Question 1

Can I try it without any API keys at all?

Accepted Answer

Yes, and that is the documented path. cargo test expects 60 passing tests, with tests/routing.rs driving the full Jev to route to catalog loop offline, including tool calls and tier filtering. Then cargo run with a plain English brief and a monthly-cents figure prints a route, its reason, the eligible model count and your remaining budget without any key.

Question 2

What does Jev decide here, and what does it not?

Accepted Answer

Jev decides which agent to use, scores complexity and sensitivity, and returns nouls for whether a private TEE, risk or urgency applies. The policy engine combines those with your budget in code and then filters the NEAR catalog by privacy tier, so the model proposes a route and your cap, not Jev, decides what you can afford.

Question 3

How much am I allowed to spend, and who controls the cap?

Accepted Answer

You do. The CLI takes a monthly cap in cents as an argument of any size, and the offline route prints the remainder against it. An estimate of the plan comes from the catalog rather than a real call, so the number is an estimate until you run with --live and a NEAR key.

Question 4

Is the privacy tier actually enforced anywhere?

Accepted Answer

The client verifies a TEE attestation report and its nonce, and fails closed in live mode when that check does not pass, which is a real client-side gate. What it cannot prove is the remote side's behaviour: an offline demo run exercises the loop with fixtures, so treat a passing check as verification of a report, not as a hardware guarantee.

Question 5

What does the Rust toolchain requirement look like?

Accepted Answer

Rust 1.70 or newer, tested on 1.96.0, with serde, serde_json and ureq as the dependencies. The layout is modular: policy.rs holds the budget by quality by sensitivity engine, budget.rs the client-controlled monthly caps, attest.rs the TEE report and nonce verification, and route.rs the orchestrator that ties the Jev answers to the catalog filter and the plan estimate.

Question 6

What happens to function tools passed through to NEAR?

Accepted Answer

They are forwarded verbatim rather than rewritten, so the tool definitions reach the model exactly as you wrote them and the tool calls flow back normally. The catalog itself comes from list_models with an estimated_cents() calculation attached per entry, which is what lets the budget check run before a completion rather than after it.

Question 7

Is there a server component, and what state is it in?

Accepted Answer

There is a design and not a build. docs/SERVER.md describes jear as a per-tenant router-proxy and is labelled as a design rather than something implemented, so nothing in that document is running code today. For new users the README points at docs/NEWBIE.md for a five-minute path to a first routed answer and docs/KEYS.md for where each key comes from and why nothing needs hosting.

Jear

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community