Supported Agents#

Chrome ACP works with any ACP-compatible agent. Here are setup instructions for popular agents.

#Claude Code

Anthropic's agentic coding tool.

# Install
npm install -g @chrome-acp/proxy-server @anthropic-ai/claude-code @zed-industries/claude-code-acp

# Run
acp-proxy --no-auth claude-code-acp

Requirements: ANTHROPIC_API_KEY environment variable

Features: Image support, Model selection, Extended thinking

#Codex CLI

OpenAI's coding agent.

# Install
npm install -g @chrome-acp/proxy-server @openai/codex @zed-industries/codex-acp

# Run
acp-proxy --no-auth codex-acp

Requirements: OPENAI_API_KEY environment variable

#OpenCode

Open-source terminal AI assistant with multi-provider support.

# Install
npm install -g @chrome-acp/proxy-server
curl -fsSL https://opencode.ai/install | bash

# Run
acp-proxy --no-auth opencode acp

#Gemini CLI

Google's AI agent with a generous free tier.

# Install
npm install -g @chrome-acp/proxy-server @google/gemini-cli

# Run
acp-proxy --no-auth gemini -- --experimental-acp

Requirements: GOOGLE_API_KEY or Google Cloud credentials

#Qwen Code

Free coding agent using Qwen models (no API key required).

# Install
npm install -g @chrome-acp/proxy-server @qwen-code/qwen-code@latest

# Run
acp-proxy --no-auth qwen -- --acp

#Augment Code

AI coding agent by Augment.

# Install
npm install -g @chrome-acp/proxy-server @augmentcode/auggie

# Run
acp-proxy --no-auth auggie -- --acp

#Agent Capabilities

Agents can declare their capabilities when creating a session. Chrome ACP supports:

#Image Support

Agents can accept images in prompts:

promptCapabilities: {
  image: true  // Enables image attachment button
}

When enabled, users can:

• Attach images via button or drag-and-drop
• Paste screenshots from clipboard
• Images are compressed to <2MB and sent as base64

#Model Selection

Agents can expose multiple models:

modelState: {
  availableModels: [
    { id: "claude-3-opus", displayName: "Claude 3 Opus" },
    { id: "claude-3-sonnet", displayName: "Claude 3 Sonnet" }
  ],
  currentModelId: "claude-3-sonnet"
}

When available, a model selector appears in the chat input footer.

#Extended Thinking

Some agents support "thinking" or "reasoning" modes where they show their thought process:

{ sessionUpdate: "agent_thought_chunk", content: { type: "text", text: "..." } }

Thought chunks are displayed in a collapsible "Thinking" section.

#Custom Agents

Any ACP-compatible agent works with Chrome ACP:

acp-proxy --no-auth /path/to/your/agent

#Requirements

Your agent must:

1. Accept communication via stdin/stdout
2. Implement the ACP protocol (NDJSON messages)
3. Handle session lifecycle (new_session, prompt, etc.)

#Minimal Protocol

← {"request":"new_session","requestId":"1"}
→ {"requestId":"1","sessionId":"abc123","promptCapabilities":{"image":false}}
← {"request":"prompt","requestId":"2","sessionId":"abc123","content":[{"type":"text","text":"Hello"}]}
→ {"requestId":"2","sessionUpdate":"agent_message_chunk","content":{"type":"text","text":"Hi there!"}}
→ {"requestId":"2","promptComplete":{"stopReason":"end_turn"}}

#Adding Browser Tools

To use browser tools, your agent must support MCP and connect to:

http://localhost:9315/mcp

See Architecture for MCP protocol details.

#Troubleshooting

#Agent doesn't start

# Test the agent directly
claude-code-acp

# Check if it accepts stdin
echo '{"request":"new_session","requestId":"1"}' | claude-code-acp

#Missing API key

Error: ANTHROPIC_API_KEY not set

Set the required environment variable:

export ANTHROPIC_API_KEY="sk-ant-..."

#Agent crashes immediately

Enable debug mode to see detailed logs:

acp-proxy --debug --no-auth claude-code-acp

Check logs in .acp-proxy/ directory.