openai-agents-contextfetch

ContextFetch web search and page reading as function tools for the OpenAI Agents SDK.

pip install openai-agents-contextfetch
export CONTEXTFETCH_API_KEY="your-api-key"  # optional: without a key, requests use the keyless tier

from agents import Agent, Runner

from openai_agents_contextfetch import contextfetch_tools

agent = Agent(name="Researcher", instructions="Answer with sources.", tools=contextfetch_tools())
result = Runner.run_sync(agent, "What changed in the latest Python release?")
print(result.final_output)

Tools

• contextfetch_search(query, include_domains?, exclude_domains?, start_published_date?): web search. Returns a list of {title, url, published_date, content}, where content is the page as Markdown: by default only the parts relevant to the query.
• contextfetch_read(url, query?): one page as Markdown. With query, only the parts relevant to it.

API errors (bad URL, no credits, rate limit) go back to the model as the tool output, with the API's message.

Options

contextfetch_tools(
    api_key=None,              # defaults to CONTEXTFETCH_API_KEY
    num_results=5,             # results per search, 1 to 25
    content="pruned_context",  # or "text" for whole pages
)

Without this package

The same tools are available over MCP: npx -y @contextfetch/mcp with the SDK's MCPServerStdio.

Tests

uv run --group test pytest tests                                   # unit tests, no network
CONTEXTFETCH_API_KEY=... uv run --group test pytest tests/test_live.py  # live, 2 credits

Tested with openai-agents 0.20.0, 0.22.0 and 0.24.0. MIT licensed.