Jev PlaygroundIndependent playground

JEV GETTING STARTED

How to Use Jev

You can try Jev online, call it through the TypeSafe API, use an official SDK, or give your coding agent the TypeSafe skill. Choose the path that matches what you want to do.

Choose how to use JevTry Jev now
PLAYGROUND TUTORIALNo visitor signupNo visitor API keyReal Jev result

00 / START HERE

Choose how you want to use Jev

Start in the browser to understand the output, or move directly to the integration path that matches your project. The no-signup note applies to this independent Jev Playground tutorial; Direct API and SDK access require your own TypeSafe account and API key.

PATH 1 · TRY ONLINE

Try Jev in your browser

Best for: Learning what Jev returns before creating an API account.

  • No visitor signup or API key
  • Real Jev model
  • Editable State and typed Questions
  • Visible answer probabilities
Start the live tutorialOpen full Playground
PATH 3 · OFFICIAL SDKs

Use an official SDK

Best for: Python or JavaScript/TypeScript applications.

Pythonpip install typesafe-sdk
JavaScriptnpm install @typesafe-ai/sdk

Both official clients read TYPESAFE_API_KEY from the environment.

See API & SDK examplesPython examples
PATH 4 · CODING AGENT

Give your coding agent the TypeSafe skill

Best for: Building a Jev integration with Codex, Claude Code, or another supported coding agent.

npx skills add typesafe-ai/skills --skill typesafe-ai
Claude Code plugin commands
claude plugin marketplace add typesafe-ai/skills
claude plugin install typesafe@typesafe-ai

The skill teaches the coding agent current TypeSafe API patterns and integration conventions. It does not turn Jev into a coding agent.

Open official Agent Skill docs (opens in a new tab)

How Jev access works today

Trying Jev here uses this independent Playground's server-side access, so visitors need neither a signup nor an API key. Building with TypeSafe Direct starts at the TypeSafe Console and requires an account with API-key access. TypeSafe's September 15 launch announcement described early access and a waitlist; today its SDK docs link directly to key creation. That does not establish immediate API access for every account.

Jev is hosted: there is no TypeSafe Jev model to download for local use. HTTP needs no SDK; the Python and JavaScript SDKs are client libraries, and the Agent Skill gives a coding agent integration guidance. None is the model itself.

GoalStart here
I just want to try JevJev Playground
I want to add Jev to my appTypeSafe API
I write Python or JavaScriptOfficial SDK
I want Codex/Claude to help integrate itTypeSafe Agent Skill (opens in a new tab)

LIVE PLAYGROUND TUTORIAL

Learn the browser-based Jev Playground path

The rest of this tutorial teaches State, Question, decision types, probabilities, and the application boundary through the live browser Playground. For API or SDK integration, use the paths above.

No TypeSafe account is needed for this Jev Playground tutorial.

For a worked agent-history decision, try the editable Context Compaction tool.

For a bounded code-change review, screen an editable diff and inspect original hunks.

To find evidence in text you provide, search original passages without a generated answer.

  1. 01LOADStart with a complete example
  2. 02RUNSend it to the real Jev model
  3. 03UNDERSTANDRead the decision and probabilities
  4. 04EDITChange one fact and compare

LIVE PLAYGROUND TUTORIAL · 01 / RUN YOUR FIRST DECISION

Your first Jev decision is ready

The Support Routing example is already loaded. Its State describes a duplicate subscription charge. Its Choice question asks which team should handle the request, with four allowed answers. No setup. No TypeSafe account. Example ready.

The example is already filled in. Click Run Jev first. You do not need to change anything.

1 Confirm Support Routing is selected.2 Press Run Jev.3 Read the returned decision and probability bars.
CHOOSE AN EXAMPLE — READY TO RUN
Evaluation inputs
INPUT

Example loaded — run it as-is, or edit any field.

118 / 8,000

The context Jev will read — for example a message, record, or report.

The decision you want Jev to make. The next step defines the shape of the allowed answer.

Decision type
SELECT ONE

Jev chooses one label from your options and returns probabilities for each.

Choices4 of 10

Run the loaded example now — no changes needed.

