Bridge Discord to Claude Code / Codex / Gemini CLI — zero agent tokens, zero noise.
CLI Bridge 把 Discord 命令直连本地三大 AI CLI（Claude Code、Codex、Gemini），不消耗 agent token，不经过 AI 润色。你在 Discord 里发
/cc 帮我查个 bug、/codex 优化这段代码、/gemini 解释这个概念，任务直接到你本机对应的 CLI，结果由 worker 回调到 Discord 频道。
Running AI coding CLIs locally is powerful, but controlling them from your phone or another device means SSH tunnels, token juggling, or leaving a terminal open. OpenClaw bots can invoke tools, but every round-trip costs agent tokens — and the bot "helpfully" rephrases your CLI output.
本地跑 AI 编程 CLI 很强，但想从手机远程操控就得折腾 SSH 隧道或者开着终端。OpenClaw bot 虽然能调工具，但每次都消耗 agent token，而且 bot 还"贴心地"把 CLI 输出重新润色一遍。
CLI Bridge registers Discord commands (/cc, /codex, /gemini, plus subcommands) via OpenClaw's registerCommand API. Commands are handled directly by the plugin — no agent dispatch, no token cost, no AI rewriting. Results are pushed back to Discord via openclaw-worker's callback mechanism.
CLI Bridge 通过 OpenClaw 的
registerCommandAPI 注册 Discord 命令。命令由插件直接处理——不经过 agent、不消耗 token、不被 AI 改写。结果通过 openclaw-worker 的回调机制推回 Discord。
Discord OpenClaw Bot CLI Bridge Plugin openclaw-worker AI CLI
| | | | |
| /cc fix the bug | | | |
|-------------------->| registerCommand match | | |
| |------------------------->| POST /claude | |
| | |---------------------->| spawn Claude Code |
| | | |-------------------->|
| | | | ...working... |
| | | |<--------------------|
| | callback to Discord | | |
|<--------------------------------------------------------------------- | |
| | | | |
| /codex optimize it | | | |
|-------------------->|------------------------->| POST /codex | |
| | |---------------------->| spawn Codex CLI |
| | | |-------------------->|
| | | |<--------------------|
|<--------------------------------------------------------------------- | |
| | | | |
| /gemini explain X | | | |
|-------------------->|------------------------->| POST /gemini | |
| | |---------------------->| spawn Gemini CLI |
| | | |-------------------->|
| | | |<--------------------|
|<--------------------------------------------------------------------- | |
- 所有命令由
registerCommand拦截，不进入 agent 处理流程（shouldContinue: false）- 任务通过 HTTP 提交给 openclaw-worker 的 Task API（
/claude、/codex、/gemini三条路径）- CLI 完成后，worker 直接用 Discord Bot API 推结果到指定频道
- 全程零 agent token 消耗
| Command | Description |
|---|---|
/cc <prompt> |
Submit a task to Claude Code (auto-continues last session) |
/cc-new |
Reset CC session (next /cc starts fresh) |
/cc-new <prompt> |
Reset CC session and immediately submit |
/cc-recent |
List recent CC sessions |
/cc-now |
Show current CC session ID |
/cc-resume <id> <prompt> |
Switch to a specific CC session and continue |
/codex <prompt> |
Submit a task to OpenAI Codex CLI (auto-continues last session) |
/codex 新会话 |
Reset Codex session |
/codex 接续 <id> <prompt> |
Switch to a specific Codex session |
/gemini <prompt> |
Submit a task to Google Gemini CLI (auto-continues last session) |
/gemini 新会话 |
Reset Gemini session |
/gemini 接续 <id> <prompt> |
Switch to a specific Gemini session |
连续发同一命令自动在同一会话里继续对话，不需要手动带 session ID。想换话题就用
/xx 新会话。
CLI Bridge also registers tools for other channel agents to call programmatically:
| Tool | Description |
|---|---|
cc_call |
Submit a task to Claude Code — result delivered via callback, not through the agent |
codex_call |
Submit a task to Codex CLI — result delivered via callback |
gemini_call |
Submit a task to Gemini CLI — result delivered via callback |
这些工具供其他频道的 agent 使用。agent 调用后结果由 worker 回调直推 Discord，不经过 agent 润色。
- OpenClaw bot running (Docker recommended)
- openclaw-worker running locally with Task API enabled
- At least one of: Claude Code, Codex CLI, or Gemini CLI installed locally
作者环境：MacBook Air M4 (16GB) · Docker 运行 OpenClaw bot · 本地 openclaw-worker 提供 Task API · Claude Code (Max subscription) + Codex CLI + Gemini CLI · Bun 运行时
- Clone into your OpenClaw extensions directory:
cd ~/.openclaw-<your-bot>/extensions/
git clone https://github.com/AliceLJY/openclaw-cli-bridge.git cli-bridge- Configure in
openclaw.json:
Add the plugin to plugins.allow and configure under plugins.entries:
{
"plugins": {
"allow": ["cli-bridge"],
"entries": {
"cli-bridge": {
"enabled": true,
"config": {
"apiToken": "your-task-api-token",
"callbackChannel": "your-discord-channel-id"
}
}
}
}
}| Config Key | Required | Description |
|---|---|---|
apiToken |
Yes | Auth token from your openclaw-worker config |
callbackChannel |
Yes | Discord channel ID where results are delivered |
apiUrl |
No | Task API URL (default: http://host.docker.internal:3456) |
discordBotToken |
No | Bot token for callback delivery (uses bot's own token if omitted) |
- Restart your OpenClaw bot. The plugin auto-registers on startup.
