Browser-Use MCP Server

# Modernize MCP Browser-Use ## 1. Primary Request and Intent Igor requested modernization of mcp-browser-use MCP server: - Update browser-use dependency from 0.1.41 to 0.10.1 - Remove LangChain ecosystem dependencies - Simplify codebase by removing custom wrappers - Remove deep research feature - Use browser-use's native LLM providers - Reduce codebase from ~3200 lines to ~300 lines **Current handoff purpose**: Create a validation plan before committing the changes. ## 2. Key Technical Concepts - **browser-use 0.10.1**: Complete API rewrite with native LLM providers - **MCP (Model Context Protocol)**: Server exposes browser-use as tools - **FastMCP**: MCP server framework used - **BrowserProfile**: Replaces old BrowserConfig - **Native Chat models**: ChatOpenAI, ChatAnthropic, ChatGoogle, ChatOllama (no LangChain) - **Pydantic Settings**: Configuration management ## 3. Files and Code Sections ### `/Users/igortarasenko/Projects/mcp-browser-use/pyproject.toml` - **Why important**: Defines dependencies - core of the simplification - **Changes made**: Reduced from 9+ dependencies to 4, version bumped to 0.2.0 - **Code snippet**: ```toml dependencies = [ "browser-use>=0.10.1", "mcp>=1.10.1", "pydantic-settings>=2.0.0", "typer>=0.12.0", ] ``` ### `/Users/igortarasenko/Projects/mcp-browser-use/src/mcp_server_browser_use/server.py` - **Why important**: Main MCP server - exposes `run_browser_agent` tool - **Changes made**: Complete rewrite using native browser-use API (86 lines) - **Code snippet**: ```python from browser_use import Agent, BrowserProfile from mcp.server.fastmcp import Context, FastMCP from .providers import get_llm @server.tool() async def run_browser_agent(ctx: Context, task: str, max_steps: Optional[int] = None) -> str: llm = get_llm(provider=settings.llm.provider, model=settings.llm.model_name, api_key=settings.llm.get_api_key()) profile = BrowserProfile(headless=settings.browser.headless) agent = Agent(task=task, llm=llm, browser_profile=profile, max_steps=steps) result = await agent.run() return result.final_result() or "Task completed without explicit result." ``` ### `/Users/igortarasenko/Projects/mcp-browser-use/src/mcp_server_browser_use/providers.py` - **Why important**: LLM factory using browser-use native providers - **Changes made**: New file (49 lines), replaces 327-line LangChain-based llm_provider.py - **Code snippet**: ```python from browser_use import ChatAnthropic, ChatGoogle, ChatOllama, ChatOpenAI def get_llm(provider: str, model: str, api_key: str | None = None) -> "BaseChatModel": if provider == "openai": return ChatOpenAI(model=model, api_key=api_key) elif provider == "anthropic": return ChatAnthropic(model=model, api_key=api_key) # ... etc ``` ### `/Users/igortarasenko/Projects/mcp-browser-use/src/mcp_server_browser_use/config.py` - **Why important**: Simplified configuration (61 lines vs 226 lines) - **Changes made**: Reduced to 4 LLM providers, removed deep research settings, removed CDP settings - **Code snippet**: ```python ProviderType = Literal["openai", "anthropic", "google", "ollama"] class LLMSettings(BaseSettings): provider: ProviderType = Field(default="anthropic") model_name: str = Field(default="claude-sonnet-4-20250514") api_key: Optional[SecretStr] = Field(default=None) ``` ### `/Users/igortarasenko/Projects/mcp-browser-use/src/mcp_server_browser_use/cli.py` - **Why important**: CLI interface for testing - **Changes made**: Simplified from 320 lines to 63 lines ### Deleted Files (via `trash` command) - `_internal/agent/deep_research/deep_research_agent.py` (1015 lines) - `_internal/agent/browser_use/browser_use_agent.py` (177 lines) - `_internal/browser/custom_browser.py` (130 lines) - `_internal/browser/custom_context.py` (116 lines) - `_internal/controller/custom_controller.py` (178 lines) - `_internal/utils/llm_provider.py` (327 lines) - `_internal/utils/mcp_client.py` (267 lines) ## 4. Problem Solving - **Solved**: Identified browser-use API changes (0.1.x → 0.10.x) - **Solved**: Removed LangChain dependency entirely - **Solved**: All linting/type checking passes - **Ongoing**: Need validation before commit ## 5. Pending Tasks - Validate implementation works end-to-end before committing - Create commit with proper message - Optionally update README for breaking changes ## 6. Current Work Implementation is COMPLETE but NOT COMMITTED. All checks pass: - `ruff format .` ✓ - `ruff check .` ✓ - `pyright` ✓ (0 errors) - Import test ✓ - CLI config command ✓ - Line count: 298 lines total Git status shows all changes staged but uncommitted. ## 7. Validation Plan (Next Steps) Before committing, validate: ### 1. Test missing API key error handling ```bash uv run python -c " from mcp_server_browser_use.providers import get_llm from mcp_server_browser_use.exceptions import LLMProviderError try: get_llm('openai', 'gpt-4', None) except LLMProviderError as e: print('✓ Error handling works:', e) " ``` ### 2. Test MCP server starts ```bash MCP_LLM_API_KEY=test-key timeout 5 uv run mcp-server-browser-use || echo "Server started (timeout expected)" ``` ### 3. Test CLI with real API key (if available) ```bash MCP_LLM_PROVIDER=anthropic MCP_LLM_API_KEY=$ANTHROPIC_API_KEY MCP_BROWSER_HEADLESS=false uv run mcp-browser-cli run "Navigate to example.com" ``` ### 4. Verify browser-use Agent API compatibility ```bash uv run python -c " from browser_use import Agent, BrowserProfile from mcp_server_browser_use import get_llm, settings # Test that we can construct the objects profile = BrowserProfile(headless=True) print('✓ BrowserProfile works') # Test LLM construction (will fail without API key, but tests imports) try: llm = get_llm('anthropic', 'claude-sonnet-4-20250514', 'test-key') print('✓ LLM construction works') except Exception as e: print('LLM construction:', e) " ``` ### 5. After validation, commit with: ```bash git add -A git commit -m 'feat: modernize to browser-use 0.10.1, remove LangChain - Update browser-use 0.1.41 → 0.10.1 - Remove LangChain ecosystem (langchain-*, langgraph) - Use native browser-use LLM providers - Remove deep research feature - Simplify from ~3200 lines to 298 lines - Support 4 providers: OpenAI, Anthropic, Google, Ollama BREAKING CHANGES: - Deep research tool removed - CDP connection removed - keep_open mode removed - LLM providers reduced to 4 core providers - Environment variable MCP_LLM_*_API_KEY → MCP_LLM_API_KEY ' ```