Agent Client Protocol (ACP)

The Agent Client Protocol (ACP) is a standardized JSON-RPC protocol for AI agent communication. Goose implements ACP for rich, bidirectional integration with IDEs, desktop applications, and other tools.

Overview

ACP provides:

Bidirectional communication: Agents can request permissions and receive cancellations
Rich tool call handling: Detailed status updates, locations, and content
Session management: Create, load, and resume sessions with full history
MCP server integration: Dynamically add MCP servers to sessions
Streaming responses: Real-time updates as the agent works

ACP vs REST API

Feature	ACP (stdio/WebSocket)	REST API (goose-server)
Use case	IDE plugins, desktop apps, embedded agents	Web apps, simple clients
Transport	stdio, WebSocket, HTTP	HTTP only
Permissions	Interactive prompts	Not supported
Tool locations	File paths with line numbers	Not provided
Streaming	JSON-RPC notifications	Server-Sent Events
Session resume	Full conversation history	Session ID only

Use ACP for desktop applications and IDE integrations where you need rich tool feedback and permission prompts. Use the REST API for simpler web-based integrations.

Quick Start

Running ACP Server

# Start Goose as an ACP server on stdio
goose acp --with-builtin developer,memory

# Or programmatically
cargo run -p goose-cli -- acp --with-builtin developer

The server communicates via stdin/stdout using JSON-RPC.

Python Client Example

See test_acp_client.py in the source repository for a complete example:

import subprocess
import json

class AcpClient:
    def __init__(self):
        self.process = subprocess.Popen(
            ['goose', 'acp', '--with-builtin', 'developer'],
            stdin=subprocess.PIPE,
            stdout=subprocess.PIPE,
            text=True
        )
        self.request_id = 0
    
    def send_request(self, method, params=None):
        self.request_id += 1
        request = {
            "jsonrpc": "2.0",
            "method": method,
            "id": self.request_id
        }
        if params:
            request["params"] = params
        
        self.process.stdin.write(json.dumps(request) + "\n")
        self.process.stdin.flush()
        
        response = json.loads(self.process.stdout.readline())
        return response

# Initialize connection
client = AcpClient()
client.send_request("initialize", {
    "protocolVersion": "v1",
    "clientInfo": {"name": "my-client", "version": "1.0.0"}
})

# Create session
session = client.send_request("session/new", {
    "cwd": "/path/to/project"
})

# Send prompt
client.send_request("session/prompt", {
    "sessionId": session["result"]["sessionId"],
    "prompt": [{"type": "text", "text": "List files"}]
})

Protocol Methods

Connection Lifecycle

initialize

Establish connection and exchange capabilities. Request:

{
  "jsonrpc": "2.0",
  "method": "initialize",
  "id": 1,
  "params": {
    "protocolVersion": "v1",
    "clientCapabilities": {},
    "clientInfo": {
      "name": "vscode-goose",
      "version": "1.0.0"
    }
  }
}

Response:

{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "protocolVersion": "v1",
    "agentCapabilities": {
      "loadSession": true,
      "promptCapabilities": {
        "image": true,
        "audio": false,
        "embeddedContext": true
      },
      "mcpCapabilities": {
        "http": true
      }
    },
    "authMethods": [
      {
        "id": "goose-provider",
        "name": "Configure Provider",
        "description": "Run `goose configure` to set up your AI provider"
      }
    ]
  }
}

Session Management

session/new

Create a new session with optional MCP servers. Request:

{
  "jsonrpc": "2.0",
  "method": "session/new",
  "id": 2,
  "params": {
    "cwd": "/home/user/project",
    "mcpServers": [
      {
        "type": "stdio",
        "name": "github",
        "command": "uvx",
        "args": ["github-mcp-server"],
        "env": [{"name": "GITHUB_TOKEN", "value": "ghp_..."}]
      }
    ]
  }
}

Response:

{
  "jsonrpc": "2.0",
  "id": 2,
  "result": {
    "sessionId": "550e8400-e29b-41d4-a716-446655440000",
    "models": {
      "current": "claude-sonnet-4-20250514",
      "available": [
        {"id": "claude-sonnet-4-20250514", "name": "Claude Sonnet 4"},
        {"id": "gpt-4o", "name": "GPT-4o"}
      ]
    }
  }
}

session/load

Load an existing session by ID. Request:

{
  "jsonrpc": "2.0",
  "method": "session/load",
  "id": 3,
  "params": {
    "sessionId": "550e8400-e29b-41d4-a716-446655440000",
    "cwd": "/home/user/project"
  }
}