apiUrl默认指向host.docker.internal:3456（Docker 容器访问宿主机）。如果 worker 不在本机，改成对应地址。
OpenClaw's registerCommand API registers text commands that are handled before the agent processes the message. When a command matches, the framework returns shouldContinue: false — the agent never sees the message. This means:
- Zero token cost — no LLM inference for command handling
- Zero noise — no AI rephrasing of CLI output
- Instant response — no agent thinking time
registerCommand注册的命令在 agent 之前处理。匹配到命令后框架返回shouldContinue: false，agent 根本看不到这条消息。
OpenClaw's matchPluginCommand splits command name from arguments by space. This means /cc-recent works (it's a registered command name), but /cc最近 (no space) would fail to match /cc and fall through to the agent.
Solution: All subcommands use independent ASCII names (cc-recent, cc-now, cc-new, cc-resume) following the pattern established by HappyClaw.
框架用空格分割命令名和参数。
/cc最近（连写、无空格）匹配不到/cc，会穿透给 agent。解决方案：子命令用独立的 ASCII 命名，学自 HappyClaw 的设计模式。
HappyClaw is a PTY bridge — it multiplexes local terminal sessions and streams I/O to OpenClaw. CLI Bridge takes a completely different approach:
| HappyClaw | CLI Bridge | |
|---|---|---|
| Mechanism | PTY terminal multiplexing | HTTP Task API |
| Scope | Any CLI tool via PTY | Claude Code + Codex + Gemini via dedicated endpoints |
| Session model | Persistent PTY processes | Stateless API calls with per-channel session tracking |
| Unique feature | Real-time terminal streaming | Auto-session continuation (just keep sending /cc, /codex, /gemini) |
| Design | General-purpose terminal bridge | Purpose-built for AI CLI + Discord workflow |
CLI Bridge was born from a real daily workflow: controlling AI CLIs from Discord on a phone while away from the desk. The auto-session continuation (no manual session ID juggling) and dual-mode design (slash commands for humans, *_call tools for agents) came directly from dogfooding the tool.
CLI Bridge 诞生于真实的日常场景：离开电脑时用手机 Discord 远程操控本地 AI CLI。自动会话续接（不用手动带 session ID，连着发就行）和双模式设计（命令给人用、
*_call工具给 agent 用）都是在实际使用中磨出来的。HappyClaw 教会了我registerCommand的正确姿势，而 CLI Bridge 是在这个基础上针对 AI CLI + Discord 场景的独立实现。
This section originates from openclaw-cli-pipeline (now archived). The protocol is fully implemented in this plugin's session management.
多轮编排协议——原始设计来自 openclaw-cli-pipeline（已归档），现在由本插件的会话管理完整实现。
The core insight: single-dispatch is not enough for complex tasks. A typical content pipeline needs 3 rounds with human review between each. The protocol preserves full AI CLI context across rounds via sessionId.
