTHROPIC_API_KEY=your-key \
  -e TAVILY_API_KEY=your-key \
  -- uvx second-read

Changes since v0.1.0:

• Published to PyPI as second-read; uvx second-read starts the MCP server.
• Package metadata (description, keywords, project links).

See the v0.1.0 release for what Second Read does.

Second Read v0.1.0

First release of Second Read: evidence-backed document review for Claude, as an MCP server.

Give it an article, a blog post or a design doc and it returns one report:

• Facts checked against the web. Each claim is researched (search, fetch, read; up to three rounds), and every "this is wrong" comes with the page and passage it relies on. Claims it cannot confirm are listed separately under "Could not verify".
• The strongest counter-case. A devil's advocate builds the best opposing argument from retrieved sources.
• Copy edits that keep your voice, and originality notes for sentences that closely match text online.

Install

Needs uv, an Anthropic API key and a Tavily search key.

claude mcp add second-read \
  -e ANTHROPIC_API_KEY=your-key \
  -e TAVILY_API_KEY=your-key \
  -- uvx --from git+https://github.com/sbanthia92/second-read@v0.1.0 secondread-mcp

Claude Desktop and command-line setup are in the README.

Good to know

• Runs on your machine with your own keys. A review of a 700-word article takes about two minutes and roughly 150,000–180,000 tokens.
• On a small test of articles whose errors need looking up, it caught 13 of 19 planted wrong facts with a cited source for each, against 7 of 19 (no sources) for a single prompt. Details in the README.