How to Set Up Jev AI With Claude Code, Cursor & Codex

Last verified September 20, 2026. Commands are quoted from TypeSafe's official skills repository. Re-check them if you are reading this much later.

What you need first

• A TypeSafe account and an API key, or access through a provider that carries Jev.
• Claude Code, Cursor, Codex or another supported coding agent.
• A project folder and a terminal.

Step 1 — Try the primitives before you wire anything

Jev has three question types and they are the whole interface. Spend ten minutes in TypeSafe's playground getting a feel for them before you integrate, because the integration is easy and the question design is the part that takes judgment.

• Noul — the probability a yes/no statement is true. "Is this customer asking for a refund?"
• Choice — one option from a list you define, up to 255. "Which queue: billing, technical, sales, other?"
• Score — a position on an ordered scale, up to 10 levels. "How urgent, 1 to 5?"

Step 2 — Install the official skill

TypeSafe publishes an agent skill so your coding agent knows how to build with Jev correctly. For Claude Code:

claude plugin marketplace add typesafe-ai/skills
claude plugin install typesafe@typesafe-ai

For Cursor, Codex and other supported agents:

npx skills add typesafe-ai/skills --skill typesafe-ai

Once installed you can invoke it explicitly with /typesafe:typesafe-ai. This is the step that makes the rest easy — without it, an agent writing Jev code is guessing at an API it has probably never seen.

Step 3 — Put the key in the environment

The SDKs read TYPESAFE_API_KEY from the environment. Set it there and nowhere else. Do not paste it into a source file, a config file you commit, or a prompt.

Step 4 — Tell the agent to use the skill

The skill only helps if the agent knows to reach for it. A starter prompt that works:

"Use the TypeSafe skill. Build a function that takes a support email and returns which queue it belongs in — billing, technical, sales or other — along with an urgency score from 1 to 5 and whether the customer is asking for a refund. Return the raw probabilities, not just the top answer."

That last sentence matters. The top answer alone throws away the most useful thing in the response.

Step 5 — Build one Noul, then add to it

Start with a single yes/no question against one piece of state and confirm you get a probability back. Then add a Choice. Then a Score. Then put all three in one request against the same state and watch the latency barely move — that is the parallel evaluation doing its work, and it is the moment the economics make sense.

The direct API, if you are skipping the SDK:

POST https://api.typesafe.ai/v1/systemone
{
  "model": "jev-latest",
  "state": "...",
  "questions": [ ... ]
}

jev-latest is an alias. Responses currently echo back the concrete version it resolves to. For anything you plan to keep, pin the version rather than the alias, so a model update does not silently change your results.

Step 6 — Choose a threshold, and know what you are choosing

Every automated decision needs a rule for when the probability is high enough to act. TypeSafe's published guidance on this is thin — a comment in their examples noting that 0.5 is a starting point and the right number depends on your use case.

No calibration curve appears in TypeSafe's public documentation as of September 20, 2026, which means a stated 0.9 has no public evidence attached about how often it is right. The practical consequence: pick a conservative threshold, run it against a set of examples you have labeled yourself, and move it only when you have measured what moving it costs.

The advanced pattern: routing before the expensive model

The integration worth building once you are comfortable is a router. Claude Code exposes a UserPromptSubmit hook that fires when a prompt is submitted, before the model processes it. That gives you a place to intercept.

The shape: a prompt arrives, Jev classifies what kind of request it is, and cheap requests get handled by a tool or a lookup while everything else continues to the full model. You pay frontier prices only for the work that needs frontier reasoning.

About the MCP servers you will find

Search for a Jev MCP server and you will find several — jev-mcp, typesafe-mcp and a number of forks. They are worth knowing one thing about: none of them is TypeSafe's. As of this writing there is no official MCP server under the TypeSafe organization. The official integration is the agent skill in Step 2.

That does not make the community servers bad. It does mean you are installing a third party's code, with your API key, into an agent that can act on your machine — and the usual questions apply. Who publishes it, what does the change history look like, what scope does it ask for, how does it handle credentials, and does it require approval before acting.

Troubleshooting

The two that actually matter are in the middle of that table. TypeSafe's own guidance is that reviewing your questions and thresholds closely is among the most important parts of an implementation, and that matches experience: setup takes minutes, and question design is where the quality lives.