A lightweight but powerful AI Agent framework in Python.
# Install
pip install fastclaw-ai
# Run
fastclaw start
# Persistent running (recommended for server)
nohup fastclaw start > /tmp/fastclaw.log 2>&1 &After starting, visit http://localhost:8765 , click Settings in the top right corner to configure main_agent, recommended: deepseek-chat.
Or edit the config file:
vim ~/.fastclaw/workspace/data/agents/main_agent/metadata.json# Interactive chat
fastclaw chat
# New session
fastclaw chat --new
# Check status
fastclaw statusgit clone https://github.com/kandada/fastclaw.git
cd fastclaw
pip install -r requirements.txt# Run
python main.py start
# Persistent running
nohup python main.py start > fastclaw.log 2>&1 &After starting, visit http://localhost:8765 , click Settings in the top right corner to configure main_agent, recommended: deepseek-chat.
Or edit the config file:
vim workspace/data/agents/main_agent/metadata.json# Interactive chat
python main.py chat
# New session
python main.py chat --new
# Check status
python main.py statusFastClaw is built on the core concept of event-driven agent orchestration. Key design principles:
- Graph-based Agent: Uses a directed graph to orchestrate agent behavior, supporting conditional branching between nodes (e.g., tool execution, response generation)
- Event Streaming: Real-time streaming output of LLM responses for better user experience
- Tool System: Extensible tool framework supporting Shell commands, skills, and custom integrations
- Session Management: Persistent conversation history with multi-session support
- Cron Scheduling: Built-in cron-style task scheduling for automated workflows
- π€ LLM-powered - Built on FastMind framework with streaming support
- π§ Tool Calling - Execute Shell commands, skills, and more
- β° Cron Jobs - Schedule tasks with cron expressions
- π¬ Multi-channel - Feishu, iMessage integrations
- π¨ Extensible - Easy to add custom skills and agents
- π Python Ecosystem - Seamlessly call professional libraries like numpy and pandas, enabling AI to use Python ecosystem tools as skillfully as humans
- β¨ New: Bundled
fastbrowserskill β real-browser automation (JS/SPA rendering, screenshots, downloads) and OS accessibility (read/act on web pages and native desktop apps) via thefastbrowserCLI driven withrun_shell
- β¨ New: Default context (max_tokens / unload_threshold_tokens) raised to 256000 β most mainstream LLMs now support this window or larger
- π§ Improvement: Default agent personality files (SOUL.md / USER.md / AGENT.md) translated to English; USER.md language preference now defaults to English for international users
- β¨ New: Full Anthropic gateway support β Messages API streaming, tool calling, extended thinking (adaptive β enabled β none state machine), and auto base_url adjustment for MiniMax / DeepSeek / Moonshot Anthropic-compatible endpoints
- β¨ New: OpenAI gateway thinking compatibility β unified extraction of
reasoning_content/thinkingfield /<think>...</think>tags (MiniMax-M3, Qwen3, etc.) - π Fix: WebUI in-progress view no longer drops tool calls β repeated same-name tools across ReAct rounds and parallel calls now each get their own card (id-based matching)
- π Fix: WebUI message block ordering β content now appears before tool_call in history replay and resume, consistent with live streaming (thinking β content β tool_call β tool_result)
- π Fix: WebUI tool result cards unified to a generic "tool" label, consistent with the stored message format
- β¨ New: Cron jobs now support Feishu (Lark) β scheduled tasks can deliver results to Feishu chats
- π§ Improvement: WebUI refactor β optimized conversation list interaction
- π Fix: Cron scheduler fixes and regression fixes
- β¨ New:
skill_creatorbundled meta-skill β guides the agent to create, update and optimize skills via chat - β¨ New: Skills management tab in WebUI β create and edit user skills with Remote Endpoint / Secret support
- β¨ New: Skills CRUD API endpoints (
/api/skills/user) β create, read, update, delete user skills - β¨ New: Hot-reload for skills β new skills are auto-registered at runtime, no restart needed
- β¨ New: "Open Skills Dir" button in WebUI to open the skills directory in file explorer
- π§ Improvement: All bundled skill documentation (SKILL.md) translated to English
- π§ Improvement: Template
_templateexcluded from skill lists - π Fix: Skills list in system prompt now updates when new skills are created
- π Fix: Added
python-socks[asyncio]dependency to fix Lark (Feishu) connection failure when using a SOCKS proxy
- π Fix: Resolved
SyntaxErroron Python 3.10/3.11 caused by backslash (\) inside f-string{}expressions (a Python 3.12+ syntax) - π§ Improvement: Added version compatibility comments at f-string usage sites to prevent future regressions
- π§ Improvement: Added copyright and GPLv3 license headers to all source files
- π Fix: Resolved
UnicodeDecodeErroron Windows Chinese systems by addingencoding="utf-8"to allread_text()/write_text()calls, fixing the crash on second startup - π Fix: WebUI agent edit/create buttons now show error messages on failure instead of silently doing nothing
- π Fix: Removed 3 extra
</div>tags from WebUI HTML that could cause DOM parsing issues - π§ Improvement: Startup messages now show
http://localhost:{port}alongsidehttp://0.0.0.0:{port}for easier access on Windows - π§ Improvement: Added diagnostic logging to silent
exceptblocks in router.py for easier debugging
- π§ Improvement: Made
ripgrepan optional dependency to avoid compilation issues on Windows without MSVC - π§ Improvement: Added Python 3.13 / 3.14 support
- π Fix: WebUI bugs and cron scheduler bugs
- β¨ Feature: LLM Gateway selector (OpenAI / Anthropic) in WebUI agent settings
- β¨ Feature: Auto-bootstrap default agents from GitHub when workspace is empty
- β¨ Feature: Auto-generate settings.json with defaults when missing
- π§ Improvement: Increased default Unload Threshold from 80k to 156k
- π Fix: Context unloading now only affects LLM input, messages.jsonl always keeps full history
- π Fix: Message timestamps are no longer overwritten on every save
- π Fix: CancelledError no longer triggers duplicate response generation
- π§ͺ Test: Added context unloading, sliding window, and timestamp preservation tests
GPL-3.0
Copyright (c) 2024-2026 xiefujin. All rights reserved.
https://www.fastclaw.world/ | https://fastclaw-ai.com/
FastClaw is part of the FastMind ecosystem. Other projects built on FastMind:
- FastClaw β A lightweight, general-purpose AI Agent framework with event-driven agent orchestration, graph-based agent design, and extensible tool system
- FastBot β A dual-loop embodied intelligence robot simulation demo, exploring perception-action closed-loop control with LLM-driven decision making
Inspired by OpenClaw. Special thanks to the open source community.
FastClaw is powerful, but its use depends entirely on the user - all responsibility lies with the user yourself.
AuthorοΌxiefujin email: 490021684@qq.comοΌwelcome to contact me.