Jev picks the tools each AI request needs.
From TypeSafe. Proved on our platform, in a client’s build 2 days later.
A banking question went from 19 tools to 8, a general one from 38 to 17. Each choice takes about 249 ms.
Out of 112 registered tools at the time.
Typical response. 95% arrive within 333 ms.
First tool-routing commit on our platform, then the same pattern in the client's AI layer.
The assistant on our own platform can call 112 tools: finance, banking, CRM, tasks, monitoring, legal research. Every tool definition sent with a request costs input tokens, and a model with too many tools to choose from picks worse.
Before Jev, a general LLM classified each message into 1 of 10 intents, and the intent decided the tool set. The sets were wide: a general question carried 38 tools, anything operational 39.
So the expensive model read dozens of tool definitions it would not use, on every message. We wanted it to see only the tools the request needs, chosen by something much cheaper and faster than itself.
TypeSafe released Jev on 15 September 2026. Five days later it was routing tools on our platform.
Step by step.
- 01
The 10 broad intents became 17 narrower tool groups.
- 02
Jev scores every group for each message. Its answers are typed, so a score always comes back in the shape the code expects. Groups with a high score open.
- 03
An open group stays open for the rest of the conversation while it is still relevant, so a follow-up like “do it” keeps the tools the previous turn needed.
- 04
Only the open groups reach the large model. A banking question now carries 8 tools instead of 19; a general one, 17 instead of 38.
- 05
When every score is weak, the assistant asks the person to confirm what they want instead of guessing.
- 06
If Jev fails or is slow, the request goes ahead safely without live data, and the answer says so. New Jev versions are tested before we switch to them.
Built with Jev by TypeSafe · Our AI assistant · Groups of tools · Our own database
Routing worked from the first week. The tool list the large model sees shrank by more than half on the questions we measured, and each routing decision took about 249 ms.
Our measurement matched TypeSafe’s. About 249 ms per routing decision on our platform sits inside the 70 to 500 ms per call TypeSafe publishes, against seconds for a general LLM.
A slow routing call could keep a chat waiting for about 30 seconds. We limited it to one short call per turn.
Jev reads literally and is strongest in English. We write exact conditions into each question, and at least half of every test set is in Spanish.
What we decided.
Adopted. Every request in our assistant goes through Jev before the large model sees a tool. Permissions still decide which tools a person may use; Jev decides which of those are relevant now.
Anywhere a system has to make a quick judgment about text, a small judgment model can do it in a fraction of a second, for far less than a large model. In a system we build for you, that can mean:
- Routing: the large model sees only the tools, documents or data each request needs.
- Triage: incoming emails and messages sorted by topic and urgency before anyone reads them.
- Lead follow-up: replies classified as interested, not now or not interested, so the next step goes out on time.
- Safety: risky actions flagged for a person’s approval before an agent runs them.
- Quality: an AI answer checked against its sources before it reaches a customer.
- Bookkeeping: transactions and expenses classified to the right account, with a person confirming the doubtful ones.
- Routing narrows the tool list; it never grants access. Permissions decide that.
- Arithmetic, dates and exact matches stay in code. Jev only judges meaning.
- If routing fails, the request goes on without live-data tools, never with all of them.
- Each new Jev version is tested before we switch to it.
Bring the same routing step to the client systems we build with large tool sets, starting with the one already running it.
What is Jev from TypeSafe?
TypeSafe’s first System One model, released in September 2026: a small model built for decisions inside software. It returns typed answers with probabilities instead of generated text. We use it to decide which tools a large model sees on each request.
How do you reduce the cost of AI tool calling?
Send fewer tool definitions per request. A fast judgment step picks the relevant tool groups first, so the large model sees 8 tools instead of 19, or 17 instead of 38. The judgment call itself is cheap: TypeSafe prices Jev’s input far below general LLMs and doesn’t charge for output.
Does routing tools add latency?
A little. A typical decision takes about 249 ms, 95% take under 333 ms, and we cap it at one short call per turn.
What happens if Jev is down?
The request goes ahead without live-data tools, and the answer says so. It never falls back to sending every tool.