Round 1 (New task):
/cc "Stage 1: mine topics for X"
→ Plugin submits to Task API (auto-generates sessionId)
→ Worker invokes: claude --print --session-id <id> "prompt"
→ Worker delivers result to Discord (callback with sessionId)
→ User reviews output, provides feedback
Round 2+ (Resume):
/cc "Continue to Stage 2. User chose angle: Y"
→ Plugin reuses sessionId from Round 1 (auto-continuation)
→ Worker invokes: claude --print --resume --session-id <id> "prompt"
→ AI CLI resumes with full context from all previous rounds
Each round preserves full AI CLI context. The user has explicit control between rounds — the plugin never auto-advances.
OpenClaw natively supports MAS via sessions_spawn — multiple agents working in parallel. We retired MAS in favor of this pipeline:
OpenClaw 原生支持 MAS 多 agent 协作，但我们已经用 CLI Pipeline 替代了它。
| MAS (Multi-Agent) | CLI Pipeline | |
|---|---|---|
| How it works | Main agent spawns sub-agents inside OpenClaw | Plugin dispatches tasks to external AI CLI |
| Context cost | Every sub-agent's full output stays in main agent's context | Only result summaries flow back to Discord |
| Human control | Agents auto-advance, user is passive | User confirms each round before next one starts |
| Agent count | N agents x full context each | 1 AI CLI x full context, summaries back |
Bottom line: For creative tasks (writing articles), the lighter single-CLI approach wins.
简单说：MAS 是 N 个 agent 各背一套上下文，CLI Pipeline 是 1 个 AI CLI 背上下文、只回传摘要。
All bugs discovered in a single evening of testing. 以下所有坑都是一个晚上踩出来的。
| Pitfall | Symptom | Fix |
|---|---|---|
| Timeout needs buffer | CLI ran 300,322ms, timeout was 300,000ms. Killed by 322ms. | Add 30s buffer. Default to 600s (10 min), never 300s. |
| SessionId must auto-generate | Round 2 --resume exits in 596ms — session doesn't exist. |
Task API auto-generates sessionId when not provided. |
| Callback must have retry | Bot container was restarting during callback. Result lost. | Worker retries 3 times, 5s apart. |
| DM channels don't work | Callback always fails in Discord DMs. | Only use server (guild) channels for callback. |
| Sessions are directory-scoped | Worker creates session from $HOME, manual CLI from another dir can't find it. |
Always start AI CLI from the same working directory. |
| Don't clean results too fast | 5-min cleanup timer deleted result before Worker could fetch. | Separate timers: 15 min (incomplete), 30 min (completed). |
This plugin is part of a toolchain for controlling AI coding CLIs from Discord and Telegram:
| Project | Role |
|---|---|
| content-alchemy | 5-stage content pipeline — a primary use case driven through CC |
| content-publisher | Image generation, layout formatting, and WeChat API publishing |
| openclaw-worker | Task API + Docker compose for OpenClaw |
| digital-clone-skill | Build digital clones from corpus data |
| local-memory | Local AI conversation search (LanceDB + Jina) |
| cc-shell | Lightweight Claude Code chat UI |
| telegram-ai-bridge | 3 Telegram bots for Claude / Codex / Gemini via SDK |
| telegram-cli-bridge | Telegram CLI bridge (Gemini CLI path) |
- HappyClaw by @rwmjhb — the PTY bridge plugin that pioneered the
registerCommandpattern for OpenClaw. CLI Bridge's architecture (ASCII subcommands, zero-token command handling, direct callback delivery) was directly inspired by studying HappyClaw's source code. Thank you for sharing! - Built for the OpenClaw ecosystem
- Uses openclaw-worker Task API for local AI CLI integration
特别感谢 @rwmjhb 的 HappyClaw 项目。CLI Bridge 的架构设计（ASCII 子命令、零 token 命令处理、回调直推）直接学自 HappyClaw 的源码。开源社区的知识共享让每个人都能站在前人的肩膀上。
Built by 小试AI (@AliceLJY) · WeChat: 我的AI小木屋
医学出身，文化口工作，AI 野路子。公众号六大板块：AI实操手账 · AI踩坑实录 · AI照见众生 · AI冷眼旁观 · AI胡思乱想 · AI视觉笔记
Six content pillars: Hands-on AI · AI Pitfall Diaries · AI & Humanity · AI Cold Eye · AI Musings · AI Visual Notes
Open-source byproducts: content-alchemy · content-publisher · openclaw-worker · openclaw-cli-bridge · digital-clone-skill · local-memory · cc-shell · telegram-ai-bridge · telegram-cli-bridge
v0.4.0 by Claude Code (Opus 4.6)