cc-connect: Plugging Claude Code into Feishu

Today I installed a new agent bridge — cc-connect. It maps your local Claude Code to a Feishu bot, wiring it into a chat app. Once it’s connected, each bot maps to one working directory, which is basically like opening Claude Code inside that folder. It’s been a great experience, so here’s a write-up.

First try with cc-connect

Installation is dead simple — just follow the official docs. It doesn’t support Copilot, but it does support Claude Code and OpenCode, which is basically enough.

请参考 https://raw.githubusercontent.com/chenhg5/cc-connect/refs/heads/main/INSTALL.md 帮我安装和配置 cc-connect

After installing, you need to manually add a startup item with cc-connect daemon install, and then you can call it from Feishu.

cc-connect itself doesn’t configure any LLM — it only bridges messages. The LLM it calls is whatever Agent CLI (Claude Code or OpenCode) you write in the config file.

The key advantage: one bot per project directory. You can isolate by project and work on several projects at once from outside using multiple bots.

Permission control: fully in your hands

The thing I cared about most upfront was the permission approval flow. Once it’s hooked up to push messages, will it ask for confirmation at every single step?

After trying it, permissions are fully in your hands. It works exactly like using Claude Code directly on the command line — every confirmation prompt that should appear does appear. Those permission settings live in Claude’s config file, so it’s a white box.

So it’s really worth considering cancelling Hermes.

Configuration and managing multiple projects

cc-connect’s config files live under ~/.config/cc-connect/, with one subdirectory per project (because it’s not just config files — there are temp data files too).

One instance is enough — a single instance can keep multiple projects separate and run commands in parallel, so several projects can run at once. Multiple instances running simultaneously are supported too.

For multi-project management, one Feishu bot per project: tasks run in parallel, which is efficient.

My choice: two bots, split by purpose

• Coding bot: multiple directories configured, using OpenCode (compatible with Copilot’s agents files)
• Journal/knowledge base bot: uses Claude Code with a cheap, high-volume model (minimax), switching to deepseek when it expires

Feishu quick commands

cc-connect supports a rich set of slash commands. I put a few into the Feishu shortcut menu:

• /new - new session
• /model [name] - view/switch model
• /reasoning [level] - view/switch reasoning effort
• /quiet - mute toggle
• /stop - stop the current run

Typing /model returns Feishu’s interactive selection menu directly, without going through the Agent CLI or the LLM.

Common model switches:

• minimax m2.7: /model minimax-cn-coding-plan/MiniMax-M2.7
• deepseek flash: /model deepseek/deepseek-v4-flash

cc-connect vs Hermes: each has its strengths

From my experience, the two are positioned completely differently:

Conclusion:

• Tasks like coding that need remote control and frequent edits → cc-connect
• Things like the library and the journal that need to call bash frequently → Hermes running from Docker is safer

Downside: cc-connect constantly asking for approval gets tiring. But waving everything through risks security problems — say, when it’s collecting data and spins up a headed browser that gets in the way of what I’m doing. The isolation of a Docker environment, Hermes still feels more reassuring.

opencode vs Claude Code

Finally I gave opencode a try. It has more models, but no superpower. In the end I switched back to Claude Code.

Come to think of it, the one with multi-agent collaboration is still Claude Code + Superpower — even GitHub Copilot hasn’t opened up multi-agent.

The most cost-effective setup for me going forward is probably deepseek pro + Claude Code.

So right now I have three tiers:

1. Out and about: use cc-connect to direct Claude Code — I state the requirements, the AI writes the code
2. At home: use Claude Code — I state the requirements, the AI writes the code
3. Tricky problems: use Copilot to write the code + WebStorm to review.