Bringing your own coding agent (Claude Code, Codex, Cursor, OpenCode, Kimi Code, etc.) into custom tooling is pretty common now. What still bothered me about most agent-to-IM bridges was the cloud relay: if a third-party server sits between your chat apps and your agent, your private work conversations pass through someone else's infrastructure—bringing both the risk of a server breach and the risk of your data being logged or sold.
So I built Mingle (MIT-licensed)—a zero-server desktop app that lets you bring your local coding agents directly into the chat apps you already use: Slack, Discord, Telegram, iMessage, WeChat, Feishu/Lark, WeCom, DingTalk, QQ, and GitHub Issues/PRs.
#
Why zero-server matters
Mingle has no backend service at all:
- No Mingle server, no accounts, no telemetry.
Direct platform connections : connects straight to each chat platform's official bot interface (or your local Apple ID for iMessage)—never routed through a hosted Mingle relay. #
Localhost-only surface : the local console and API bind to127.0.0.1 only. #
OS keychain for secrets : bot tokens and credentials stay inside macOS Keychain, Windows DPAPI, or Linux Secret Service. #
Plain-file memory : each agent's persona and memory are ordinary files in a local folder (ROLE.md ,MEMORY.md , andnotes/ ). #
Optional 24/7 mode on your own hardware : if you want your agent online when your laptop sleeps, you can deploy the runtime to your own SSH host directly from the app—still your machine, no middleman.
#
What makes it usable in real chats
Connecting a CLI agent to a webhook is easy to demo; making it behave well in daily group chats took most of the real engineering:
One agent, unified memory across chat apps The same agent can join your Slack workspace, Discord server, and Telegram/WeChat groups while sharing one continuous memory. You can ask it in one app what was decided in another, or ask it to forward a summary across platforms.
2.
Parallel side tasks without blocking chat When you hand it heavy work—writing code, investigating a bug, running long shell commands—it forks the work into a background task (opening native threads in Slack, Discord, and Feishu) while staying responsive in the main room.
3.
Group-chat etiquette It listens passively, responds when addressed by name or nickname without needing an@ on every message, and automatically throttles itself in busy channels instead of spamming every turn.
4.
14 coding-agent harnesses + one-click session handoff Supports Claude Code, Codex, OpenCode, Kimi Code, Cursor, and 9 more harnesses via OAR. You can also hand off an existing session's context and preferences into a Mingle agent in one click.
#
Honest current status
Mingle started about a week ago and is currently at v0.4.8—definitely early:
End-to-end verified harnesses : Claude Code, Codex, OpenCode, and Kimi Code. #
Chat platforms : Feishu is battle-tested on a live tenant; the other connectors are currently verified against platform APIs and automated tests. #
Installers : macOS (Apple-signed & notarized), Windows (unsigned for now), Linux, plus a one-line install script.
#
Links
If you try it out, I'd love feedback on the zero-server architecture, the SSH deployment flow, or any quirks you hit with a specific chat platform—feel free to drop a comment here or open an issue on GitHub.