jev-deepresearch
Depth-crawling research agent where Jev makes every per-page relevance and next-link decision and an LLM only plans and writes.
Overview
Unfold is a depth-crawling research agent built to test what a System One model is good for. One question goes in, pages fan out across a canvas, each is scored for relevance, and the survivors converge into a written report. Qwen 3.6 Plus is called twice per run to plan searches and write the report; Jev 1.13.0 is called once per page to decide whether it is relevant and which link to follow next; SearXNG finds pages.
The design point is that every per-page decision is a choice between known options, so the expensive model runs twice instead of dozens of times. The README estimates a depth-8 run costs roughly a cent. It needs a TypeSafe key and an OpenRouter key, plus access to a SearXNG instance with JSON output (the author tunnels to one over SSH). All outbound calls come from the server, not the browser.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is this page relevant to the research question? | — |
| Choice | Which link should the crawl follow next? | — |
How it works
- Qwen turns the question into 4 to 6 searches
- SearXNG returns pages
- Jev scores each page for relevance and picks the next link
- Qwen writes the report from surviving pages
Strengths
- Expensive LLM called twice per run instead of per page
- README estimates about a cent per depth-8 run
- Keys stay server-side
Caveats
- Requires a SearXNG instance
- Needs both TypeSafe and OpenRouter keys
- No licence stated
Getting started
`npm install`, copy .env.example to .env with TypeSafe and OpenRouter keys, set SEARXNG_URL (or `npm run tunnel`), then `npm start` and open http://localhost:8790.At a glance
- Type
- Project
- Status
- prototype
- Language
- JavaScript
- Runtime
- Node server with browser canvas
- Has code
- yes
Use cases and integrations
How curated lists describe it
Deep research crawl where Jev (a System One model) makes every per-page decision and an LLM only plans and writes.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide?
Whether each page is relevant and which link to follow next.
Why not let the LLM do it?
Per-page decisions are choices between known options, so Jev handles them and the LLM is called only twice.
What does a run cost?
The README estimates roughly a cent for a depth-8 run.
Related projects
Toy research agent where Jev makes every keep/skip and delve decision over search results, and one expensive call writes the report.
ProjectprototypePython- jev-social★ 169
Local social research agent: Jev picks each bounded Instagram, TikTok or LinkedIn step and socai CLI runs it in your Chrome.
CLIbetaTypeScript Research agent where GPT plans and Jev finds evidence in parallel across document regions with Choice and Noul.
Browser extensionresearchTypeScriptWeb search CLI for agents that has Jev score results and scraped chunks against the query and goal, passing only relevant ones on.
BenchmarkbetaGoMulti-agent orchestration layer for Claude Code with opt-in Jev hooks for fast judgment points, decision caching and egress controls.
PluginproductionTypeScript- winnow★ 110
Calibrated context sieve for Claude Code: Jev judges each block of large tool output and hides confident-no blocks behind recallable stubs.
Projectbeta