Researchguy

What it is

Researchguy is a command-line research tool. Given a question, it runs multi-backend research workflows, generates structured reports, and - the part that makes it more than a wrapper - keeps an evidence ledger: checkable claims extracted from documents carry source quotes and verification results.

The problem it solves

LLM research tools produce confident prose with citation-shaped holes. Researchguy treats the report as the visible layer of a pipeline that preserves where each fact came from: run records keep raw tool results on disk, claims carry the quote that states them, and contested or single-origin claims get flagged as such.

My contribution

I designed and built the whole system: the Go CLI, the multi-backend worker orchestration (Claude, Codex, Goose, Ollama, or hybrid fan-out), the claims pipeline, the knowledge graph, and the grepai-backed semantic search over saved research.

Architecture

• Workflows: deep research, literature reviews, comparative analysis, document enrichment, and one-shot Q&A, all searchable against existing research before generating.
• Evidence: run records persist raw tool results per run; claims are extracted with a local model, quoted, quote-checked, and linked across independent origins.
• Knowledge graph: entities, sources, claims, and reports persist as referenceable nodes with typed edges.
• Scheduling: cron-based recurring research with a background daemon, plus a batch task queue.
• MCP server: every tool exposed to MCP clients over stdio.

Try the sample

The report explorer on this site shows the claim-to-source structure with a small prepared example about hosting a static portfolio - no model calls happen in your browser.

Engineering decisions

Claims are checked, not trusted. A claim’s quote is verified against the document text as part of the claims workflow; quote_verified: false is a first-class, visible state.

Flags over silent synthesis. Contested, superseded, single-origin, and possibly-outdated are flags the tool attaches rather than judgments it hides.

Local-first storage. Reports and run records are stored as local files, while graph relationships use SQLite. Model backends and web research can still depend on external services.

Stack

Go, multiple LLM CLI backends, SQLite, grepai semantic search.