workspace-mcp

# OpenBB Workspace MCP OpenBB Workspace Companion App that exposes the connected browser session as an MCP server. What it does: - exposes a stateless streamable HTTP MCP endpoint at `/mcp` - forwards MCP tool calls to the connected Workspace tab over websocket - returns fresh workspace snapshots and command results from the browser Important security considerations: - use only if you treat local MCP clients as trusted; they can read and mutate the connected Workspace - do not expose the sidecar on `0.0.0.0`, a LAN, a tunnel, or a public reverse proxy - `http` is fine for local deployments (`localhost`/`127.0.0.1`) ## How to use ### 1. Install & Run Install: ```bash uv tool install --python 3.13 git+https://github.com/OpenBB-finance/workspace-mcp.git # If installing from a PR uv tool install --force --editable --python 3.13 . ``` Run: ```bash workspace-mcp --host 127.0.0.1 --port 8787 ``` ### 2. Connect from Workspace - open OpenBB Workspace - click the hamburger icon in the top left of the Workspace UI - find `Workspace MCP Companion` - set the companion base URL to your local sidecar URL, for example `http://127.0.0.1:8787` - connect the companion after the sidecar is running ### 3. Connect your AI agent Use the `http://127.0.0.1:8787/mcp` as the MCP url that you pass into the `mcp add` command of your agent. Example `.mcp.json` for claude code: ```json { "mcpServers": { "workspace_mcp": { "type": "http", "url": "http://127.0.0.1:8787/mcp" } } } ``` ## Features MCP tools: - `get_workspace_snapshot` - `list_available_widgets` - `get_widget_schema` - `get_widget_data` - `get_params_options` - `manage_dashboard` - `navigate_workspace` - `update_widget_layout` - `read_widget` - `create_widget` - `update_widget` - `delete_widget` - `manage_navigation_bar` - `add_generative_widget` - `manage_backends` - `manage_apps` - `assign_tasks_to_agents` - `get_skill_content` Scope: - single user - one connected Workspace browser session - local-only sidecar - flat tool list ## Development Reload on code changes: ```bash python -m workspace_mcp --host 127.0.0.1 --port 8787 --reload ```