Set up Open Agent Search

Use the official npm or Python package open-agent-search. No paid search API key or account is required. Search queries and fetched URLs go to upstream providers and websites; running locally does not make those requests private.

Check uv --version. If missing, follow https://docs.astral.sh/uv/getting-started/installation/ for the user's OS. Use an isolated environment with managed Python 3.12:

uvx --python 3.12 --from open-agent-search==1.0.4 oas-mcp

Alternatively, with Node.js 22+ and uv installed, use npx --yes open-agent-search@1.0.4 mcp or install open-agent-search@1.0.4 with npm and run oas-mcp. The npm launcher still requires uv and uses the same managed Python package. Its --version and --help work without starting Python. OPEN_AGENT_SEARCH_UVX can select an absolute uvx path.

The first launch downloads Python if needed and dependencies. The Python entry point is a stdio server, not a CLI that prints search results or a --help check. Let the MCP client manage its lifetime.

For a plugin installation, the bundled MCP manifest already supplies this command. Otherwise add a server named open-agent-search with command uvx and args ["--python", "3.12", "--from", "open-agent-search==1.0.4", "oas-mcp"] to the current client's supported MCP configuration. Merge that entry and preserve all other servers and settings. If the client cannot find uvx, resolve its absolute executable path. Do not wrap stdio with commands that print to stdout.

Install the setup and usage skills for the current agent:

npx --yes skills add Jayanth-MKV/open-agent-search --skill open-agent-search-setup open-agent-search --agent codex --copy --yes

Replace codex with the current agent's supported identifier and default to project scope. A plugin supplies both skills, so skip duplicate installation. If Node.js is unavailable, copy only these two repository skill directories into the client's supported skills directory, checking existing destinations before replacing them.

Verify through the MCP client: initialize, list tools, and confirm search_web, search_images, search_videos, search_news, search_books, search_everything, fetch_content, and fetch_multiple_contents. Run one short web search for a non-sensitive query. Distinguish connection success from an upstream search failure; include the actual error and avoid repeated retries. If tools become available only after a restart, state that verification is still pending.

HTTP hosting is optional. uv tool install --python 3.12 open-agent-search==1.0.4 exposes open-agent-search; the npm command is open-agent-search serve. Both listen on all interfaces at port 8000 and mount HTTP MCP at /ai/mcp. Do not start HTTP hosting or deploy a service for a task that only asks for local MCP tools.