Agents Council
The simplest way to bridge and collaborate across AI Agent sessions like Claude Code, Codex, Gemini, or Cursor. It allows your agents to combine their strengths to solve your most difficult tasks without leaving their current context.
README
<p align="center"> <img src="./.github/agents-council.jpg" alt="Agents Council" width="1280" height="714" /> </p>
<h1 align="center">Agents Council</h1> <p align="center"><strong>The simplest way to bridge and collaborate across AI Agent sessions</strong></p>
<p align="center"> Status: <code>Experimental</code> </p>
🏛️ Overview
Designed for developers who already have active agent sessions and want them to collaborate without extra infrastructure.
Agents Council is the simplest way to bridge and collaborate across AI Agent sessions like Claude Code, Codex, Gemini, Cursor or others. It allows your agents to combine their strengths to solve your most difficult tasks without leaving their current context.
The most powerful way to use the council is by connecting your existing, active sessions. You can initialize them with the specific context you want, let them brainstorm or peer-review the matter, and then seamlessly take over the session once the council has finished.
Inspired by Andrej Karpathy's LLM Council, it provides an MCP-based CLI tool that lets multiple agents communicate with each other and find solutions to your most complex tasks.
✨ Features
- Centralized agent communication via MCP stdio server (no complex peer-to-peer networking).
- Summon Claude or Codex: Instantly summon Claude or Codex into your council when needed. Reuses local CLI authentication where available.
- Session Preservation: Start agents with your specific context, let them collaborate, and resume when they are done.
- Human Participation: A local desktop Council Hall app to monitor or join the discussion.
- Private & Local: State is stored on disk at
~/.agents-council/state.json. - Flexibility: Markdown or JSON text output for agent readability or automation.
🚀 Getting Started
Agents Council requires Node.js or Bun.
1. MCP Mode (Zero Install)
No installation is needed when using only the MCP mode. You can add the agents council MCP server to your agents using npx (or bunx). See the MCP Setup section below for specific commands.
If you only need agent-to-agent communication, skip to MCP Setup. Install the CLI only for the desktop app and global council command.
2. Desktop App & CLI
If you want to participate via the desktop app or use the council command globally, install the package:
npm install -g agents-council
Then run one of these:
council # desktop default launch
council chat # compatibility alias: also launches desktop
council mcp # terminal MCP server mode
3. Distribution Model
Releases keep a single user-facing npm package (agents-council) backed by platform optional dependencies:
agents-council-linux-x64agents-council-linux-arm64agents-council-darwin-x64agents-council-darwin-arm64agents-council-windows-x64
Each platform package ships:
- the
councilCLI binary for terminal mode (--help,--version,mcp), and - Electrobun desktop artifacts in
desktop-artifacts/for native installer/update distribution.
GitHub releases publish the Electrobun desktop-launchable artifacts for macOS, Windows, and Linux.
🔌 MCP Setup
Add the council to your favorite MCP client using the commands below.
<details> <summary>Claude Code</summary> Use the Claude Code CLI to add the Agents Council MCP server (<a href="https://docs.anthropic.com/en/docs/claude-code/mcp">guide</a>):
claude mcp add council npx agents-council@latest mcp
or use a predefined Agent Name:
claude mcp add council -s user -- npx agents-council@latest mcp -n Opus
</details>
<details> <summary>Gemini CLI</summary> Use the Gemini CLI to add the Agents Council MCP server (<a href="https://geminicli.com/docs/tools/mcp-server/#adding-a-server-gemini-mcp-add">guide</a>):
gemini mcp add council npx agents-council@latest mcp
</details>
<details> <summary>Codex</summary> Use the Codex CLI to add the Agents Council MCP server (<a href="https://developers.openai.com/codex/mcp/#add-a-mcp-server">guide</a>):
codex mcp add council npx agents-council@latest mcp
</details>
<details> <summary>amp</summary> Use the amp CLI to add the Agents Council MCP server (<a href="https://ampcode.com/manual#mcp">guide</a>):
amp mcp add council npx agents-council@latest mcp
</details>
<details> <summary>Other MCP integrations</summary>
{
"mcpServers": {
"council": {
"command": "npx",
"args": [
"agents-council@latest",
"mcp"
]
}
}
}
</details>
📖 Quick Start
- Start a council session in one terminal (e.g. via Claude Code) and describe the complex topic you need help with.
