Getting started with the Jev backend

Set it up

1. Install thunc

   The jev backend is in thunc 0.1.3 and later. It needs no extra Python packages.

   Copy

   pip install --upgrade thunc

2. Get a TypeSafe API key

   Create one in the TypeSafe console. You'll paste it into jev login in step 4, and nowhere else.

3. Install the jev CLI

   On a Mac with Apple Silicon, use TypeSafe's Homebrew tap. Homebrew asks you to trust a tap's formulas before it installs them; trusting only jev is enough.

   macOS (Apple Silicon)

   Copy

   brew tap model-clis/packages
   brew trust --formula model-clis/packages/jev
   brew install jev

   On Linux, or a Mac without Homebrew, the install script checks the download's SHA-256 and puts jev in ~/.local/bin. For Windows and other options, see the jev README.

   Linux and macOS

   Copy

   curl -fsSL https://raw.githubusercontent.com/model-clis/jev/main/scripts/install.sh | sh

4. Log in

   jev login asks for your key, checks it with TypeSafe and stores it in ~/.model-clis/jev/credentials.json, readable only by you.

   Copy

   jev login

   In CI or a container, set JEV_API_KEY instead. It takes precedence over the stored login. thunc never sends configure(api_key=...) to Jev, because that is usually the key of another backend.

5. Check that it works

   One call costs a fraction of a cent. A probability near 1 means the login works.

   Copy

   jev noul "Is the sky blue?" --state "A clear summer day"

   Output

   {"answers":{"q":{"noul":0.95,"type":"noul"}},"latency_ms":265,"model":"jev-1.13.0","ok":true,...}

6. Point thunc at it

   Pick Jev for one function with backend="jev", or for everything with thunc.configure(backend="jev") or THUNC_BACKEND=jev. thunc never picks it on its own.

   Copy

   from typing import Literal
   import thunc
   
   @thunc.function(backend="jev")
   def team(ticket: str) -> Literal["bug", "billing", "feature-request", "other"]:
       """Which team should handle this customer support ticket?"""
       ...
   
   team("I was charged twice this month.")  # -> "billing"

What Jev can answer

Jev picks from answers you define, so only these return types work. Any other type raises ThuncError before a request is sent.

For a rating, use Literal[1, 2, 3, 4, 5] instead of int with ensure=.

Use it next to Claude

Jev can't write text, so most programs use it for the decisions and another backend for the writing. Each function picks its own backend. Here Jev filters every message, and Claude drafts a reply only for the ones that need one.

thunc.configure(backend="anthropic")  # the default, for text

@thunc.function(backend="jev")
def needs_reply(message: str) -> bool:
    """Does this message ask a question or report a problem we should answer?"""
    ...

@thunc.function
def draft_reply(message: str) -> str:
    """Write a short first reply to this message. Never promise a refund."""
    ...

for message in inbox:
    if needs_reply(message):  # ~0.3 s, fractions of a cent
        print(draft_reply(message))

How thunc's options behave

inputs

Sent as Jev's state. The docstring or instructions become the question.

system=

Your own system prompt goes before the question as context. thunc's default prompt is written for text models, so it isn't sent.

model=

Ignored. The CLI always uses jev-latest, which the cache and trace record.

ensure=

A rejected answer raises ThuncError at once. Asking again would get the same answer.

cache=, trace=

Work as for any backend.

thunc.map

Works as usual. 8 workers stay under Jev's limit of about 20 requests per second.

Try the examples

From a clone of the repository, after step 4:

python3 -m examples.jev_hello        # a yes/no, a label and a rating
python3 -m examples.jev_inbox        # 8 tickets, 24 calls, about 1 s
python3 -m examples.jev_with_claude  # Jev decides, Claude Code writes the replies

To run thunc's live tests on Jev: THUNC_BACKEND=jev pytest live_tests. Tests that need text answers are skipped.

If something goes wrong