Response: Returns full conversation history as a series of session/notification notifications, followed by the response:

{
  "jsonrpc": "2.0",
  "id": 3,
  "result": {
    "models": { /* same as session/new */ }
  }
}

session/prompt

Send a prompt and receive streaming responses. Request:

{
  "jsonrpc": "2.0",
  "method": "session/prompt",
  "id": 4,
  "params": {
    "sessionId": "550e8400-e29b-41d4-a716-446655440000",
    "prompt": [
      {"type": "text", "text": "What files are in this directory?"},
      {"type": "image", "data": "base64...", "mimeType": "image/png"}
    ]
  }
}

Response:

{
  "jsonrpc": "2.0",
  "id": 4,
  "result": {
    "stopReason": "endTurn"
  }
}

The response is sent only after all streaming notifications complete. Use notifications to receive real-time updates.

session/cancel

Cancel an in-progress prompt. Notification:

{
  "jsonrpc": "2.0",
  "method": "session/cancel",
  "params": {
    "sessionId": "550e8400-e29b-41d4-a716-446655440000"
  }
}

Notifications (Agent → Client)

Notifications stream updates during session/prompt.

session/notification - Agent Message

{
  "jsonrpc": "2.0",
  "method": "session/notification",
  "params": {
    "sessionId": "550e8400-e29b-41d4-a716-446655440000",
    "update": {
      "sessionUpdate": "agentMessageChunk",
      "chunk": {
        "content": {"type": "text", "text": "Let me check the files"}
      }
    }
  }
}

session/notification - Tool Call

{
  "jsonrpc": "2.0",
  "method": "session/notification",
  "params": {
    "sessionId": "550e8400-e29b-41d4-a716-446655440000",
    "update": {
      "sessionUpdate": "toolCall",
      "toolCall": {
        "id": "call_abc123",
        "title": "Developer: List Files",
        "status": "pending"
      }
    }
  }
}

session/notification - Tool Call Update

Includes file locations for editor integration:

{
  "jsonrpc": "2.0",
  "method": "session/notification",
  "params": {
    "sessionId": "550e8400-e29b-41d4-a716-446655440000",
    "update": {
      "sessionUpdate": "toolCallUpdate",
      "toolCallUpdate": {
        "id": "call_abc123",
        "fields": {
          "status": "completed",
          "content": [
            {"type": "text", "text": "Found 5 files"}
          ],
          "locations": [
            {"path": "/home/user/project/main.rs", "line": 1}
          ]
        }
      }
    }
  }
}

The locations field enables IDE integrations to highlight or navigate to files modified by tools.

requestPermission - Interactive Approval

For sensitive operations, the agent requests permission:

{
  "jsonrpc": "2.0",
  "method": "requestPermission",
  "id": 100,
  "params": {
    "sessionId": "550e8400-e29b-41d4-a716-446655440000",
    "toolCallUpdate": {
      "id": "call_xyz789",
      "fields": {
        "title": "Execute Command",
        "status": "pending",
        "rawInput": {"command": "rm -rf /"},
        "content": [{"type": "text", "text": "Delete all files?"}]
      }
    },
    "options": [
      {"id": "allow_always", "name": "Allow Always", "kind": "allowAlways"},
      {"id": "allow_once", "name": "Allow Once", "kind": "allowOnce"},
      {"id": "reject_once", "name": "Reject Once", "kind": "rejectOnce"},
      {"id": "reject_always", "name": "Reject Always", "kind": "rejectAlways"}
    ]
  }
}

Client Response:

{
  "jsonrpc": "2.0",
  "id": 100,
  "result": {
    "outcome": {
      "type": "selected",
      "optionId": "allow_once"
    }
  }
}

Custom Methods

Goose extends ACP with custom methods for additional functionality.

_extensions/add

Add an MCP extension to a running session:

{
  "jsonrpc": "2.0",
  "method": "_extensions/add",
  "id": 5,
  "params": {
    "sessionId": "550e8400-e29b-41d4-a716-446655440000",
    "config": {
      "type": "stdio",
      "name": "slack",
      "cmd": "uvx",
      "args": ["mcp-slack"]
    }
  }
}

_session/list

List all sessions:

{
  "jsonrpc": "2.0",
  "method": "_session/list",
  "id": 6
}

Response:

