feizhuniu-infja/lift/firecrawl

Firecrawl (web scraping & search)

Reach the live internet through Firecrawl's official hosted MCP server, straight from the IPython kernel. This is the only real "go online" capability in the LIFT eval container (there is no browser and no bundled websearch), so prefer it over guessing when a task depends on current external facts.

Setup

This integration uses a static bearer token, not OAuth. It is connected whenever the FIRECRAWL_API_KEY environment variable is set (the LIFT image bakes it from the build-time secret). If a call raises NotEnabled, the key is missing — tell the user to set FIRECRAWL_API_KEY in the environment and rebuild the image. Do not point them at /mcp login; that flow has no provider for a bearer-only server and reports "Unknown MCP integration".

Usage

The tool set is defined by the server, not by this skill, so discover before you call — don't assume tool names or argument names:

import firecrawl
# 1. Discover available tools
for tool in await firecrawl.list_tools():
    print(tool["name"], "-", tool["description"])
# 2. Inspect a specific tool's arguments (rendered from its JSON Schema)
help(firecrawl.firecrawl_scrape)
# 3. Call it; keyword args must match the tool's input schema
page = await firecrawl.firecrawl_scrape(url="https://example.com")
print(page)
hits = await firecrawl.firecrawl_search(query="prime intellect prime-agent release")
print(hits)

Notes:

• Every tool is an async method — always await.
• Results are already-parsed Python (a dict for structured output, otherwise a

string). No need to json.loads them.

• For tools whose names aren't valid Python identifiers, use the escape hatch:

await firecrawl.call_tool("tool-name", {"arg": "value"}).

• Run list_tools() before relying on help() or assuming a tool exists — it

populates the schemas help() shows, and the server is the source of truth for tool names and arguments.

• Scrapes can be large; when you only need the main text, pass

onlyMainContent=True (check the tool schema) and summarize rather than dumping the whole page back into context.