ACP Bridge — HTTP gateway for multi-agent orchestration over ACP

[xiwan]

ACP Bridge

I built ACP Bridge, an open-source Python/uvicorn gateway that exposes local ACP-compatible CLI agents as HTTP services, with multi-agent orchestration and IM-driven async workflows.

What it does

• HTTP → ACP: Incoming REST calls (/runs) are routed to agent handlers over stdio JSON-RPC, managed by a process pool with LRU eviction and OOM protection
• Multi-agent pipelines: /pipelines endpoint supports sequence, parallel, race, and conversation modes across different agents
• Async jobs: /jobs submits background tasks with webhook callbacks to Discord/Feishu via OpenClaw or any HTTP endpoint
• Zero-config: Auto-detects installed agent CLIs in PATH — no config file needed to get started

Supported ACP agents

Agent	Mode
Kiro CLI	ACP (native)
Claude Code	ACP (via adapter)
OpenClaw	ACP (openclaw acp)
Qwen Code	ACP
OpenCode	ACP
Hermes Agent	ACP
Harness Factory	ACP
Codex CLI	PTY fallback
Trae Agent	PTY fallback

Use case

One EC2 instance runs multiple agents; the team (or bots) calls them via HTTP. No local agent install needed for end users. IM platforms like Discord send a message → agent executes → result pushed back to chat.

Links

• Repo: https://github.com/xiwan/acp-bridge
• Docs: https://github.com/xiwan/acp-bridge/tree/main/docs
• AWS Blog: https://aws.amazon.com/cn/blogs/china/enable-kiro-and-claude-code-for-im-with-acp-bridge-async-ai-workflow/

Feedback and contributions welcome!

[jingchang0623-crypto]

这玩意儿解决了我的"半夜惊魂"问题

凌晨2点43分，我的5个Agent在Discord群里吵了起来。

事情是这样的：我用OpenClaw跑了一个cron任务链，Agent A要等Agent B的结果，Agent B在等MCP server的响应，然后MCP server挂了。于是整个链条开始疯狂重试，Discord收到了300+条错误消息。

我醒来看到手机时，以为公司被黑了。

ACP Bridge这个东西，本质上是在说：HTTP是Universal Adapter。

这让我想起我写过的一篇踩坑实录：MCP协议的黑暗面 —— MCP很好，但它不是银弹。你依然需要：

• Circuit breaker（熔断器）
• Job queue（任务队列）
• Webhook callbacks（异步回调）

而ACP Bridge把这些都包进去了。Zero-config 这个feature是最骚的 —— 因为我之前花了两天配置MCP server，结果发现忘记export PATH。

一个小建议：
我在MCP vs A2A：AI界的USB vs蓝牙之争里提到过，协议战的本质是"谁掌握入口"。ACP Bridge的聪明之处在于：它不站队，它只是把所有Agent都变成HTTP endpoint。

这样你就不需要：

• ❌ 等Anthropic开放Claude Code API
• ❌ 等OpenAI开放Codex WebSocket
• ❌ 等Google开放Agent orchestration

你只需要：curl http://localhost:8080/runs

P.S. 我也在尝试用类似思路做一个多Agent调度系统，但我的版本是让Agent们"开视频会议"（其实是LLM chain of thought共享），结果它们聊了8小时没出结果 —— 多Agent会议地狱

👍 for this project!