CountBot is a lightweight AI agent framework that better fits Chinese users' habits, and also a runtime hub built for local deployment and long-running operation.

It connects roles, teams, workflows, tool invocation, memory, LLMs, IM channels, and local workspaces so AI can run continuously, collaborate across entry points, and execute real tasks.

You can think of CountBot as both a framework and a hub:

Upward, it connects to 100+ LLM providers and different model strategies
Outward, it connects to Web, WeChat, Lark, DingTalk, Telegram, Discord, QQ, Weibo, and more
Inward, it organizes roles, teams, workflows, memory, and security boundaries
At the execution layer, it connects files, Shell, Web, screenshots, file transfer, and external professional tools such as Claude Code, Codex, and OpenCode

In one sentence: CountBot is an AI agent framework and runtime hub that connects models, channels, teams, and tools.

CountBot was born from natural language. Its vision is not to require more people to learn complex configuration and programming before they can use AI, but to let ordinary users interact with AI directly through natural language to retrieve information, generate content, break down tasks, call tools, orchestrate workflows, and even build their own assistants, team collaboration flows, and automation systems.

Latest Updates

v0.7.0
- Addressed multiple issue reports, fixed a batch of known bugs, and improved overall stability
- Added file upload and processing flows to the WEBUI, improved Mermaid rendering, and made chart presentation more user-friendly
- Optimized frontend pages and interaction flows so everyday operations around skills, configuration, and tools feel smoother
- Optimized the tool invocation chain and context organization, significantly reducing token usage compared with previous versions
- Added a model thinking control switch so you can adjust reasoning intensity by scenario and get faster perceived responses
- Added find-skills with full Tencent Cloud SkillsHub integration, enabling conversational search, install, enable, disable, and delete for skills
- Added ima-knowledge-base and ima-notes with full IMA knowledge-base and note capabilities, including search, upload, web import, reading, creation, and append
- Release notes: https://654321.ai/docs/releases/v0.7.0
v0.6.0
- Added WeChat ClawBot integration with support for binding multiple accounts
- Added external coding tool integration for Claude, Codex, and OpenCode, which can either be called by the LLM as tools or act as agents connected to IM channels
- Added the secure first-time remote initialization entry /setup/<random>
- Added REMOTE_SETUP_SECRET_TTL_MINUTES so the validity period of the remote setup entry can be controlled
- Strengthened remote authentication boundaries, covering /api/* and /ws/chat
- Release notes: https://654321.ai/docs/releases/v0.6.0
v0.5.0
- Agent Teams became a first-class capability, supporting multi-role collaboration, context handoff, and team-level orchestration
- The configuration system evolved from session-level configuration to role-level, team-level, and multi-bot coordination
- The channel system was further strengthened to support more enterprise and social entry points
- Frontend panels for chat, configuration, skills, teams, and tools were upgraded as a whole
- Release notes: https://654321.ai/docs/releases/v0.5.0
v0.4.0
- Introduced session-level configuration so different sessions can use different models and prompts
- Expanded channel support for Lark, WeCom, Weibo, Xiaozhi AI, and more
- Improved multi-agent collaboration and added usage entry points such as /help
- Release notes: https://654321.ai/docs/releases/v0.4.0
v0.3.0
- Multi-agent collaboration modes (pipeline / graph / council) were first systematized
- Scheduled tasks, skill configuration, and workspace management were significantly enhanced
- Release notes: https://654321.ai/docs/releases/v0.3.0
v0.2.0 · 2026-02-26
- Focused on bug fixes, interaction improvements, and cleanup of the build workflow
- Pushed CountBot from "basically runnable" toward "stable and ready to iterate"
Official Open Source Release · 2026-02-21
- CountBot was open-sourced for the first time, including the base framework, startup scripts, agent core, channel system, tool system, skill system, and foundational frontend capabilities
- Public iteration has continued from that day onward

Why CountBot

CountBot starts from the broader trend of software being driven by natural language, and aims to let AI truly land in real tasks and everyday work.

Software development and software usage are changing:

Interaction between people and software is moving from complex commands and code toward natural language.
Tools and systems are moving from standardized products toward a future where more people can define and create personalized tools.

The core challenge for AI agents is no longer just model capability alone. It is how to combine LLMs, tool invocation, multi-channel access, permission boundaries, and long-running operation into a system that can actually be deployed and maintained.

OpenClaw has already validated a viable path for local execution and autonomous agents. On top of that, CountBot focuses on filling key gaps in Chinese scenario adaptation, lightweight deployment, usability expansion, security control, and governance.

CountBot is not merely a conversational AI application. It is open infrastructure for real tasks: deployable, extensible, and governable, so ordinary users can use natural language to drive AI, organize tools, connect channels, and actually land business and daily workflows.