Free to try · Usage limits apply

OUTPUTTypeSafe Jev
Your Jev decision will appear here.

Your Jev decision will appear here.

Give Jev a state and a typed question.
Run it to see the decision and probabilities.

State + QuestionJevDecision + Probability
Real model execution. No simulated answers.

Run the loaded example, then edit the inputs or choose another example to experiment.

What just happened?

The Playground sent the visible State and typed Question to the real Jev model. Jev selected one of the four choices and returned its probability distribution. The result panel renders that live response; this tutorial does not predict or hard-code the answer. Actual probabilities can vary between evaluations.

Step-by-step browser tutorial

02 / WRITE THE STATE

State is the evidence Jev can use

State describes the situation at decision time. Include the facts that separate one answer from another. Clear evidence matters more than length; you do not need prompting tricks or a long persona.

  • Customer message
  • Email
  • Incident report
  • Document
  • Agent status
  • Browser page description
GOOD STATE

“The customer says they were charged twice for their monthly subscription and wants one duplicate charge refunded.”

Names the event, product context, and requested outcome.
WEAK STATE

“A customer has a problem. Think carefully and act as an expert support manager.”

Adds a role, but leaves out the evidence needed to route the request.

03 / ASK ONE DECISION

Make the Question bounded and specific

A good Question has a decision your application recognizes. Ask what you need to choose, verify, or score—not for an essay about the situation.

GOOD QUESTION

“Which team should handle this customer request?”

One decision, paired with Billing, Technical, Sales, and Other.
WEAK QUESTION

“Tell me what you think.”

Combines classification and response writing without a bounded answer.

04 / CHOOSE THE ANSWER TYPE

Boolean, Choice, or Score

BOOLEAN

How likely is this true?

Use for a proposition such as “Was a refund issued?” The result is P(true), shown as Yes and No probabilities.

CHOICE

Which allowed option fits?

Use for routing or classification. You define 2–10 distinct choices; Jev selects one and may return probabilities for them.

SCORE

Where is it on this scale?

Use an ordered rubric. This Playground uses five named levels, from 0 to 4; Jev may return a fractional score between them.

Jev Playground presents yes/no judgments as Boolean. TypeSafe's Direct API calls the native yes/no primitive Noul. The Playground keeps Boolean in its interface; there is no Noul button here. Compare the API question types.

Which type should I use?True or false proposition? BooleanNamed alternatives? ChoiceOrdered severity or quality? Score

05 / READ PROBABILITIES

A lead can be clear or close

Probabilities show Jev’s distribution over the allowed answers. They express its estimated preference among those answers; they are not a guarantee of correctness. Your application decides when to act, ask for review, or gather more evidence based on its own risk, costs, and evaluation results.

Illustrative clear leadBilling is well ahead
Billing92%
Technical5%
Sales2%
Other1%
Illustrative close callReview may be useful
Billing44%
Technical41%
Sales10%
Other5%

Illustrative distributions for reading the interface; they are not saved Jev results or recommended universal thresholds.

06 / TRY YOUR OWN INPUT

Try a controlled comparison

Return to the live Playground, keep Support Routing selected, and replace the State and Question below. Run it again, then compare the returned choice and distribution with your first run.

PASTE THIS STATE

“My account is locked and I cannot sign in.”

Change the Question
Which team should handle this?
Keep the Choices
Billing · Technical · Sales · Other
Back to the live Playground ↑

Changing one fact at a time makes the comparison easier to understand. The exercise does not prescribe which option Jev must select.

NEXT / TRY THE OTHER PRESETS

Run Boolean and Score next

BOOLEAN

Refund Verification

Evaluates how likely it is that the supplied state says a refund was issued.

SCORE

Incident Urgency

Places an incident on an ordered rubric from No urgency to Critical.

Both presets are already available above. Select one, read its State and Question, then run it as-is before editing.

07 / ASK MORE THAN ONE QUESTION

Reuse one State for focused decisions

In the Playground, choose Multiple questions and load Agent Action Gate. The preset sends one shared State with three independent questions: whether important data is at risk, whether confirmation is required, and which next action to choose.

