Start with a question

• “Find competitors for my product and compare their public capabilities, pricing and positioning.”
• “Which open-source tools could become my private AI second brain?”
• “Find original videos about a topic and compare the evidence for what gained traction.”
• “Investigate this technical claim. Find supporting evidence, counterexamples and what is still unknown.”

Radar discovers sources, inspects their content, asks the same research questions across records, and builds an evidence-linked comparison. You can inspect every saved Jev decision, open its supporting passages, filter results, replay the investigation, and export your findings.

What Jev does here

Part	Responsibility
Jev	Selects a research method, accepts or declines proposed questions and searches, chooses which observed leads to inspect, assesses collected evidence, and checks answer claims.
Search and browser tools	Find public URLs and collect permitted page content. Independent selected leads can be inspected concurrently.
Optional text model	Proposes goal-specific questions and searches, then drafts an answer from saved findings for Jev to check.
Radar	Enforces budgets and acquisition rules, preserves source links and unknowns, and shows actual activity, request timings and estimated spend.

The live scene follows recorded requests and decisions. Multiple requests appear together when they are actually in flight. Saved replay makes no new provider calls. A comparison matrix shows how each inspected record answers your questions, including missing evidence.

Jev is hosted by TypeSafe. Radar itself runs on a normal CPU machine, with no GPU frameworks or local model weights. Jev does not supply a search index; automatic discovery uses Brave Search.

Get started

Requirements: Python 3.11+, Node.js 22.12+ (or Node 20.19+ within Node 20), npm and Git. Use Linux, macOS or WSL.

git clone https://github.com/Eliovp-BV/Jev-Radar.git
cd Jev-Radar
./scripts/install.sh

The installer creates .env from the blank template if it is missing. Open that file locally and add your keys:

Key in .env	Purpose
TYPESAFE_API_KEY	Required for Jev's decisions and evidence assessment.
BRAVE_SEARCH_API_KEY	Required for automatic discovery from a prompt.
OPENAI_API_KEY, GEMINI_API_KEY, ANTHROPIC_API_KEY or OPENROUTER_API_KEY	Optional: add only your chosen text provider's key.

Keys stay on the server. They are never entered in the browser or included in source control. An OpenAI-compatible provider is also supported; see the configuration guide.

./scripts/run.sh

Open http://127.0.0.1:8787, or use this machine's LAN IP from another device. The default bind address is 0.0.0.0; the launcher prints the actual port.

Trusted networks only: there is no login, so anyone who can reach the app can read research and start paid requests. Set RADAR_HOST=127.0.0.1 for local-only access. Security details.

1. In Settings → API connections, reload keys. If using a text model, select its provider and model, check its price estimates, and save.
2. Return to Live → What are you looking for? Enter a goal and press Start research.
3. Watch the investigation, then open Results for findings, comparisons, source passages and remaining unknowns.

Settings → Spend limits defaults to $5 for Jev and $20 for the text model per investigation. These are estimated ceilings, not typical costs or provider billing caps. Search charges are separate; page, request, token and time limits also apply.

Chromium is optional for JavaScript page rendering:

./scripts/install.sh --with-browser

For provider setup, model options, network access and troubleshooting, see Setup. The app also includes a searchable Guide.

A real investigation

The included media comes from a new public research run:

Find open-source AI tools that turn my documents into a private, searchable second brain. Compare local-model support, document search, self-hosting and license evidence using official sources.

The recording shows the actual interface at normal speed. It uses real Brave discovery, hosted Jev decisions and a configured text model. No results, decisions or timings were scripted. Read the demo notes for its scope, measured results and limitations. The demo database is not included: your installation starts empty.

See the starting screen and research results

Built for inspectable research

• Goal-specific research: product landscapes, technical artifacts, content, individual videos and open questions.
• Evidence you can follow: saved source excerpts, collection coverage, Jev decisions, uncertainty and review states.
• Collection comparisons: common questions across records, filters, supported and unknown cells, and measured assessment timings.
• A visible workflow: live spatial visualization, concurrent activity, pause/resume controls and recorded replay.
• Your research stays organized: local SQLite persistence, saved investigations, reruns, imports and retention controls.
• Portable results: Markdown, HTML, CSV, structured JSON and a research-matrix CSV.

Research is bounded. Search appearance is not popularity; a supported passage is not independent proof; inaccessible content remains unknown. Video research assesses the text and metadata actually collected—it does not watch video or infer unseen footage. A run may finish with partial coverage. See current limitations.

Documentation and contributing

Read	For
Setup	Installation, keys, providers, settings and troubleshooting
Architecture	Research loop, Jev integration and modules
Data provenance	Evidence, imports, metrics and export semantics
Development	Tests, UI checks and opt-in live checks
Contributing	Changes and clean public-source releases
Security	Network boundaries and data handling
Limitations	What the application can and cannot establish

# After installing with --with-browser:
./scripts/test.sh

Normal tests use isolated fixtures and block unmocked model calls. Paid live checks are separate and opt-in. Keep .env, databases, imports, exports, browser profiles and local notes out of commits. Public demo media is explicitly reviewed and allowlisted by the source exporter.

License

MIT · Copyright © 2026 Eliovp BV. Dependency and service acknowledgements are in Third-party notices.