Core Capabilities

Module	Description
Agent Loop	ReAct reasoning, tool invocation, result feedback, and iterative control
Agent Teams	Three collaboration modes: `pipeline`, `graph`, and `council`
Multi-bot channel matrix	One workspace can serve multiple channels, multiple bots, and multiple business entry points
Configuration layers	Global defaults, roles, teams, and runtime session configuration working together
Tools and skills	Files, Shell, Web, screenshots, memory, workflows, media delivery, and skill extensions
External execution tool integration	Connect external professional tools such as Claude Code, Codex, and OpenCode into one unified runtime
Memory and sessions	Long-term memory, summaries, context injection, and session isolation
Cron and Heartbeat	Scheduled tasks, proactive reminders, and long-running background operation
Security and workspace	Local control, path restrictions, audit logs, timeouts, and remote authentication boundaries
Multi-model access	Compatible with major domestic and international models, with team-level model override support

Use Cases

Breaking down complex tasks into multiple roles that collaborate through natural-language instructions
Running your own AI assistant, AI team, or automation workflow locally or in a private environment
Serving multiple entry points at the same time, including Web, WeCom, Lark, DingTalk, Telegram, and Discord
Unifying tool invocation, file operations, message delivery, and scheduled tasks in one runtime
Bringing LLMs, message channels, tool usage, and team collaboration into a single operating hub
Going beyond a simple AI assistant to build a sustainable, long-running agent system

Quick Start

Option 1: Start From Source

git clone https://github.com/countbot-ai/CountBot.git
cd CountBot

# Default installation
pip install -r requirements.txt

# If you are in mainland China, you can use the Aliyun mirror
pip install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/

python start_app.py

After startup, CountBot opens at http://127.0.0.1:8000 by default.

If GitHub access is limited, you can switch to Gitee:

git clone https://gitee.com/countbot-ai/CountBot.git

Option 2: Use the Desktop Build

Gitee Releases: https://gitee.com/countbot-ai/CountBot/releases
GitHub Releases: https://github.com/countbot-ai/CountBot/releases
Supported platforms: Windows / macOS / Linux

Key Changes in v0.7.0

1. Focused on bug fixes and stability improvements

This release first tackles known issues coming from GitHub issues and daily usage feedback, covering tool invocation, configuration flows, skill integration, and interaction details, making CountBot more stable and predictable in complex scenarios.

2. Frontend interaction keeps getting tighter and faster

The frontend continues to be optimized around common operation paths, reducing the cognitive load of configuration and skill management so daily use feels more direct.

3. The tool invocation chain is now leaner

The tool flow, context assembly, and execution chain were reworked again. Without sacrificing capability coverage, token usage dropped significantly.

4. Model thinking control is more precise

A new thinking control switch lets you decide whether to enable or reduce reasoning intensity based on task complexity, which improves perceived response speed in many high-frequency scenarios.

5. SkillsHub is now part of the conversational management flow

The new find-skills skill integrates Tencent Cloud SkillsHub into the conversation flow. You can now search, install, enable, disable, and remove skills directly through natural language without manually handling underlying files.

6. IMA knowledge base and notes are now built in

CountBot adds two new skills:

ima-knowledge-base: search knowledge bases, inspect hit details, browse entries, upload files, and import web pages
ima-notes: search notes, read notes, create new notes, and append content

This allows CountBot to write material directly into the IMA knowledge base or persist content into the IMA notes system, further strengthening the idea of "conversation as management".

Documentation

Document	Description	Link
Quick Start	Installation, configuration, and startup	https://654321.ai/docs/getting-started/quick-start-guide
Configuration Manual	Full configuration reference	https://654321.ai/docs/getting-started/configuration-manual
Deployment and Operations	Startup, deployment, and troubleshooting	https://654321.ai/docs/advanced/deployment
Remote Access Guide	Remote initialization, authentication, and troubleshooting	https://654321.ai/docs/advanced/remote-access
Authentication	Password setup and access boundaries	https://654321.ai/docs/advanced/auth
API Reference	REST API and WebSocket	https://654321.ai/docs/api-reference
Release Notes	Version history and release changes	https://654321.ai/docs/releases/v0.7.0

For the full documentation site, see: https://654321.ai/docs

Development and Contribution

Local Development

uvicorn backend.app:app --reload --host 0.0.0.0 --port 8000

Community

QQ group: 1028356423
Discussion topics: CountBot usage, issue feedback, secondary development, and scenario co-creation

Issue Reporting

GitHub Issues: https://github.com/countbot-ai/CountBot/issues

License and Acknowledgements

License

MIT License

Project Inspiration

OpenClaw
NanoBot
ZeroClaw
anthropics/skills

Technical Thanks

Thanks to FastAPI, Vue.js, SQLAlchemy, Pydantic, LiteLLM, and other open-source projects.

An AI agent runtime hub that connects models, channels, teams, and tools

Official Website · GitHub · Gitee · Full Documentation

Description

百度 AI 搜索。支持网页搜索、百度百科、秒懂百科、AI 智能生成四种模式。自动包含当前日期上下文。当用户要求搜索信息、查询百科、获取最新资讯、搜索新闻、查找资料时使用。

agent-skills ai-agents ai-assisted assistant deepseek-v4 official

Readme MIT 21 MiB

Languages

Python 53.1%

Vue 23.1%

CSS 15.2%

TypeScript 7.1%

HTML 1.3%

Other 0.2%