claude-pipe 0.1.0

A pip-installable Python CLI that pipes stdin into Claude with Pydantic-validated structured outputs.

claude-cli

A pip-installable Python CLI that pipes stdin into Claude with structured outputs.

cat samples/error.log | claude-cli --schema schemas.RootCause

…and you get back validated JSON, not a paragraph you have to parse.

Why

Built to scratch a real itch — I kept reaching for Claude in shell scripts and writing the same boilerplate every time. Production LLM tools live or die on output discipline, so this one bakes Pydantic validation in at the boundary.

Install

pip install claude-pipe
export ANTHROPIC_API_KEY=sk-ant-...

Or, from source:

git clone https://github.com/nishithabottu/claude-cli && cd claude-cli
pip install -e .

Worked example

schemas.py defines a RootCause Pydantic model. Pipe a log line through the CLI with --schema, and the response is a JSON object that matches the schema exactly:

echo '2026-05-09T14:32:11Z [ERROR] auth-api: jwt.ExpiredSignatureError at token_service.py:142' \
  | claude-cli --schema schemas.RootCause

{
  "severity": "error",
  "service": "auth-api",
  "summary": "JWT signature was expired during token issuance.",
  "likely_cause": "Stale signing key or clock skew in token_service.py.",
  "evidence": ["jwt.ExpiredSignatureError", "token_service.py"],
  "error_code": null
}

Flags

flag	what it does
--system	system prompt prepended to the request
--schema	dotted path to a Pydantic model (e.g. schemas.RootCause)
--stream	stream tokens to stderr while assembling the response
--model	Anthropic model name (default: claude-sonnet-4-6)
--max-tokens	response token cap (default: 1024)

Eval

The success metric: samples/error.log has 10 hand-authored log lines. scripts/eval.sh pipes each through the CLI with --schema schemas.RootCause and counts how many produce valid Pydantic-typed JSON. Target: ≥ 9 / 10.

./scripts/eval.sh
# Result: 10/10 valid responses (run on 2026-05-10, claude-sonnet-4-6)

Tests

pip install -e ".[dev]"
pytest

License

MIT.