deepclaw-ai 0.1.0

A local AI agent powered by DeepSeek — reads, writes, executes, and browses the web from your terminal.

DeepClaw

DeepClaw is a local AI agent powered by DeepSeek models, running in your terminal. It reads/writes files, executes shell commands, searches codebases, and browses the web — autonomously. Features streaming output, context compression, a Jinja2-driven Skill system, and a VS Code-style Plugin architecture for unlimited extensibility. Zero environment variables, one-command setup, ready to work.

Software Architecture

User → CLI (Rich TUI) → Agent Core (streaming + tool-call loop) → DeepSeek API
                              ↓
                    Tool Execution Layer
           (file I/O, shell, grep, web search, HTTP)
                              ↓
              Skill Engine (Jinja2) ←→ Plugin System (hot-load)

Built with Python, using openai SDK for API calls and Rich for terminal UI. Skills use Jinja2 templates with YAML front matter; Plugins are dynamically-loaded Python modules following a VS Code Extension-like contract.

Installation

git clone https://github.com/your/repo.git
cd DeepClaw
pip install -r requirements.txt
python -m deepclaw

On first run, you'll be guided through setup: choose DeepSeek official API (API Key only) or third-party API (Key + URL + model).

Instructions

Command	Description
python -m deepclaw	Launch the interactive CLI
python -m deepclaw --model deepseek-v4-pro	Start with a specific model
python -m deepclaw --plugin weather	Load plugins at startup
python -m deepclaw --resume	Restore last session

In-session commands:

Command	Description
/help	Show all commands
/skill <name>	List / load / unload skills (interactive)
/plugin list	Manage plugins
/model	View / switch models
/config	View / edit configuration
/back	Load / delete saved sessions
/export	Export conversation to Markdown
/exit	Quit (auto-saves session)

Built-in Skills

Skill	Description
python-coder	Python coding assistant — PEP 8, type hints, testing
git-helper	Git operations — conventional commits, safety checks
reviewer	Code reviewer — configurable focus (security / performance / style), strict mode, max issues
supercoder	Full project workflow — requirements → design → framework → implement → test → iterate

Skills use Jinja2 templates with YAML parameter specs. Load via /skill <name> — interactive forms appear for skills with parameters.

Plugin System

Plugins run arbitrary Python code with full freedom:

# ~/.deepclaw/plugins/my-plugin/main.py
def on_load(api):
    api.register_tool({"function": {"name": "my_tool", ... }}, handler)
    api.status_bar.set("my-plugin", "Ready")
    api.on("chat:after", lambda resp: print(resp))

Extensions can: inject tools, register commands, subscribe to events, run background services, or even take over the UI entirely (mode: gui).

Contribution

1. Fork the repository
2. Create Feat_xxx branch
3. Commit your code
4. Create Pull Request

Skills and plugins can be contributed by adding a directory to ~/.deepclaw/skills/ or ~/.deepclaw/plugins/ with the required files — no fork needed for content contributions.

License

MIT License — see LICENSE.