Second Read

Second Read

Second Read reviews a draft the way a careful editor would, and shows its sources. You give it an article, a blog post or a design doc. It returns one report:

• Facts checked against the web. Each claim is looked up, and every "this is wrong" comes with the page that says so.
• The strongest counter-case. A devil's advocate builds the best argument against your piece from sources it retrieves.
• Copy edits. Grammar and clarity fixes that keep your voice.
• Originality notes. Sentences that closely match text found online.

It runs as an MCP server with one tool, review_document, so you can ask for a review from Claude Code or Claude Desktop. It also has a command-line tool. It runs on your machine with your own API keys; nothing is sent to the author of this project.

A finding looks like this:

Factual error — "Mjällby clinched the title on 20 October 2025 by winning

2-0 at IFK Göteborg, with two games to spare"

The date, the 2-0 score and the opponent are confirmed. ESPN and Flashscore

both say three rounds were left (27 of 30 played), so "two games to spare" is

wrong.

What you need

• uv (it installs Python 3.12 for you).
• An Anthropic API key: ANTHROPIC_API_KEY.
• A Tavily search key: TAVILY_API_KEY.

Getting your keys

You need two keys. Each takes a couple of minutes: sign up, create a key, copy it.

The Anthropic key is pay-as-you-go: reviews fail with a clear message if the account has no credit. Keep keys out of chats and out of version control; put them in the MCP client's env block or in a .env file as shown below.

A review of a 700-word article takes about two minutes and uses roughly 150,000–180,000 tokens on your model key (about $0.60–$1.20 at Anthropic's list prices in October 2026) plus about 30 searches.

Install: Claude Code

claude mcp add second-read \

-e ANTHROPIC_API_KEY=your-key \ -e TAVILY_API_KEY=your-key \ -- uvx second-read

Start Claude Code with a longer tool timeout, because a review takes minutes:

MCP_TOOL_TIMEOUT=600000 claude

Then ask: "Use second-read to review this draft", and paste the text or give a public URL.

Install: Claude Desktop

Add this to claude_desktop_config.json (on macOS: ~/Library/Application Support/Claude/claude_desktop_config.json), then restart the app. If Claude Desktop cannot find uvx, use its full path (which uvx).

{

"mcpServers": { "second-read": { "command": "uvx", "args": ["second-read"], "env": { "ANTHROPIC_API_KEY": "your-key", "TAVILY_API_KEY": "your-key" } } } }

To try it without keys, replace the env block with {"SECONDREAD_FAKE": "1"}. Demo mode returns placeholder findings and makes no model calls.

Install: other MCP clients

Any client that runs local (stdio) MCP servers can use the same command. For Cursor, VS Code and similar, add a server named second-read that runs uvx second-read with ANTHROPIC_API_KEY and TAVILY_API_KEY in its environment, as in the Claude Desktop example above.

Second Read is listed in the official MCP Registry as io.github.sbanthia92/second-read, so clients and directories that read the registry can find it.

ChatGPT and claude.ai on the web connect only to remote (hosted) MCP servers, so they cannot run Second Read yet. A hosted version is on the roadmap.

Command line

export ANTHROPIC_API_KEY=your-key TAVILY_API_KEY=your-key

uvx --from second-read secondread review path/to/draft.md --profile opinion > report.md

Or keep the keys in a .env file and pass it with --env-file path/to/.env.

Progress prints to the terminal; the report goes to report.md. Use --profile design_doc for technical design docs, or leave the profile out to let the tool decide. A URL works in place of a file path.

The tool

review_document takes:

It returns the report as markdown and sends progress updates while it works.

The report has these sections: a verdict line; Must fix (factual errors and claims the sources do not back); Could not verify (claims it found no adequate source for, which is not the same as wrong); Strongest counter-case; Should fix (structure); Polish (copy edits); Originality notes; and an appendix listing every claim with its verdict and sources.

