Hi, and welcome to the first issue of The MrJev Weekly.

Jev has been public for four days. More than 1,000 GitHub repositories already mention it, so this week we read the code of 14 of the most popular Jev tools and ran ten of them in a sandbox. Here's what changed, and what we found.

What changed

Jev launched. On September 15, TypeSafe AI released Jev, the first "System One" model. You send it state and typed questions (Choice, Score, or Noul) and it returns answers with calibrated probabilities instead of text. It costs $0.042 per million input tokens, and output is free. Access is through an early-access waitlist. (Launch post · Our explainer)

$40M seed round. TypeSafe came out of stealth the same day with a $40 million seed led by DCVC. (Announcement)

Jev is on OpenRouter. typesafe/jev-1.13 is listed at the same price as TypeSafe's own API, but with a 32K context window instead of 64K. Handy if you already have OpenRouter billing. (OpenRouter)

Two SDK releases, both with breaking changes.

• JavaScript SDK 0.6.0: Score criteria are now an ordered array instead of an object keyed by integers. (Release)
• Python SDK 0.7.0: switched from msgspec to Pydantic, and system_one() gains a response_model argument that parses answers straight into your own Pydantic model. (Release)

TypeSafe published Jev's weak spots. A docs page lists nine known failure modes of jev-1.13, including arithmetic, date comparison, multi-hop questions, and large, noisy state. Read it before you build anything. (Jev 1.13 jaggedness)

Pin jev-1.13.0 instead of jev-latest once you've tuned anything.

We reviewed 14 Jev tools

Our new roundup covers what each tool is good for, what it sends where, and what to check before installing. Read the Best Jev Tools roundup →

Our picks:

• Give Claude Code or Codex a Jev tool: typesafe-mcp, one binary and one setup command.
• Cheaper models on easy turns: jev-router lets Jev pick the Claude or Codex model for each prompt, and never downgrades you when it isn't sure.
• Fast browser automation: jev-ultrafast from the Browser Use team. Point it at a separate browser, because by default it drives your everyday Chrome.
• Code review without a generative model: Jev Review screens JavaScript and TypeScript diffs with a chain of small Jev questions. Its prompts are a great example of how to write Jev questions.
• No data to TypeSafe at all: SemIf (formerly OpenJev) and Jevlike rebuild Jev's interface on open models you run yourself.

And for fun, TypeSafe Mario has Jev playing Super Mario Bros. from parsed game state, no screenshots.

What we found

• "Only the prompt" rarely means only the prompt. Depending on the tool, requests to TypeSafe include your git diff, code your agent wrote, visible page text, or the contents of on-screen form fields. The roundup has a table of what each tool sends and where.
• Keys end up in more places than you'd think: client config files in plain text, or the environment of a coding agent running with approvals off. Use a separate Jev key for each tool.
• The good ones default to not acting. The best projects start in a dry run, cap their steps, stop when Jev isn't sure, and hand off to a person when something fails.

We've reported most of what we found to the maintainers, and some of it is a one-line fix.

One practical tip: measure confidence before you trust it

Jev returns a confidence score with every Choice and Score answer, and it's tempting to automate anything above 0.9. But calibration depends on your data. The Janus project measured Jev's confidence on two public datasets, and its author found that the same confidence level was noticeably more reliable on one than the other.

Before you let confidence gate a real action:

1. Label 100 to 200 real examples.
2. Run them through Jev and check accuracy at each confidence level.
3. Pick the threshold that meets your error budget, and send everything below it to a person or a stronger model.

jevcal automates exactly this, and can fail your CI when a model update shifts the numbers. Our confidence thresholds guide walks through it by hand.

That's it for this week. Hit reply and tell me what you're building with Jev, or which tool we should review next. I read every reply.

MrJev

MrJev is independent and not affiliated with TypeSafe AI. Numbers attributed to projects are their authors' own.