- Join the council from another terminal (e.g. via Codex or Gemini) to provide feedback.
- Review feedback and take over the session once the council has provided enough insights.
<p align="center"> <img align="middle" src="./.github/cc-start-council.png" alt="Claude Code start council" width="317" /> → <img align="middle" src="./.github/codex-join_council.png" alt="Codex join council" width="381" /> </p>
💬 Desktop Council Hall
Run the local desktop interface for human participants:
council
council chat remains available as a compatibility alias and opens/focuses the same desktop interface.
The desktop interface follows the Council Hall redesign:
- Council Sidebar: session chronicle, session spawn action, and archive area.
- Council Hall: active session header, voice stream, composer, and summon entry point.
The summon workflow is part of the Hall interface. Model and agent selections persist in ~/.agents-council/config.json.
UI implementation and future UI feature work should follow docs/ui-spec.md and docs/ui-implementation-progress.md.
🪄 Summon Agents
The Summon feature lets you bring a Claude or Codex agent into the active Council Hall session. Summoned agents review the matter and prior feedback, then contribute their response to the same voice stream.
Summon Claude
Prerequisites:
- Claude Code must be installed and available in your PATH
- Run
claudeat least once to authenticate
How it works:
- A summoned Claude agent joins the active council session
- It reads the current request and any prior feedback
- It contributes a response using the council tools
- The agent is granted read-only access to the project (Read/Glob/Grep) plus council tools
- Other tools follow your Claude Code user permission settings
Advanced Configuration:
These settings are optional. By default, Agents Council looks for claude in your PATH.
| Setting | Default | Description |
|---|---|---|
| Claude Code Path (Settings UI) | claude |
Custom path to the Claude Code executable |
CLAUDE_CODE_PATH (env var) |
claude |
Alternative way to set the path |
AGENTS_COUNCIL_SUMMON_DEBUG |
(unset) | Set to 1 to enable debug logging |
Summon Codex
Prerequisites:
- Run
codex loginonce (authentication is handled by the Codex SDK/CLI) - Optional: set a default model in
~/.codex/config.toml(for example,model = "gpt-5.2-codex")
How it works:
- A summoned Codex agent reads the current request and any prior feedback
- It returns a single response that is recorded in the council
- Codex runs via the Codex SDK CLI in read-only mode by default
Advanced Configuration:
Authentication is handled by the Codex SDK/CLI (typically codex login). If your Codex setup uses API keys or a custom base URL, those will still be honored.
These settings are optional. By default, Agents Council uses the bundled Codex CLI from the SDK.
| Setting | Default | Description |
|---|---|---|
| Codex CLI Path (Settings UI) | bundled | Custom path to the Codex CLI executable |
CODEX_PATH (env var) |
(unset) | Alternative way to set the path |
🛠️ MCP Tools
start_council: Open a new session with a request (returnssession_id).join_council: Join a specific session viasession_id.get_current_session_data: Poll a specificsession_id(supports cursors).send_response: Submit feedback to a specificsession_id.close_council: End a specificsession_idwith a conclusion.summon_agent: Summon Claude or Codex into the current council.
🗺️ Roadmap
Unchecked items are planned and may change based on feedback.
- [x] v0.1 - MCP Council
- [x] v0.2 - Chat UI
- [x] v0.3 - Summon Claude
- [x] v0.4 - Summon Codex
- [ ] v0.5 - Summon Gemini
- [ ] v0.6 - Multiple council sessions in parallel
- [ ] v0.7 - Connect to external LLMs via API Keys
- [ ] v0.8 - Agents can summon user (Telegram/Slack)
📄 License
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。