acp-to-agui

KO | EN

gitlite — search

#typescript #ai-agents #ai #dsh-plugin #deepseek-harness #open-source #claude-code #codex #cli #developer-tools #react #windows

★ 19 Open GitHub ↗

A protocol bridge that sits between any coding agent(supporting ACP) and any web frontend

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<h1 align="center"><a href="https://agentclientprotocol.com">ACP</a> → <a href="https://docs.ag-ui.com">AG-UI</a></h1>

 <a href="https://agentclientprotocol.com">agentclientprotocol.com</a> &nbsp;→&nbsp; <a href="https://docs.ag-ui.com">docs.ag-ui.com</a>

A reference bridge that gives any ACP coding agent (Kiro, Claude, Codex, etc.) a web UI via the AG-UI protocol.

<img src="docs/assets/reference-ui-over-kiro.gif" alt="Reference UI powered by Kiro CLI" width="49%"/>
 <img src="docs/assets/reference-ui-over-claudecode.gif" alt="Reference UI powered by Claude Code" width="49%"/>


 Left: Kiro CLI powering the workspace. Right: same UI with Claude Code. Just swap the agent.

<img src="docs/assets/ui-toggle.gif" alt="Agent toggle on project selector" width="500"/>


 Select your agent from the frontend. No restart needed. Each session spawns the chosen binary.

---

## The Problem

There are now **33+ coding agents** that support the [Agent Client Protocol (ACP)](https://agentclientprotocol.com): Kiro, Claude Code, Codex CLI, Cursor, Gemini CLI, GitHub Copilot, OpenCode, Cline, and many more. They all speak JSON-RPC 2.0 over stdio. You can use them in terminals. You can use them in editors.

But teams keep needing **purpose-built web interfaces** on top of these agents: a micro-app creation platform where non-engineers iterate on device experiences, a report generation UI where analysts interact with agents without learning CLI commands, deployment dashboards with team-visible approval flows, domain-specific editors powered by an agent underneath.

Today, building any of these means implementing the protocol bridge yourself: parsing JSON-RPC streams, managing subprocesses, translating events into something a web frontend can render.

## The Solution

This project is a **reference bridge** that sits between any ACP agent and any web frontend:

```mermaid
graph TB
    subgraph agents["ACP Agents (33+)"]
        A1["kiro-cli"] ~~~ A2["claude-agent-acp"] ~~~ A3["codex"] ~~~ A4["gemini cli"] ~~~ A5["cursor"] ~~~ A6["ANY ACP binary"]
    end

agents <-->|"JSON-RPC 2.0 over stdio"| bridge

subgraph bridge["This Bridge (Python / FastAPI)"]
        B1["AgentRunner"] ~~~ B2["AcpToAguiBridge"] ~~~ B3["SessionManager"]
    end

bridge -->|"AG-UI Events over SSE"| frontends

subgraph frontends["Your Frontend"]
        F1["Reference UI"] ~~~ F2["CopilotKit"] ~~~ F3["AG-UI HttpAgent"] ~~~ F4["Anything"]
    end

style agents fill:#1e3a5f,stroke:#60a5fa,color:#fff
    style bridge fill:#3b1f6e,stroke:#a78bfa,color:#fff
    style frontends fill:#1a4731,stroke:#6ee7b7,color:#fff
```

Clone this repo, select your agent, and you have a working web UI with streaming chat, tool visualization, and human-in-the-loop approvals.

## Use Cases

Same agent underneath. Different frontend for each audience:

- **Micro-app creation platforms**: non-engineers iterating on apps through a web workspace
- **Report generation**: analysts interacting with agents without CLI commands
- **Deployment dashboards**: team-visible approval flows for infrastructure changes
- **Domain-specific IDEs**: focused editors (config, pipelines) powered by an agent
- **Any internal tooling**: the agent does the heavy lifting, users get a tailored UI

## Why AG-UI?

[AG-UI](https://docs.ag-ui.com) (Agent-User Interaction Protocol) is the open standard for connecting AI agents to frontends. Instead of rolling your own SSE/WebSocket protocol, you get:

