How to Use Jev With Claude Code, and the Mistakes to Avoid

Jev, from TypeSafe AI, is a decision model that reads text or JSON and returns typed answers, such as yes or no, a choice, or a score, with calibrated probabilities, and never writes prose. Use it with Claude Code for narrow, repeated judgments like filtering, ranking, and checking. Leave writing to Claude, and avoid it for maths and adversarial inputs.

What Jev is

Jev is the first model from TypeSafe AI, a San Francisco lab, launched in mid-September 2026. Firecrawl's overview notes the company was founded by Diogo Almeida, a co-author of the InstructGPT paper. TypeSafe calls Jev a System One model: it reads state, meaning text or JSON, plus predefined questions, and returns typed answers with calibrated confidence. Its documentation lists three question types: Noul for yes or no, returning a single probability; Choice, returning the selected option with a probability for each; and Score, a position on a rubric with probabilities per level.

It does not generate explanations, text, or code. Flavio Copes describes it as a smart if statement for software.

Speed and price, as published

According to TypeSafe's figures reported by Flavio Copes and Firecrawl, Jev answers in roughly 70 to 500 milliseconds end to end, and list pricing at launch was 0.042 dollars per million input tokens with output free. TypeSafe also published comparisons claiming large speed and cost advantages over frontier models on classification tasks; Firecrawl points out the launch post's own caveats that the workflows were built by TypeSafe staff. Treat headline multiples as vendor claims and measure on your own tasks.

Connecting it to Claude Code

TypeSafe's documentation lists several routes: a Claude Code plugin installed through the plugin marketplace, agent skills for coding agents, a Python SDK, and a REST endpoint. Coverage also mentions access through gateways such as Vercel's AI Gateway and OpenRouter. Signups opened publicly in September 2026 and were paused for a period because of demand, so check the current access status in TypeSafe's console before planning around it.

Whichever route you use, Jev runs as a separate hosted API with its own key. Claude Code calls it as a tool when a step needs a decision rather than prose.

Where it fits in a Claude Code workflow

The pattern is division of labour: Claude reasons and writes; Jev makes many fast, narrow judgments. Uses described in coverage include reranking search results before the agent reads them, verifying whether a source supports a citation, judging proposed tool calls before execution, routing prompts to an appropriately priced model, and classifying documentation pages during a crawl.

The token saving comes from not making Claude read everything. Instead of loading hundreds of search results or files into context to decide which matter, Claude asks Jev to score them and reads only the top few.

The mistakes that waste it

  • Asking it to write. Jev returns typed answers only. Anything needing prose, code, or explanation belongs to Claude.
  • Using it for maths, counting, or dates. TypeSafe's own list of weak spots includes arithmetic, counting, date comparison, and literal reading. Do these in code.
  • Treating it as a security boundary. Adversarial content is on the documented weakness list. Screening for prompt injection with Jev can reduce noise, but it should not be the only control on untrusted input.
  • Feeding huge, noisy state. Large noisy inputs degrade results. Send the relevant fields, not whole documents.
  • Ignoring the probabilities. The calibrated confidence is the point. Set thresholds, act automatically only above them, and escalate the uncertain middle to Claude or a person.
  • Asking open questions. Jev works best choosing from a defined set. Pick a card from the deck instead of asking it to name one.

Frequently asked questions

What is Jev?
Jev is a decision model from TypeSafe AI, launched in September 2026. It reads text or JSON and answers predefined questions with typed results, such as yes or no, a choice from options, or a rubric score, each with calibrated probabilities. It never generates prose or code, which keeps it fast and inexpensive for narrow judgments.
How do I add Jev to Claude Code?
TypeSafe documents a Claude Code plugin installed from its plugin marketplace, plus agent skills, a Python SDK, and a REST API. You need a TypeSafe API key or access through a supported gateway. Check TypeSafe's documentation and console for current install commands and signup availability, which changed during its launch period.
Does Jev replace Claude?
No. Jev makes narrow decisions; Claude reasons, plans, and writes. They complement each other: Claude delegates repetitive judgments, such as filtering or ranking many items, to Jev and reads only what matters. Tasks needing explanation, code, multi-step reasoning, maths, or counting stay with Claude or with ordinary code.
Is Jev good for detecting prompt injection?
It can help screen content quickly, but TypeSafe lists adversarial content among Jev's weak spots, so it should not be the only defence. Combine any classifier with structural controls: narrow agent permissions, approval for consequential actions, and treating fetched content as untrusted data.
How much does Jev cost?
At launch, TypeSafe's list price was reported as 0.042 dollars per million input tokens, with output free, according to Flavio Copes and Firecrawl. Prices and any free credits can change, and gateways may price differently, so check TypeSafe's console or your gateway for current rates before estimating costs.