Question 1

How is this different from an agent that just calls Jev?

Accepted Answer

Jev selects the attention mode and then evaluates each proposed next step against rubrics, with a state-bound single-use decision ticket gating tool permission instead of letting the text model act directly.

Question 2

Does it ride on my ChatGPT subscription, or do I need gateway credits?

Accepted Answer

Gateway credits. Both the text model and Jev go through the Vercel AI Gateway with AI_GATEWAY_API_KEY, and Jev uses the AI SDK evaluation API with the typesafe-ai/jev model by default; it does not reuse a ChatGPT subscription, scrape credentials, or run Pi or Codex as a nested agent. Your text model ID must be one the gateway actually lists, then passed to init, and it must support JSON-object output. The demo, tests, and doctor run without credentials; only doctor --live makes a paid call.

Question 3

Can I use it for security hunts safely?

Accepted Answer

Hunt boards persist rules, scope, dependencies, and routing history across restarts, but structural asymmetry is treated as a lead needing context rather than proof, and confidential workspaces need the documented security precautions.

Question 4

Does the demo prove the cost or accuracy claims?

Accepted Answer

No. The deterministic switchboard demo uses a fixture planner and mock evaluator to show wiring and enforcement only, with every mock evaluation labelled and stored as mock source.

Question 5

How is the hunt board staged and routed?

Accepted Answer

INTAKE, READY, ACTIVE, REVIEW, DONE, with BLOCKED branching off ACTIVE. Routing is availability-aware: Jev only ever sees currently idle specialists as candidates, so a card is not queued behind a busy worker, and review-stage cards avoid the original worker when another specialist is free. Admin drives it with native hunt_create, hunt_card, hunt_board, and hunt_route tools, and the /hunt commands show, run, pause, and resume boards.

Question 6

What does the research lane actually cover?

Accepted Answer

Read-only investigation with built-in semantic operations for code-surface mapping, symbol lookup, reference and caller tracing, function and modifier inspection, state read/write tracing, and structural function comparison. Smart-contract Solidity gets a bounded local structural scanner while TypeScript, JavaScript, Rust, and Move receive lighter function discovery. The README calls these research aids rather than a compiler or a proof engine, and structural asymmetry is recorded as a lead needing context rather than a vulnerability.

Question 7

What does the desktop surface add, and what are its limits?

Accepted Answer

A Tauri interface with Dashboard, Swarm, Profile Library, Memory, Tools, Settings, approvals, and a localhost runtime bridge exposing hunt-board state, running the same root runtime as the CLI rather than a second backend. It is experimental and Linux-first; the Omarchy installer builds the binary with --no-bundle and installs eutrya-desktop, keeping eutrya as the CLI command and adding eutrya-cli as a snapshot.

Eutrya

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community