deepclaw-ai 0.1.1

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

# PyPI (recommended)
pip install deepclaw-ai
deepclaw

# From source
git clone https://github.com/boyin111-1/DeepClaw.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).

Usage

Command	Description
deepclaw	Launch interactive CLI (when installed via pip)
python -m deepclaw	Launch from source
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
/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)

Skill System

Skills are plain Markdown files with YAML front matter and Jinja2 templates. They define custom AI behavior — code review, project workflow, git operations, and more.

Place skills in ~/.deepclaw/skills/<name>/SKILL.md and load with /skill <name>. Example:

---
name: My Reviewer
params:
  focus: { type: enum, values: [安全, 性能, 风格], default: 安全 }
---
# {{ name }}
Focus on **{{ focus }}** when reviewing code.

Load with /skill reviewer — interactive parameter 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

Share skills and plugins via the Discussions tab.

License

MIT License — see LICENSE.