agent-channels

Talk to your CLI agents in Discord, get the work done

Discord bridge + MCP tools for CLI coding agents. When someone @mentions your bot, a small bridge wakes a warm ACP agent session; the bridge posts the reply as your bot (default), or the agent can use Discord MCP tools in legacy mode.

Supported agents (any CLI that speaks ACP works — set CDC_AGENT_BIN):

Discord  →  bridge  →  ACP agent (cursor / kimi / devin)  →  bridge posts reply (default)
                              ↳ or Discord MCP when CDC_BRIDGE_OUTBOUND=mcp

Uses your agent subscription — no separate API bill. No VPS required to get started.

Choose your setup

Related MCP server: agy-discord-mcp

Quick start (local)

git clone https://github.com/lilyzhng/agent-channels.git
cd agent-channels
npm install

1. Create a Discord bot → save token to ~/.cursor/channels/discord/.env

2. agent login

3. export CURSOR_CWD=/path/to/your/project and export CDC_STATE_DIR=~/.cursor/channels/discord

4. npm run bridge — or bash scripts/start-bridge-local.sh for background

Full walkthrough: docs/LOCAL_SETUP.md

Using Kimi Code (Kimi K3)

uv tool install --python 3.13 kimi-cli
kimi login                      # ACP mode needs the OAuth login, not an API key
export CDC_AGENT_BIN=$(which kimi)
export CDC_AGENT_MODE=acp

Set the model in ~/.kimi/config.toml (default_model = "k3", provider type = "kimi", base_url = "https://api.kimi.com/coding/v1"). Keep default_thinking = false for fast turns. The bridge auto-approves the agent's permission prompts by ACP option kind, so Kimi's tool calls run unattended just like Cursor's.

What you get

• @mention your bot → real reply in the thread

• Subscription auth — same quota as Cursor IDE / CLI

• Access control — pairing, allowlists, trusted bots

• Self-hosted — your token, your rules

Benchmarks

Measured on a live server against the real agents. The warm ACP session (v0.2) cuts steady-state latency from ~10 to 12s (cold spawn per message) to ~1.2 to 2.5s, roughly 5 to 10x, and adds shared cross-channel memory.

We also raced three models on one full agentic task (read PR #5 → author an SVG → render with rsvg-convert → post the image back to Discord):

Composer 2.5 won on speed and held up on quality; Opus produced the richest-looking diagram. Full writeup, the three diagrams, and the shell-harness root-cause fix: docs/BENCHMARKS.md.

More help

• docs/DISCORD_BOT.md — create a bot in Discord Developer Portal (start here if new)

• docs/LOCAL_SETUP.md — Mac/laptop + nohup (recommended first)

• docs/VPS_SETUP.md — systemd, multi-agent on a server

• docs/AUTH.md — subscription vs API key

• docs/BENCHMARKS.md — warm-bridge speedup + cross-model agentic race

• skills/discord-access/SKILL.md — who can talk to the bot

License

MIT