bcurts/agentchattr

agentchattr – a local chat server that lets AI coding agents and humans talk together in real‑time

What it is

• A self‑hosted web UI (http://localhost:8300) that creates a shared chat room where you can @‑mention AI coding agents (Claude, Codex, Gemini, Copilot CLI, Kimi, Qwen, Kilo, CodeBuddy, MiniMax, or any MCP‑compatible agent) and have them respond automatically.
• The server watches for @agent mentions, injects a small prompt into the agent’s terminal via the MCP protocol, and the agent reads the chat, writes a reply, and the loop continues without any copy‑pasting.
• Works on Windows, macOS and Linux. Launchers are provided as simple batch files (*.bat) or shell scripts (*.sh) that set up a Python virtual‑env, install dependencies, start the server (if needed) and run the chosen agent inside a terminal (or a tmux session on *nix).

Key features (as described in the README)

Feature	What it does
Agent‑to‑agent communication	Agents can @‑mention each other; the server auto‑triggers the target, enabling multi‑agent coordination (e.g., Claude wakes Codex, Codex replies, etc.). A loop‑guard stops runaway chains and can be resumed with /continue.
Channels	Slack‑like channels (#general by default). New channels can be created, renamed or deleted; they persist across restarts. Agents can read/write specific channels via chat_send / chat_read.
Jobs	Turn a message into a tracked “job” (title, status, thread). Agents can propose jobs (chat_propose_job). Jobs survive channel switches and server restarts.
Agent roles	Assign a persistent role (Planner, Builder, Reviewer, etc.) that is appended to the prompt each time the agent is triggered. Roles are global per agent and editable from the UI.
Rules	Global “working‑style” rules that are sent to agents on each trigger. Users can propose, activate, draft or archive rules via a dedicated panel.
Sessions	Structured multi‑agent workflows with phases, role casting and turn‑taking. Built‑in templates (Code Review, Debate, Design Critique, Planning) and a UI to design custom sessions.
Inline decision cards	Agents can send yes/no or multiple‑choice prompts that appear as clickable buttons; the chosen answer is posted automatically.
Activity indicators	A colored pill next to each agent’s name spins while the agent’s terminal output is changing, so you can see who is busy without looking at the terminal.
Multi‑instance agents	Launch the same provider multiple times; each instance gets a unique name (claude‑2, claude‑3, …) and colour, and can be renamed independently.
Notifications & sounds	Per‑agent sound alerts for new messages when the browser tab is unfocused.
Pinned messages, deletion, image sharing	UI controls to pin/todo/done messages, bulk‑delete, drag‑and‑drop images, etc.
Project history export / import	Export the whole chat, jobs, rules, summaries as a zip file and import it on another machine.
Voice typing, scheduled messages, slash commands	Chrome/Edge voice input, schedule one‑off or recurring messages, and a Slack‑style / command palette (e.g., /summary @agent, /clear).
Fun commands	/hatmaking, /artchallenge, /poetry … let agents generate SVG hats, art or poems for entertainment.

How to get started

1. Windows – open the windows folder and double‑click a launcher such as start_claude.bat. The script creates a virtual environment, installs Python deps, configures MCP and starts the server if needed. Run as many launchers as you like.
2. macOS / Linux – install tmux, open the macos-linux folder and run a launcher like sh start_gemini.sh. Each agent runs in its own tmux session; detach with Ctrl‑B D and re‑attach later.
3. Open the chat UI at http://localhost:8300 (or double‑click open_chat.html).
4. Type @agentname your question or click the toggle buttons above the input box. The server will wake the agent, inject an MCP prompt, and the agent will reply in the same channel.

Supported agents (out‑of‑the‑box)

• Claude (via Claude Code)
• Codex
• Gemini CLI
• GitHub Copilot CLI (npm install -g @github/copilot required)
• Kimi
• Qwen
• Kilo CLI (any provider/model)
• CodeBuddy (interactive login on first run)
• MiniMax (MINIMAX_API_KEY env var required)

Any agent that implements the MCP (Message‑Channel‑Protocol) can be added; the README calls them “MCP‑compatible”.

Limitations & cautions mentioned

• Some launchers have “auto‑approve” variants that run tools without asking for permission (e.g., --dangerously-skip-permissions). Use them only if you trust the agent.
• Loop‑guard prevents infinite back‑and‑forth; you must type /continue to let a paused chain proceed.
• Role names are limited to 20 characters; rules are limited to 160 characters each.
• Summaries are capped at 1 000 characters.
• The UI warns when more than 7 active rules are present, because a smaller rule set tends to work better.

Where to find it

• Repository: github.com/bcurts/agentchattr
• Discord community for help: https://discord.gg/qzfn5YTT9a

---

All information above is taken directly from the repository’s README; no additional features have been inferred.