- **~16 standard event types**: streaming chat, tool calls, state sync, generative UI, interrupts
- **Transport agnostic**: works over SSE, WebSockets, or webhooks
- **Rich ecosystem**: supported by CopilotKit, LangGraph, Google ADK, AWS Strands, Pydantic AI, and 20+ frameworks
- **Frontend SDKs**: TypeScript, Python, Kotlin, Go, Rust, and more
- **Human-in-the-loop built in**: pause, approve, reject, or redirect agent execution mid-flow

By emitting AG-UI events, your frontend becomes portable across the entire agent ecosystem. See [`docs/why-agui.md`](docs/why-agui.md) for a deep dive on what this unlocks: CopilotKit integration, shared state, generative UI, and more.

## Quick Start

Works on macOS, Linux, and Windows (PowerShell or cmd). Prereqs: Python 3.11+, Node.js 18+, pnpm.

```bash
git clone https://github.com/namanrajpal/acp-to-agui.git
cd acp-to-agui
pnpm install
pnpm dev
```

Open **http://localhost:3000**. Select your agent (Kiro, Claude, Codex, or OpenCode), enter a project path, and start chatting. The bridge spawns the agent subprocess, translates its output to AG-UI events, and streams them to the React frontend.

Prefer side-by-side logs? Run each half in its own terminal:

```bash
pnpm dev:backend   # FastAPI on :8000
pnpm dev:ui        # Vite on :3000
```

Both commands run identically on every platform. The launcher resolves `PYTHONPATH` to the repo root, so you can run them from any working directory.

<details>
<summary>Platform notes</summary>

- **macOS / Linux**: `pnpm dev:backend` runs uvicorn with `--reload`, so Python edits hot-reload automatically.
- **Windows**: `--reload` is skipped automatically because uvicorn's reload supervisor installs an asyncio event loop policy that can't spawn subprocesses (you'd hit `NotImplementedError` from `_make_subprocess_transport` the moment you tried to start a session). Restart `pnpm dev:backend` manually after backend edits. The Vite frontend hot-reloads everywhere.
- **Windows + non-`.exe` agents** (`npx`, `claude-agent-acp`, etc.): the bridge auto-wraps `.cmd`/`.bat` shims with `cmd.exe /c` so you can use the same `agentCommand` config as on Unix. Native `.exe` agents like `kiro-cli` pass through unchanged.

</details>

## Configuration

There are two ways to select an agent:

### 1. Frontend selector (per-session)

The workspace UI has an agent toggle on the project selector page. Pick Kiro, Claude, Codex, or OpenCode. the selected agent command is sent with the session creation request and the backend spawns that specific binary.

This means you can switch agents between sessions without restarting anything.

### 2. Default via `bridge.config.json`

The config file sets the fallback agent when no command is specified per-session:

```json
{
  "projectName": "acp-to-agui",
  "displayTitle": "ACP → AG-UI Bridge",
  "agentCommand": ["kiro-cli", "acp"],
  "backendPort": 8000,
  "corsOrigins": ["http://localhost:3000", "http://localhost:3001"]
}
```

### Supported agents

| Agent | Command | Auth |
|-------|---------|------|
| Kiro CLI | `["kiro-cli", "acp"]` | AWS Builder ID |
| Claude Agent | `["claude-agent-acp"]` | `ANTHROPIC_API_KEY` |
| Codex CLI | `["codex-acp"]` | ChatGPT subscription or API key |
| OpenCode | `["opencode", "acp"]` | OpenCode Zen or provider API key |
| Gemini CLI | `["gemini", "cli", "acp"]` | Google auth |
| Cursor | `["cursor", "--acp"]` | Cursor subscription |
| GitHub Copilot | `["github-copilot-cli", "--acp"]` | GitHub auth |
| Goose | `["goose", "--acp"]` | Provider API key |
| Any ACP binary | `["your-agent", "acp"]` | Varies |

### API-level control

You can also pass `agentCommand` directly when creating a session via REST:

```bash
curl -X POST http://localhost:8000/v2/tasks \
  -H "Content-Type: application/json" \
  -d '{"cwd": "/your/project", "agentCommand": ["claude-agent-acp"]}'
```

