Several agents, one cited answer.
A command-line research pipeline over a notes vault: a planner splits the question, workers search in parallel, and a judge sends weak answers back.
In plain words
Ask one question of a large pile of notes and a single pass will miss half of it. This tool splits the question into parts, researches each part separately, writes one answer with its sources, and then grades its own answer. A weak answer is sent back for another try.
How it works
Ask a question
In the command line, in plain language.
Watch it work
The plan, the parallel searches and the draft appear live.
Get an answer with sources
It is graded first; below the pass mark it is revised before you see it.
The problem
One model call over a large body of notes either misses half of it or answers with confidence it has not earned.
What we built
A planner breaks the question into sub-questions and a router decides how deep each one goes. Workers read the vault through a standard tool protocol in parallel, a synthesizer writes a structured answer with sources, and a judge scores it. Below the threshold, the answer is revised.
How we know it works
| Evaluation set | 15 questions |
|---|---|
| Judged on | groundedness, relevance, completeness |
| Regression gate | the run fails below a set score |
- Built with
- TypeScript, Claude, Model Context Protocol
- Runs as
- A command-line tool; a demo vault is bundled
- Good fit for
- Research over internal notes and documentation
Under the hood
- PlannerSplits the question into 2 to 5 parts
- RouterQuick, deep or impossible, per part
- WorkersSearch the notes in parallel through MCP
- SynthesizerOne structured answer with sources
- JudgeScores it; a low score sends it back
It runs from a clean clone with a bundled demo vault and a mock model, so no API key is needed to see the pipeline.
Decisions and why
A small model plans, a larger one writes
Planning and routing use the fast model; synthesis and judging use the stronger one.
"Impossible" is a valid route
A part the notes cannot answer is marked as such instead of being forced into an answer.
The notes are reached through a standard protocol
Workers use MCP tools, so the same pipeline can be pointed at another source.
A score gates the run
The evaluation set fails the run when the judged score drops below the threshold.