carp

Unix-style tooling for building, running, and observing AI agents.

This repository is the CARP CLI + Python SDK. v0 is a Python package:

pip install -e ".[dev]"
carp doctor
carp init my-agent
cd my-agent
carp validate
carp run --input '{"message":"ping"}'
echo '{"message":"ping"}' | carp run --input -

Implementation reports for each usage milestone live in registry/.

Layout

src/carp/          # library (protocol, manifest, SDK, runtime, CLI, interop)
schemas/           # JSON Schema for carp.yaml and the invoke envelope
examples/          # repo-architect, rag-agent, interop-*
docs/interop.md    # OpenAI / Anthropic / LiteLLM / LangChain / CrewAI / Pi
tests/
registry/          # standalone HTML reports, one per milestone

Frameworks: wrap the LLM loop in a carp.Agent handler and call ctx.tool for side effects. Official Pi cannot be runtime.entrypoint (TypeScript). See docs/interop.md.

Commands (U0)

Command	What it does
carp version	Print package version
carp doctor	Check Python, install, config dir
carp init	Scaffold prompt + tools + history-aware loop
carp validate	Lint the manifest
carp secret ls	Names the manifest expects (never values)

Verify

make install
make test

GitHub Actions runs pytest on Python 3.11 and 3.12, Studio unit tests, and a Studio production build on every push and pull request. Locally, make ci runs typecheck, Studio tests, and pytest without rebuilding committed static assets.