## Architecture

```mermaid
graph TB
    subgraph frontend["React Frontend (Vite)"]
        direction LR
        CP["ChatPanel"]
        AD["ApprovalDialog"]
        TC["ToolCard"]
        SS["SessionSidebar"]
    end

frontend -->|"SSE stream + REST"| backend

subgraph backend["Python Backend (FastAPI)"]
        SM["SessionManager"]
        BR["AcpToAguiBridge"]
        ST["SessionStore"]
        AR["AgentRunner"]
        PE["PolicyEngine"]
        API["REST APIs"]
        SM --> AR
        SM --> BR
        SM --> ST
        BR --> PE
    end

AR <-->|"JSON-RPC 2.0 / stdio"| agent

subgraph agent["ACP Agent (subprocess)"]
        AG["kiro-cli / claude-agent-acp / any"]
    end

style frontend fill:#1a4731,stroke:#6ee7b7,color:#fff
    style backend fill:#3b1f6e,stroke:#a78bfa,color:#fff
    style agent fill:#1e3a5f,stroke:#60a5fa,color:#fff
```

## Protocol Translation

| ACP Event | AG-UI Event(s) | Notes |
|-----------|---------------|-------|
| `agent_message_chunk` | `TEXT_MESSAGE_START` + `TEXT_MESSAGE_CONTENT` | Opens message on first chunk |
| `tool_call` | `TOOL_CALL_START` + `TOOL_CALL_ARGS` | Closes open text message first |
| `tool_call_update` | `TOOL_CALL_ARGS` or `TOOL_CALL_END` | Based on status field |
| `turn_end` | `TEXT_MESSAGE_END` + `TOOL_CALL_END`(s) + `RUN_FINISHED` | Closes everything |
| `session/request_permission` | `STATE_UPDATE` (approval pending) | Uses asyncio.Future for async bridge |
| Vendor extensions (`_*.dev/*`) | `CUSTOM` events | Normalized to `agent:*` namespace |

See [`docs/protocol-translation.md`](docs/protocol-translation.md) for the full mapping with diagrams.

## The Tricky Parts

ACP and AG-UI do not map one-to-one. These required a normalization layer:

**Tool Approvals:** ACP's SDK calls `request_permission()` and blocks waiting for a return value. But our approval comes asynchronously from a REST endpoint. We bridge this with `asyncio.Future`: the SDK callback awaits the future, the REST endpoint resolves it.

**Message Boundaries:** ACP streams `agent_message_chunk` continuously. AG-UI needs explicit `TEXT_MESSAGE_START` and `TEXT_MESSAGE_END` events. The bridge tracks open message state and auto-closes before tool calls or turn end.

**Vendor Extensions:** ACP agents send custom notifications (e.g., `_kiro.dev/mcp_servers_ready`). The SDK routes these to `ext_notification()`. We normalize them into `CUSTOM` AG-UI events with a clean `agent:*` namespace.

## Three Ways to Build Your Frontend

| Approach | How it consumes AG-UI | Lines of Code | Use When |
|----------|----------------------|--------------|----------|
| **Custom Workspace** (`example-frontends/custom-workspace-ui-demo/`) | Direct REST + raw SSE parsing | ~2000 | You want full control over every pixel |
| **CopilotKit** (`example-frontends/copilotkit-demo/`) | AG-UI via framework (zero UI code) | ~20 | You want to ship fast with production features |
| **AG-UI HttpAgent** (`example-frontends/httpagent-demo/`) | AG-UI via client library (you build UI) | ~50 | You want the raw protocol with your own UI framework |

## How It Works

1. **Select an agent**. via the UI toggle, the `agentCommand` API field, or the `bridge.config.json` default
2. **Create a session**: `POST /v2/tasks` spawns the agent subprocess, initializes ACP
3. **Start a run**: `POST /v2/tasks/{id}/run` sends your prompt via JSON-RPC
4. **Stream events**: `GET /v2/tasks/{id}/events?runId=...` returns AG-UI SSE stream
5. **Or use the standard endpoint**: `POST /ag-ui` (what CopilotKit and AG-UI HttpAgent use)
6. **Handle approvals**: `POST /v2/tasks/{id}/approval` resolves pending tool permissions