{
  "jsonrpc": "2.0",
  "id": 6,
  "result": {
    "sessions": [
      {
        "sessionId": "550e8400-e29b-41d4-a716-446655440000",
        "title": "My Project",
        "cwd": "/home/user/project",
        "updatedAt": "2026-03-04T12:00:00Z"
      }
    ]
  }
}

Other Custom Methods

_session/get: Get full session details
_session/delete: Delete a session
_session/export: Export session data
_session/import: Import session data
_extensions/remove: Remove an extension
_tools: List available tools
_resource/read: Read an MCP resource
_working_dir/update: Change session working directory
_config/extensions: Get extension configuration

Transport Options

stdio (Default)

Communicate via standard input/output:

goose acp --with-builtin developer

Ideal for:

CLI tools
Editor plugins (VS Code, Neovim)
Process-based integrations

WebSocket (Future)

WebSocket transport for browser-based clients:

goose acp --transport websocket --port 8080

HTTP (Future)

HTTP transport for stateless clients:

goose acp --transport http --port 8080

Implementation Details

Source Code

ACP server: crates/goose-acp/src/server.rs
CLI integration: crates/goose-cli/src/cli.rs (Command::Acp)
Protocol library: sacp crate (vendored)
Custom methods: crates/goose-acp/src/custom_requests.rs
Test client: test_acp_client.py

Tool Location Extraction

Goose automatically extracts file locations from tool responses for developer tools:

// crates/goose-acp/src/server.rs
fn extract_tool_locations(
    tool_request: &ToolRequest,
    tool_response: &ToolResponse,
) -> Vec<ToolCallLocation> {
    // Parse tool output for file paths and line numbers
    // Returns locations for IDE navigation
}

Supported tools:

write: File creation (line 1)
edit: File modification (parsed from output)
view: File viewing (parsed from output)

Session State Management

ACP sessions are isolated per connection:

struct GooseAcpSession {
    agent: Arc<Agent>,
    messages: Conversation,
    tool_requests: HashMap<String, ToolRequest>,
    cancel_token: Option<CancellationToken>,
}

Sessions accumulate until the transport closes.

Best Practices

Error Handling

Handle JSON-RPC errors gracefully:

response = client.send_request("session/prompt", params)
if "error" in response:
    print(f"Error {response['error']['code']}: {response['error']['message']}")
    # Fallback logic

Notification Buffering

Buffer notifications during streaming:

def collect_response(client, request_id):
    notifications = []
    while True:
        line = client.process.stdout.readline()
        msg = json.loads(line)
        
        if "method" in msg:  # Notification
            notifications.append(msg)
        elif msg.get("id") == request_id:  # Response
            return msg, notifications

Cancellation

Implement cancellation for long-running operations:

import threading

def prompt_with_timeout(client, session_id, prompt, timeout=30):
    def cancel():
        time.sleep(timeout)
        client.send_notification("session/cancel", {"sessionId": session_id})
    
    threading.Thread(target=cancel, daemon=True).start()
    return client.send_request("session/prompt", {...})

Resources

ACP Specification: Anthropic Cookbook
Goose ACP Implementation: crates/goose-acp/
Test Client: test_acp_client.py
sacp Library: vendor/sacp/ (Rust ACP implementation)

​Overview

​ACP vs REST API

​Quick Start

​Running ACP Server

​Python Client Example

​Protocol Methods

​Connection Lifecycle

​initialize

​Session Management

​session/new

​session/load

​session/prompt

​session/cancel

​Notifications (Agent → Client)

​session/notification - Agent Message

​session/notification - Tool Call

​session/notification - Tool Call Update

​requestPermission - Interactive Approval

​Custom Methods

​_extensions/add

​_session/list

​Other Custom Methods

​Transport Options

​stdio (Default)

​WebSocket (Future)

​HTTP (Future)

​Implementation Details

​Source Code

​Tool Location Extraction

​Session State Management

​Best Practices

​Error Handling

​Notification Buffering

​Cancellation

​Resources

Overview

ACP vs REST API

Quick Start

Running ACP Server

Python Client Example

Protocol Methods

Connection Lifecycle

initialize

Session Management

session/new

session/load

session/prompt

session/cancel

Notifications (Agent → Client)

session/notification - Agent Message

session/notification - Tool Call

session/notification - Tool Call Update

requestPermission - Interactive Approval

Custom Methods

_extensions/add

_session/list

Other Custom Methods

Transport Options

stdio (Default)

WebSocket (Future)

HTTP (Future)

Implementation Details

Source Code

Tool Location Extraction

Session State Management

Best Practices

Error Handling

Notification Buffering

Cancellation

Resources