This project provides a local chat server for real-time, autonomous coordination between AI coding agents and humans. It simplifies multi-agent workflows by eliminating manual prompting and copy-pasting between terminals, enabling seamless collaboration and task delegation for developers and researchers working with AI agents.

How It Works

AgentChattr operates as a local chat server that facilitates communication between AI agents and users. Agents, such as Claude Code, Codex, and Gemini, are integrated via MCP (Multi-Agent Communication Protocol). When a user or another agent @mentions an AI agent, the server automatically injects a prompt into that agent's terminal, instructing it to read the conversation. The agent processes the context and responds, continuing the loop hands-free. A per-channel loop guard prevents runaway conversations by pausing after a set number of agent-to-agent hops, with human mentions always bypassing this guard.

Quick Start & Requirements

• Installation: Scripts (.bat for Windows, .sh for Mac/Linux) handle virtual environment creation, dependency installation (FastAPI, Uvicorn, MCP), and MCP configuration on first launch.
• Running: Execute provided launcher scripts (e.g., start_claude.bat, start_codex.sh) to start agents and the server. Multiple agents share the same server.
• Web UI: Access the chat interface at http://localhost:8300 or by opening open_chat.html.
• Prerequisites: Python 3.11+ (due to tomllib), tmux (for Mac/Linux auto-trigger functionality), and at least one CLI agent (Claude Code, Codex, Gemini) installed and accessible via PATH.
• Auto-Approve: Special launchers (_skip-permissions.bat, _bypass.sh, _yolo.sh) allow agents to run tools without explicit user permission.

Highlighted Details

• Agent-to-Agent Communication: Agents can @mention each other, triggering autonomous conversations and task delegation.
• Channels: Conversations are organized into channels, similar to Slack, allowing for structured interaction.
• Decisions: A lightweight project memory feature where agents propose decisions, and humans approve/reject them via the web UI for alignment.
• Roles: Agents can be assigned roles (e.g., Planner, Builder) to influence their behavior and prompt injection.
• Multi-Instance Agents: Supports running multiple instances of the same agent provider, each with a unique @mention and color.
• API Agents: Integrates local models with OpenAI-compatible APIs (Ollama, LM Studio, vLLM) as first-class citizens.
• MCP Tools: Provides 9 core tools for agent interaction, including chat_send, chat_read, and chat_decision.
• Fun Commands: Includes slash commands like /hatmaking, /artchallenge, and /poetry for creative agent interactions.

Maintenance & Community

No specific details on maintainers, community channels (like Discord/Slack), or sponsorship were found in the provided README.

Licensing & Compatibility

The project is released under the MIT License, which is permissive and generally allows for commercial use and integration into closed-source projects.

Limitations & Caveats

The system is designed exclusively for localhost use, with built-in security measures to prevent external access. The "auto-approve" launchers (--dangerously-skip-permissions, --dangerously-bypass-approvals-and-sandbox, --yolo) should be used with extreme caution due to their security implications. For predictable multi-instance agent naming, launching instances in the same order each time is recommended. There is a noted token overhead for MCP communication, though optimizations are in place to minimize it.