Official tools
Jev / TypeSafe AI
New to Jev? Start here to understand its three answer types, what you can ask it to decide, and how to try a small task before connecting a larger workflow.
Best for
Testing a first classification or routing task
TOOLAI / Jev · tools & integrations
What can Jev do?
How should you connect it?
Some jobs need a decision before they need a written answer. Here is how Jev fits into an application, which tools help you use it, and where each option makes sense.
A decision workflow
01 / Incoming message
“Can I upgrade my team’s plan?”
JevChoice
Which team should handle this?
Sales0.88
Route to the right team
Human review
01 / JEV
Start with the question your software needs answered.
An email arrives asking for a refund. Before anyone writes a reply, your system needs to know whether it belongs with billing, support or sales. Jev can make that kind of choice from the options you define. Your application still decides what to do with the result.
Choice
Choose a category
Choice selects among defined options and returns probabilities and confidence.
Score
Rate against a rubric
Score evaluates against defined levels and returns a distribution and confidence.
Noul
Assess a statement
Noul returns a probability between 0 and 1 for a proposition.
02 / Jev · tools & integrations
Read about the tools before choosing one.
These tools do different jobs. The SDKs call the model, gateways provide another access route, and agent skills help you write the application. Open a guide for a worked example, setup steps and the tradeoffs.
Official tools
TypeSafe Python SDK
If your data processing or backend already runs on Python, this is the direct route. Read how to prepare questions, handle results and test them against human labels.
Best for
Python services, data pipelines and evaluation scripts
Official tools
TypeSafe JavaScript SDK
For a Node.js or TypeScript application, the official client keeps the decision close to your existing code. See where it belongs on the server and how its types help.
Best for
JavaScript and TypeScript server applications
Model gateways
OpenRouter
Already managing several models with OpenRouter? You can keep that access route, but Jev uses a decisions interface. The guide explains what changes in your requests.
Best for
Teams already managing models through OpenRouter
Model gateways
Vercel AI Gateway
If your application uses AI SDK or AI Gateway, Jev can fit into that setup. Read which interface to choose and why example versions matter.
Best for
Applications already using Vercel AI SDK
Model gateways
Cloudflare AI
For a service running on Workers, Cloudflare AI offers a route to Jev. Learn what the binding handles, what still belongs in your code, and how to check access.
Best for
Decision steps inside a Cloudflare Workers application
Agent development
LangChain
Use Jev for a bounded choice inside an existing chain or agent. The guide walks through routing and explains why classification alone cannot authorize an action.
Best for
Adding a decision step to an existing agent workflow
Agent development
TypeSafe Agent Skills
These skills give a coding assistant project guidance. Read what they help it build, and why installing a skill does not connect or run the model for you.
Best for
Building with a coding assistant
Each guide explains the connection to Jev, links to related material on ToolAI, and provides a clearly marked route to the official project.
Which route fits the way you already work?
You do not need all eight tools. Start with the language, hosting platform or framework your team already understands, then add only the pieces your task needs.
You want to understand Jev first
Read the Jev overview and try a question with a small, explicit set of answers. A single task—such as deciding which team should read a message—is easier to check than a complete agent.
You already have a Python or JavaScript service
Use the matching official SDK to call TypeSafe directly. Your code prepares the context, reads the typed answer and handles errors. This keeps the request path straightforward.
Your team already uses a model gateway
OpenRouter, Vercel and Cloudflare provide documented routes. Look at the account you use, the interface the provider expects, and the current billing terms. These routes are alternatives; you do not need to chain them together.
You are adding decisions to an agent
LangChain can place the classifier in a workflow you already run. Agent Skills serve a different purpose: they help a coding assistant design the application. Neither replaces the permissions in your business code.
03 / A decision workflow
Follow one email from arrival to a useful label.
A useful starting point is to suggest a category while a person keeps handling the replies. The examples below describe a possible implementation; they do not turn on features in ToolAI Mail.
Define the task
Write down what each category means. “Customer enquiry” could mean a question about buying your product; “existing customer” needs evidence such as an account or order match. Include a route for messages that do not fit.
Evaluate real examples
Keep a set of messages that someone has already checked. Compare the model’s labels with those answers, especially where sales pitches resemble real enquiries. Test each language you receive.
Connect the action
Start by adding a suggested tag and a reason for review. Compare the result with account or order records before treating a sender as a customer. Only connect further actions once you know the common mistakes.
Three messages that need different treatment
| What the message says | What to establish | A sensible next step |
|---|---|---|
| “We bought the team plan last month. Can you add seats?” | A purchase is claimed, but the email alone does not prove an account relationship. | Check the sender against account or order records, then suggest a sales or customer-success queue. |
| “We can place your website on 200 high-authority blogs.” | This is a likely backlink sales pitch, even if it uses words such as partnership. | Suggest a marketing or backlink-sales label. Review ambiguous messages before filtering them out. |
| “My payment went through, but I cannot use the service.” | The message combines an access problem with a billing claim. | Look up the account and payment, then route it to support with the relevant context. |
A model label and a business fact are different things. To find real customers, combine what the email says with records you already hold. If those records are missing or conflict, keep the message for review.
CHOICE / 0.88
Try a review threshold.
This sample Choice result has confidence 0.88. Move the threshold to see how a routing rule changes.
Send for human review
Illustration only. Confidence is not a guarantee of correctness; select a threshold using your own evaluation data.
04 / TOOLAI
Keep exploring on ToolAI.
The model, the code and the context—in one place.
05 / FAQ
Before your first integration.
Can Jev write email replies or articles?
Jev returns typed decisions, not generated prose. Use it to choose a category, assess a condition or score an item; use a separate generative model or a person to write the response.
Does a multilingual guide mean Jev supports every language equally?
No. English is Jev’s primary training language. Evaluate Chinese and other languages using representative examples from your own workload.
Is Jev free or available to run locally?
The official service bills input tokens; free output does not mean free requests. The SDKs are API clients, not downloadable model weights. Check current prices and access terms for the provider you choose.
Does a typed answer guarantee the right decision?
No. A result can match the required type and still be wrong. Keep evaluation samples, review uncertain cases, and define permissions and fallback behavior in application code.
Read more before you connect anything
The detailed guides include the official sources checked on 2026-09-25. Their examples are suggested workflows, not measured performance results.