Summary

One question in, a research pass out — not a list of links.

Because the agent does the browsing itself, your client spends one tool call where it would otherwise spend twenty, and the reading of PDFs and transcripts happens on the far side. The cost is three separate accounts before anything runs, and inference billed to your own OpenAI key on every question.

What it is

The `deep_research` agent from HuggingFace's open_deep_research example, wrapped as an MCP server. Rather than a single search call, the agent browses, opens documents and reports back, so the client asks one research question and receives worked findings.

What you get

Web search and information gathering through a text-based browser
PDF and document analysis, with format conversion to Markdown for reading
Image analysis and description
YouTube transcript retrieval
Archive site search

Requirements

Python 3.11 or higher and the `uv` package manager, or Docker. Three keys are required in the environment: `OPENAI_API_KEY`, `HF_TOKEN` and `SERPER_API_KEY` (free signup at Serper.dev). Run locally with `uv run deep_research.py`, or build the image and register a `docker run` command with the three keys passed in; the container also exposes port 8080 if you would rather connect over SSE. Package version 0.1.0. Apache 2.0.

Setup effort

Run a container — pull the image and point your client at it

Deep Research