Skip to content

research-topic

🔮 Roadmap example. Depends on: a generic-web-search MCP server you've configured under mcp: in reyn.yaml, plus a web_research stdlib component that wraps it. Not runnable on Reyn v1 as of 2026-05-02.

Tracked in: post-OSS roadmap. Note: the mcp_search stdlib component that ships today is a registry search (it queries github.com/mcp for MCP servers matching a capability). It is not a generic web-research entry point — calling it with "What changed in DuckDB v1.0?" returns server candidates, not research findings. This recipe documents the intended end-to-end shape; the wiring still has to be built.

Run web research on a topic and get a structured summary back.

What this shows

  • Driving a stdlib component with natural language input.
  • How a workflow that needs an external capability (web search) is wired through MCP rather than baked into the OS (P7).

Prerequisites

  • An MCP search server configured in reyn.yaml under mcp:. The default recipe assumes a server named search exposing a web_search tool. If you don't have one, the run will fail fast at the first mcp_call op — fix reyn.yaml and retry.

Run it

reyn run mcp_search "What changed in DuckDB v1.0 vs v0.10?"

Or with explicit JSON input:

reyn run mcp_search '{"type":"user_message","data":{"text":"DuckDB v1.0 changes"}}'

Expected output

A final_output artifact with the synthesized findings plus citations to the URLs the search returned. Token / cost summary follows.

See transcript.example.txt for a sketch.

Variations

  • Different model: reyn run mcp_search "..." --model openai/gpt-4o-mini
  • Output in Japanese: reyn run mcp_search "..." --output-language ja
  • Inspect events: append --events to dump the full event log after.

Troubleshooting

  • mcp server 'search' not configured — add a mcp: block to reyn.yaml.
  • Empty search results — your MCP server's web_search tool may need an API key in its own env.

See also