GPT-AGI/Clawd-Code

📚 What is Clawd‑Code?

Clawd‑Code is a Python re‑implementation of Anthropic’s Claude Code agent. The original Claude Code was written in TypeScript; this project ports the same architecture to idiomatic Python (3.10+), adds type hints, tests, and a ready‑to‑run command‑line interface.

It isn’t just a code dump – it ships a fully functional CLI REPL that can:

stream LLM responses token‑by‑token,
invoke tools (file read/write, web fetch, bash, etc.) in a loop, and
be extended with markdown‑driven “skills” (slash commands) that limit which tools they may use.

The goal is to give developers a production‑ready, hackable Claude‑style coding assistant that works locally with any of the supported LLM providers.

✨ Core Features (as described in the README)

Feature	What it does
Streaming Agent Experience	Real‑time token streaming, `/stream` toggle, `/render-last` to re‑render the previous answer as Markdown.
Programmable Skill Runtime	Skills are defined in `SKILL.md` files (markdown front‑matter). They can declare arguments, allowed tools, and a description.
Multi‑Provider Support	Works with Anthropic Claude, OpenAI GPT, and Zhipu GLM out of the box; providers are easy to add.
Interactive REPL	Rich terminal UI with history, tab‑completion, multiline mode, session save/load, and built‑in commands (`/help`, `/save`, `/load`, …).
Complete CLI	`clawd` starts the REPL, `clawd login` configures API keys, `clawd config` shows settings, `clawd --version` prints the version.
Tool System	30+ built‑in tools (file ops, Bash, web fetch/search, task management, agent‑level tools, configuration, etc.) with validation and sandboxing.
Session Persistence	Sessions can be saved locally and re‑loaded later.
Tests & Docs	Core test suite covering skills, providers, REPL, tools, and context; >10 markdown docs and a changelog.

🚀 Getting Started

# Clone the repo
git clone https://github.com/GPT-AGI/Clawd-Code.git
cd Clawd-Code

# Create a virtual environment (uv is recommended)
uv venv --python 3.11
source .venv/bin/activate

# Install dependencies
uv pip install -r requirements.txt

Configure your LLM provider

python -m src.cli login   # interactive wizard

The wizard asks you to pick a provider (Anthropic, OpenAI, or GLM), enter the API key, optional base URL, and a default model. The configuration is saved to ~/.clawd/config.json.

Run the agent

python -m src.cli          # start the REPL
python -m src.cli --help   # show CLI options

You can now type natural language, use /stream on for live output, or invoke a skill, e.g.:

>>> /explain-code qsort.py

🛠️ Extending with Skills

Create a directory under the project (or ~/.clawd) like:
```
.clawd/skills/explain-code/SKILL.md
```

Write a markdown front‑matter block:

---
description: Explain code with analogies and diagrams
allowed-tools:
  - Read
  - Grep
  - Glob
arguments: [path]
---

Explain the code in $path. Start with an analogy, then draw a diagram.

Call it from the REPL with /explain-code <file>.

The skill system enforces the listed allowed-tools, substitutes $path with the argument, and runs inside the same sandboxed tool loop.

📦 Project Layout

Clawd-Code/
├─ src/
│  ├─ cli.py            # entry point
│  ├─ providers/        # Anthropic, OpenAI, GLM adapters
│  ├─ repl/             # interactive REPL implementation
│  ├─ skills/           # loader for SKILL.md files
│  └─ tool_system/      # registry, validation, loop logic
├─ tests/                # unit & integration tests
├─ .clawd/skills/        # example project‑level skills
└─ FEATURE_LIST.md      # detailed feature status

🤝 Contributing

The repo follows a typical Python open‑source workflow:

pip install -e .[dev]          # install in editable mode with dev extras
pytest -v tests/                # run the test suite

Guidelines are in CONTRIBUTING.md; pull requests are welcomed.

⚡ Performance & Security

Startup < 1 s, Memory < 50 MiB.
API keys are stored base‑64‑encoded in the user config and .env files are ignored.
No secret data is committed to the repository.

📄 License

MIT – see LICENSE.

Clawd‑Code gives you a Claude‑style coding assistant that runs locally, is fully scriptable, and can be extended with your own markdown‑based skills—all in pure Python.