Settings

Set these in the MCP client's env block, in the environment, or in a .env file (the server reads .env from its working directory, or the file named by SECONDREAD_ENV_FILE).

Privacy

Second Read runs on your machine. The project's author runs no server and receives no data.

• What it collects. Nothing is collected by the project. To do a review the tool handles the document you give it and the web pages it retrieves.
• What leaves your machine, and where.
• Your document goes to Anthropic, under your key and their terms.
• Search queries built from your document's claims, and up to eight sentences quoted exactly for the originality check, go to Tavily, under your key and their terms.
• Web pages named in search results are fetched directly from your machine. The fetcher refuses private and local network addresses.
• Storage and retention. Nothing is stored after the review finishes. The report is returned to your client and is not kept by Second Read. Logs go to your terminal and never contain keys or document text.
• Sharing. Nothing is shared with anyone else. The only third parties are the model and search providers above, which you choose by supplying their keys.
• Keys. Read from your environment or a .env file you control, and never logged.
• Contact. Questions or concerns: open an issue at https://github.com/sbanthia92/second-read/issues. For security problems see SECURITY.md.

How well it works

What was tested. Four articles written for the test, each with errors planted in it that cannot be settled from memory (recent results, exact figures): 19 planted wrong facts in all. Every planted error has an answer key recording the exact sentence and the true value. Each article was reviewed once by two setups:

• Single prompt: one request to the model Second Read uses for its strongest tier, "review this and argue the other side". No web search, so it cannot look anything up or cite a page.
• Second Read: the full review described below, with web search.

What "caught" means. The report has a factual-error finding on the sentence that holds the planted error. A "could not verify" note does not count, so a careful miss scores as a miss. "Rebutted" means a counter-argument was written against a deliberately weak argument; "caught" for borrowed sentences means an originality note on the copied sentence.

Reading the result. Second Read earns its cost on claims that need looking up: recent events, exact figures, and details a model cannot know from memory. The gain appears to come mainly from searching and reading sources, though this test did not separate that from the other design choices. Every finding names its source so you can check it in one click. It is slower, uses about 27 times the tokens, and did worse on borrowed sentences.

Limits of this test. It is small: 19 errors, one run each, so each error is worth about five points and the exact percentages are rough. It does not measure false alarms (true facts wrongly called wrong) for either setup. It does not say which of the six misses were careful "could not verify" results and which were missed outright. A larger run (all 40 documents, three runs each) is on the roadmap.

Getting the most from it:

• Open the cited source before acting on a "factual error". The report links the page and quotes the passage it relied on.
• Search results change, so two runs on the same draft can differ slightly.
• The originality check is a heuristic: it flags close matches found online.
• Sources are web pages; PDFs are not read.

How it works

An orchestrator runs specialist agents and settles their disagreements:

1. A claim extractor splits the document into its thesis, supporting claims and checkable facts.
2. In parallel: a fact-checker researches each fact in a bounded loop (search, fetch, read, assess; at most three rounds per claim); a devil's advocate gathers counter-evidence; a copy editor; for articles, an originality checker; and, for design docs, a structure reviewer.
3. A reaction round: a verified fact the devil's advocate disputes is re-checked, and the two argue once over the most important disagreements.
4. Fixed rules settle what is left (for example, a grammar fix on a sentence flagged as wrong is dropped), findings are ranked, and the report is written.

ARCHITECTURE.md describes how the system is built, the decisions behind it and their trade-offs, the planned hosted service, and the shared types.

Development

uv sync

uv run ruff check . && uv run ruff format --check . && uv run mypy && uv run pytest

Tests run offline with fakes. Tests that call real providers are marked live and skipped by default (uv run pytest -m live). The evaluation harness and its 40 seeded documents are in eval/ (see eval/README.md); running it against real providers spends money on your keys.

Status

The local MCP server and command-line tool work. A hosted version (sign in, add a key once, get the report by email) is designed in ARCHITECTURE.md but not built. ROADMAP.md lists what comes next.

License

MIT. See LICENSE.