On 15 September 2026 TypeSafe AI opened early access to Jev, a model that does not write text. Jev answers typed questions about the data you give it: a choice, a score or a yes/no probability, each with a calibrated distribution. Within twelve days developers had published 648 open-source repositories built around it. We audited every one of them. This report covers what Jev is, where its claims hold, where they don’t, and what those repositories are actually built with.
1. What Jev is
TypeSafe calls Jev a System One model: fast, narrow judgments, as opposed to the slow, generative “System Two” of an LLM (Wikipedia, TypeSafe docs). A request has three parts:
- State. The text or JSON being judged, such as a ticket, a page snapshot or a record.
- Questions. A map of typed questions:
- Choice picks one of up to 255 options.
- Score places the state on 2–10 ordered, described levels.
- Noul returns the probability that a statement is true.
- Model.
jev-latest, which currently resolves tojev-1.13.0.
Every question is evaluated in parallel against the same state, and answers come back under your question IDs. Choice and Score answers include per-option probabilities and a confidence value. TypeSafe lists the price at $0.042 per million input tokens with output free. A request can carry up to 64k tokens, and TypeSafe claims 70–500 ms responses. Official SDKs exist for JavaScript (@typesafe-ai/sdk) and Python (typesafe-sdk), and Vercel’s AI SDK has a provider (@ai-sdk/typesafe-ai).
2. The claims and the pushback
TypeSafe’s launch claimed Jev is 40–200× faster and 40–400× cheaper than frontier LLMs on comparable work, and that it “can’t hallucinate”. Reactions on Hacker News, from Simon Willison and in InfoQ’s coverage narrowed those claims:
- “Can’t hallucinate” means “can’t return an invalid type.” A Choice always names a real option, but it can name the wrong one.
- The speed comparisons aren’t like for like. An LLM that emits JSON and prose does more work than a model emitting one constrained decision. The latency is real; the multiplier depends on what you compare against.
- Probabilities are not invariants. Reordering options can shift the distribution, and TypeSafe’s own docs warn that a Noul and its negation, or a Noul and an equivalent Choice, need not add up.
- TypeSafe flags its own sample. The company said its speed and cost comparisons were built by its own team and are likely at the high end of real-world results.
TypeSafe’s docs are unusually direct about weak spots. The jev-1.13 jaggedness page lists literal reading, math and counting, date arithmetic, double negatives, large irrelevant state and adversarial content. Its advice for each is the same: do that part in code.
3. When to use Jev, an LLM, or plain code
| The task | Use |
|---|---|
| An exact rule, lookup, count, date or calculation | Code |
| Picking one of a known set, scoring on a rubric, checking whether a statement holds | Jev |
| Writing text, writing code, multi-step reasoning, explaining a decision | An LLM |
| Acting on an uncertain decision | Jev’s confidence decides: act, ask a human, or escalate to an LLM |
The pattern TypeSafe recommends, and most serious repositories follow, is code in control, Jev for judgments. Ask every question in one request, filter the state to what the questions need, and gate actions on confidence.
4. What 648 open-source Jev repos are built with
We merged the repository lists in awesome-jev-use-cases (658 unique entries), fetched each repo from GitHub and dropped forks, archived and missing repos. That left 648. For the 443 with a package.json, pyproject.toml or requirements*.txt, we checked which libraries they declare.
| Signal | Repos | Share |
|---|---|---|
| Primary language Python | 228 | 35% of 648 |
| Primary language TypeScript or JavaScript | 272 | 42% of 648 |
| Declares an official TypeSafe SDK | 134 | 30% of 443 |
@typesafe-ai/sdk (JavaScript) | 78 | 18% of 443 |
typesafe-sdk (Python) | 53 | 12% of 443 |
@ai-sdk/typesafe-ai (Vercel AI SDK provider) | 4 | 1% of 443 |
| MCP server or client SDK | 71 | 16% of 443 |
| React / Next.js / Vite front end | 72 / 31 / 42 | |
| Playwright (browser automation) | 25 | 6% of 443 |
| FastAPI | 23 | 5% of 443 |
| Also declares an LLM SDK (OpenAI, Anthropic or Vercel AI SDK) | 45 | 10% of 443 |
What stands out:
- Most projects call the HTTP API directly. Seventy percent declare no TypeSafe SDK, including the most-starred project, browser-use/jev-ultrafast, which uses
httpx, and fast-jev-compaction, a Claude Code plugin. The API is one POST with a JSON body, so many developers skipped the SDK. - Agents are the first market. One in six projects ships an MCP server or client, and the most-starred repositories are browser agents and coding-agent plugins. TypeSafe itself ships an official agent skill rather than an MCP server.
- Few repos pair Jev with an LLM yet. Only one in ten also declares an LLM SDK, even though the System One plus System Two split is the architecture TypeSafe describes.
- The ecosystem is young and long-tailed. 45 repos had 100 or more stars and 8 had 1,000 or more. 59% are MIT-licensed and 28% have no license at all, so check before you reuse code.
5. Three ways to build with Jev
Each stack has a free AGENTS.md, CLAUDE.md, Cursor rule and Agent Skill (SKILL.md). They cover the architecture around Jev; TypeSafe’s official skill covers the API itself.
| You are building | Stack | Rule |
|---|---|---|
| A web app where Jev routes, classifies and gates, and an LLM writes | Next.js + Vercel AI SDK + Jev | Next.js + AI SDK + Jev |
| A triage, routing or guardrail service other systems call | FastAPI + typesafe-sdk + Postgres | FastAPI Jev decision service |
| A browser agent that picks actions instead of generating them | Playwright + Jev (+ MCP) | Jev browser agent |
6. Takeaways
- Treat Jev as a typed function, not a chatbot. Its value is closed answer spaces, probabilities and latency.
- Keep math, dates and counts in code. TypeSafe says so itself.
- Use confidence as a second axis. Act on high-confidence answers, and send the rest to a person or an LLM.
- Pin and log the model version.
jev-latestmoves when TypeSafe ships a release, and thresholds tuned on one version may not carry over. - The ecosystem is two weeks old. Most repos are demos. Read the code and the license before depending on one.
Methodology. The sample is every repository in the repos.csv, more-repos.csv and new-repos-2026-09-26.csv files of awesome-jev-use-cases, deduplicated (658). Forks, archived repos and repos that no longer resolve were removed, leaving 648. Metadata and up to 20 package.json, pyproject.toml and requirements*.txt files per repo (excluding examples, tests and vendored code) were fetched from GitHub on 2 October 2026. A repo counts as using a library when any of those manifests declares it. Repositories that call Jev over plain HTTP are counted as not declaring an SDK.