<details>
<summary>Project Structure</summary>

```
├── backend/                    # Python FastAPI (the bridge)
│   ├── agent/                  # ACP SDK integration (spawn + protocol)
│   ├── bridge/                 # ACP → AG-UI event translation
│   ├── agui/                   # AG-UI event types + SSE encoding
│   ├── sessions/               # Session lifecycle, store, routes
│   ├── policy/                 # Tool approval engine
│   ├── api/                    # Side-channel REST (files, git)
│   └── agui_endpoint.py        # POST /ag-ui (AG-UI standard endpoint)
├── example-frontends/
│   ├── custom-workspace-ui-demo/  # Full workspace UI (React + Vite + Tailwind)
│   ├── copilotkit-demo/           # CopilotKit in 20 lines
│   ├── httpagent-demo/            # Raw @ag-ui/client AG-UI HttpAgent
│   └── agents.md                  # Agent configuration guide
├── docs/
│   ├── architecture.md         # Detailed system design
│   ├── integration-contract.md # REST + SSE API spec
│   ├── protocol-translation.md # Full ACP ↔ AG-UI mapping
│   └── why-agui.md            # AG-UI ecosystem benefits
├── bridge.config.json          # Your agent configuration
└── package.json                # Workspace orchestrator
```

</details>

## Tested With Real Agents

| Agent | Version | Status | Notes |
|-------|---------|--------|-------|
| **Kiro CLI** | 2.3.0 | ✅ Working | 13 modes, extension notifications, full streaming |
| **Claude Agent** (claude-agent-acp) | 0.36.1 | ✅ Working | 5 modes, prompt queueing, embedded context |
| **Codex CLI** (codex-acp) | 0.14.0 | 🟡 Supported | Via Zed adapter, tool calls + edit review |
| **OpenCode** | 1.15.6 | 🟡 Supported | Native ACP, 2 agents (build/plan), MCP servers |

Kiro and Claude Agent tested end-to-end with zero code changes between them. Codex and OpenCode are ACP-compatible and supported by this bridge; community testing welcome. Just swap `agentCommand`. See [`docs/demo-walkthrough.md`](docs/demo-walkthrough.md) for full test results and [`example-frontends/agents.md`](example-frontends/agents.md) for setup guides.

<details>
<summary>Full ACP Ecosystem (33+ agents)</summary>

| Agent | ACP Type | Notes |
|-------|----------|-------|
| Kiro CLI | Native | Full-featured, 13+ modes, custom agents |
| Claude Code | Via adapter ([claude-agent-acp](https://github.com/zed-industries/claude-code-acp)) | Permission modes, tool calls, MCP |
| Codex CLI | Via adapter ([codex-acp](https://github.com/zed-industries/codex-acp)) | Edit review, slash commands, MCP |
| OpenCode | Native | 2 built-in agents, custom tools, MCP |
| Gemini CLI | Native | Google's coding agent |
| GitHub Copilot | Native (public preview) | Copilot in terminal |
| Cursor | Native | IDE agent over ACP |
| Goose | Native | Block's open-source agent |
| Cline | Native | VS Code agent with ACP |

Also supported: Augment Code, AutoDev, Blackbox AI, Docker cagent, fast-agent, Factory Droid, Hermes Agent, Junie (JetBrains), Kimi CLI, Mistral Vibe, OpenHands, Poolside, Qwen Code, and more.

Full list: [agentclientprotocol.com/get-started/agents](https://agentclientprotocol.com/get-started/agents)

</details>

## The Talk

This repository accompanies a live demo presented at [Seattle AI Tinkerers](https://seattle.aitinkerers.org/), May 2026.

## Contributing

Contributions welcome! Areas of interest:

- Additional agent configuration examples
- Frontend components for new AG-UI event types
- Policy engine enhancements (configurable approval rules)
- Session resume/persistence improvements
- More AG-UI event types (STATE_DELTA, activities, reasoning)

## License

MIT

---

This project is independent and is not affiliated with, endorsed by, or sponsored by Amazon/Kiro, Anthropic/Claude, OpenAI/Codex, Google/Gemini, GitHub/Copilot, Cursor/Anysphere, OpenCode, Cline, or their respective owners. All product names, logos, and trademarks are property of their respective owners.

acp-to-agui

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README