Jev is a System One model from TypeSafe, and the shortest accurate description is that it turns application state into typed judgments that code can branch on. It is not a chat model, and it is not a replacement for the coding model in your editor. Nearly every confused question about Jev comes from expecting one of those two things.
This page is the index for everything we have written about it, in the order that makes sense if you are starting from zero. If you only read one section, read the table below.
The useful output is not prose. Jev answers arrive as one of three question shapes, documented on the TypeSafe primitives page:
That shape is the whole point. A threshold in code can act on a probability. Parsing a sentence to find the answer is the pattern Jev exists to replace. The documented HTTP call posts your state to `api.typesafe.ai/v1/systemone` and can carry several independent questions in one request.
TypeSafe publishes one official agent skill and two official SDKs, for JavaScript and TypeScript and Python. The installation guide is explicit about the choice it wants you to make:
Use one method per agent. Running the plugin and the skills.sh install into the same agent is the second most common way to get a confusing result.
Treating Jev as the coding model. TypeSafe says directly that Jev is not a drop-in replacement for a coding LLM, and this is where most first attempts go wrong. In OpenCode, Jev does not power the editor; it handles narrow judgments alongside the model that writes code. The OpenCode walkthrough exists mainly to draw that line clearly.
Assuming one environment variable. The official SDKs read `TYPESAFE_API_KEY`. Some host integrations and adapters read a different variable, and the error you get when the wrong one is set looks like an authentication failure rather than a configuration mismatch. Check which of your four possible configurations is actually active before you check the key itself.
There is a third, quieter mistake: installing the skill and expecting it to make calls. The skill changes how your agent designs and writes; your code still needs the key at runtime.
Our rule, and it is a rule about consequences rather than about model quality: if the step applies an explicit rule to known fields, write code. If the step needs a bounded judgment over messy context, use Jev. If the step needs prose or open-ended reasoning, use a text model.
Then measure. A typed judgment is only worth wiring into a consequential path if you have checked its error and abstention behaviour on your own cases, not on a vendor example. The approval gate pattern is what we do with the uncertain band: the judgment proposes, code validates, and a person owns the irreversible step.
If you want to see how the wider community uses it, the awesome-typesafe-jev list is a curated, source-linked directory of projects, and it is honest about what is community work.
Is Jev free?
Access and pricing are TypeSafe's to set and they change. Check the current terms at the point of signup. This page does not quote prices for a product we do not sell.
Does Jev replace my coding agent or my LLM bill?
No. It replaces the prompt-and-parse step where you asked a text model for a structured answer. In most teams that is a small share of token spend and a large share of the unreliable behaviour.
Can I use it without an agent skill?
Yes. The SDK path and the HTTP endpoint do not need a coding agent at all. The skill is a convenience for people building inside one.
Where do I start if I only have ten minutes?
Install the skill in the host you already use, then make one Noul call on a real decision from your product and look at the probability it returns. That single call tells you whether the shape fits your code.
Reading seven guides is the easy part. The work is deciding which three decisions in your product become typed judgments this quarter, wiring the thresholds so a wrong answer is contained rather than shipped, and leaving a maintainable boundary behind. That is an engineering judgment call, and it is the thing we hire for.
Maxpertise embeds an AI engineer inside your existing team for that work, starting from a written intake rather than a discovery call. If you are weighing it against hiring, what an AI engineer costs is the honest starting point, and our services page shows both lanes.
[1] TypeSafe, typesafe-ai/skills, accessed October 12, 2026. One official skill, MIT licensed, with both install paths quoted on this page.
[2] TypeSafe, Agent skill guide, accessed October 12, 2026. The guide states "Use one installation method", names `/typesafe:typesafe-ai` as the explicit invocation, and documents the update routes for each install method.
[3] TypeSafe, Quickstart, accessed October 12, 2026. `POST https://api.typesafe.ai/v1/systemone` with `Authorization: Bearer $TYPESAFE_API_KEY`, and the invitation to mix Noul, Choice and Score questions in one call.
[4] TypeSafe, Primitives, accessed October 12, 2026. The three question shapes and their result shapes.
[5] TypeSafe, typesafe-sdk-js and typesafe-sdk-python, accessed October 12, 2026. Official clients, MIT licensed.
[6] The TypeSafe statement that Jev is not a drop-in replacement for a coding LLM is quoted in our OpenCode walkthrough, which verified it against TypeSafe's own material on October 11, 2026.
[7] AbdelStark, awesome-typesafe-jev, read through the GitHub API on October 12, 2026. A community directory that labels official and community entries separately.
Maxpertise is an AI-native engineering company. We embed native AI engineers inside your team, live in about 10 days. Please enable JavaScript to view the site, or email contact@maxpertise.net.