What separates this from a code-generation prompt is the last step: the generated script goes into a Modal sandbox and the output comes back with the answer, so a script that does not run is caught before you see it. Sandboxes are spawned on demand with a small package set, and anything else is installed first — expect that install to show up as latency on the first run of a heavier request.
A multi-agent research and code assistant, exposed over MCP by Gradio. A question is broadened into sub-questions, searched on the web, turned into Python, executed in a remote sandbox, and returned with the output and citations attached.
- Seven agents in the documented protocol schema, callable as a workflow or individually — `question_enhancer`, `web_search`, `llm_processor`, `citation_formatter`, `code_generator`, `code_runner`, `orchestrator`
- A question decomposed into focused sub-questions before any search runs — the default is 3, adjustable up to 10
- Grounded answers: web results retrieved through Tavily, then formatted into APA citations
- Generated code executed in a Modal sandbox rather than assumed correct — small footprint by default (pandas, numpy, requests, scikit-learn), with extra packages installed on demand
- A hosted endpoint to connect to over `mcp-remote` — `https://agents-mcp-hackathon-shallowcoderesearch.hf.space/gradio_api/mcp/sse` — plus a Cline SSE config
To run your own instance: Python 3.12, an LLM provider key (Nebius is the default, with OpenAI, Anthropic and HuggingFace supported) set through `LLM_PROVIDER` and the matching key, a `TAVILY_API_KEY` for search, and a Modal account with `MODAL_ID` and `MODEL_SECRET_TOKEN` for code execution. The hosted Space is configured already and is reached with `npx mcp-remote` against the SSE URL. In VS Code, the README suggests naming the agent function in your prompt so the right tool is selected.
One command plus a key — npx mcp-remote https://agents-mcp-hackathon-shallowcoderesearch.hf.space/gradio_api/mcp/sse, then supply credentials