One shared State
BooleanBooleanChoice
Your app combines the answers

One answer is not automatically used as context for another. Your software combines the returned decisions under its own policy. See the browser agent worked example for a complete decision boundary.

08 / JEV DECIDES. SOFTWARE ACTS.

Jev decides; software acts

JEV

Chooses Billing

YOUR SOFTWARE

Assigns the ticket after checking queue rules.

JEV

Returns a high P(true) for refund issued

YOUR SOFTWARE

Shows the status or requests review; it does not issue money.

JEV

Returns CLICK e3

BROWSER CODE

Validates the current page and performs the click.

09 / COMMON MISTAKES

Eight mistakes to avoid

  1. Using Jev like a chatbotAsk for one bounded decision instead of an open-ended reply.
  2. Asking an open-ended questionName the exact choice, claim, or scale you need evaluated.
  3. Using overlapping choicesMake each Choice option lead to a distinct meaning or action.
  4. Leaving important facts out of StatePut the evidence needed for the decision in State.
  5. Treating a 100% display as guaranteed correctnessProbability is Jev’s distribution over the allowed answers, not proof.
  6. Asking Jev to perform the actionKeep permissions, side effects, and enforcement in your software.
  7. Using Score for unordered categoriesUse Choice when the allowed answers have no lowest-to-highest order.
  8. Adding many questions too earlyUnderstand one focused question before combining several.

30-SECOND CHEAT SHEET

Use Jev in seven parts

  1. STATEWhat Jev knows.
  2. QUESTIONWhat Jev must decide.
  3. CHOICEPick one option.
  4. BOOLEANProbability a statement is true.
  5. SCOREPosition on an ordered scale.
  6. PROBABILITYHow strongly the model favors an answer.
  7. APPLICATIONWhat actually performs the action.

BEGINNER FAQ

Questions about using Jev

How do I access Jev?

Try it online in this no-signup Playground, call the Direct API with a TypeSafe account and API key, use an official Python or JavaScript SDK, or give a supported coding agent the official TypeSafe skill.

Can I try Jev without signing up?

Yes. Jev Playground lets you use the interactive tutorial without supplying a TypeSafe account or API key. Usage limits apply. Try Jev now.

Do I need a TypeSafe account to try Jev here?

No. Jev Playground provides its own server-side access for this interactive tutorial.

Do I need a TypeSafe account to use the API?

Yes. Direct TypeSafe API access requires your own TypeSafe account and API key. Open the Jev API guide.

Where do I log in, and is Jev still early access?

Sign in at the TypeSafe Console. The September 15 launch described early access and a waitlist; current docs link to API-key creation, but Direct API availability depends on your TypeSafe account.

Do I need to install Jev?

The hosted Jev model does not require local installation. Raw HTTP needs no SDK; Python uses typesafe-sdk, JavaScript uses @typesafe-ai/sdk, and coding agents can optionally install the official TypeSafe Agent Skill.

Is Jev a chatbot?

No. Jev evaluates supplied State against typed Questions and returns bounded answers with probabilities. Read What Is Jev.

Can I use Jev for free here?

Yes, within this independent Playground’s usage limits. TypeSafe API usage has separate published pricing. See Jev pricing.

What is State?

State is the information Jev reads before deciding, such as a customer message, incident report, record, agent status, or page description.

What is the difference between Choice and Boolean?

Choice selects one supplied option. Boolean is Jev Playground’s yes/no surface and returns P(true); the TypeSafe Direct API calls its native yes/no primitive Noul. Compare the API types.

What does a 100% display mean?

It means the displayed distribution strongly favors that answer after rounding. It is not a guarantee that the answer is correct.

Can I ask several questions about the same State?

Yes. Multiple Questions mode sends focused Questions with one shared State. Jev evaluates each independently; your application decides how to combine the answers. See the browser agent example.

Can Jev take actions for me?

Jev returns decisions. Your software validates permissions and policy, chooses any threshold or review path, and performs the action.

Where do developers find the API?

The Jev API guide covers authentication, request contracts, SDK examples, response handling, and errors. Open the Jev API guide.

READY TO BUILD?

Move the same decision pattern into your application.

The API guide covers the real request contract, server-side key handling, response shapes, and failure states.

Read the Jev API guide