Guide for coding agents

If you use Claude Code, Codex, Cursor or another coding agent, point it at the repository. It reads AGENTS.md first, which maps every kind of question (using it, results, research, design choices) to the file that answers it, and llms.txt for a short index.

Fastest path

1. pip install "readanydoc[engine]" and set ANTHROPIC_API_KEY.
2. Pick one interface:
   • Python in the same process: readanydoc.parse / readanydoc.extract (Python reference).
   • HTTP from another language or service: readanydoc serve (HTTP reference).
   • An existing Reducto integration: change the SDK's base_url (Migrate from Reducto).
3. For structured data, define a pydantic class and call extract; read values from out.data and their sources from out.citations.

Prompt you can give your agent

Use the open-source library readanydoc (pip install "readanydoc[engine]", docs at
https://arpitpatel25.github.io/ReadAnyDoc/docs/). Define a pydantic model for the fields we need, call
readanydoc.extract(path, schema=Model), and store out.data together with out.citations so a reviewer can
see where each value came from. Log out.usage.cost_usd and out.warnings.

Facts worth getting right

• Only Claude Haiku 5.5 (Sonnet 5.5 for handwriting) is benchmarked. Other models work; their accuracy is unmeasured.
• Benchmark scores come from each benchmark's own grader, not an AI judge.
• Confidence values are tiers (high, low) from how a value was found, not calibrated probabilities.
• Split, classify and edit are server features; the library covers parse and extract.