nova‑act‑mcp‑server is a zero‑install Model Context Protocol (MCP) server that exposes Amazon Nova Act browser‑automation tools.

• On-Demand Screenshots: New inspect_browser tool to explicitly request screenshots only when needed
• Reduced Token Usage: Browser actions no longer automatically include screenshots, saving context space
• More Efficient Workflows: Agents can now control when to get visual feedback
• Better Performance: Smaller response payloads improve overall agent experience

# Start a browser session
start_result = await control_browser(action="start", url="https://example.com")
session_id = start_result["session_id"]

# Execute an action without getting a screenshot
execute_result = await control_browser(
    action="execute",
    session_id=session_id,
    instruction="Click on the 'More information...' link"
)

# Now explicitly request a screenshot to see the result
inspect_result = await inspect_browser(session_id=session_id)

# Example output from inspect_browser:
{
  "session_id": "f8a53291-b3a7-4e1e-8c9d-9a12b3c45d67",
  "current_url": "https://www.iana.org/domains/reserved",
  "page_title": "IANA — IANA-managed Reserved Domains",
  "content": [
    {
      "type": "image_base64",
      "data": "data:image/jpeg;base64,/9j/4AAQSkZJRgABAQEASABIAAD/2wBDAAMCA...",
      "caption": "Current viewport"
    },
    {
      "type": "text",
      "text": "Current URL: https://www.iana.org/domains/reserved\nPage Title: IANA — IANA-managed Reserved Domains"
    }
  ],
  "agent_thinking": [],
  "success": true
}

• Improved Screenshot Reliability: More dependable screenshot delivery in responses
• Enhanced Log Path Discovery: Smart, efficient path tracking for logs and screenshots
• Better Agent Communication: Clear messaging when screenshots can't be embedded
• Improved Performance: Eliminated inefficient directory scanning for faster responses

• Enhanced Inline Screenshots: Screenshots now appear directly in the response content array
• Improved compatibility with vision-capable models like Claude
• Screenshots include descriptive captions based on the executed instruction
• Each screenshot is delivered as { type: "image_base64", data: "..." } in the content array

• Automatic Inline Screenshots: Every browser action now includes an optimized screenshot
• Improved screenshot quality and reliability for AI agents
• Added environment variables to customize screenshot quality and size limits
• Comprehensive test coverage ensuring screenshots work in all scenarios

Every successful execute response now contains inline_screenshot, a base64-encoded JPEG of the current viewport:

• Quality ≈ 45, hard-capped at 250 KB (configurable via NOVA_MCP_MAX_INLINE_IMG env variable)
• If the raw JPEG is larger than the cap, the field is null
• No extra API calls needed - screenshots are included automatically
• For full-resolution images and HAR/HTML logs, use the compress_logs tool

• Added compatibility with NovaAct SDK 0.9+ by normalizing log directory handling
• Improved test organization with clear markers for unit, mock, smoke and e2e tests
• Moved mock HTML creation logic from production code to test helpers
• Fixed several syntax errors and incomplete code blocks
• Added SCREENSHOT_QUALITY constant for consistent compression settings

Add it to your MCP client configuration:

{
  "mcpServers": {
    "nova-act-mcp-server": {
      "command": "uvx",
      "args": ["nova-act-mcp-server@latest"],
      "env": { "NOVA_ACT_API_KEY": "<your_api_key>" }
    }
  }
}

That's all you need to start controlling browsers from any MCP‑compatible client such as Claude Desktop or VS Code.

git clone https://github.com/madtank/nova-act-mcp.git
cd nova-act-mcp
uv sync
uv run nova_mcp.py

MIT