If you're hunting for a free Jev API key, here's the answer straight away: the free route doesn't need a key at all.
OpenCode Zen lists a model called jev-1.13-free, and you can call it with a plain POST request and no Authorization header.
I re-checked that on 5 October 2026, and it came back with an answer, a probability and a cost of "0".
So this guide skips the hunt and gets you to a working Jev call, then shows you how to wire it into an agent properly.
I'll also show you exactly where a real key starts to matter, because the free door is labelled limited time.
What Jev is, in one sentence
Jev is a decision model from TypeSafe AI that reads a situation and answers typed questions with a value and a probability, instead of writing text.
It launched on 15 September 2026, and TypeSafe calls it the first of its "System One" models.
You hand it two things.
The first thing is the state, which is the email, ticket, log line or record you want judged.
The second thing is a set of questions with the only answers you'll accept.
Jev hands back one answer per question, plus the odds, and it never writes you a paragraph.
That's why it's so useful inside an agent, because most of what an agent does all day is make small decisions rather than write essays.
Is this email worth a reply?
Is this task finished?
Should this go to a human?
Those are exactly the questions Jev was built for.
Who makes Jev and what a key normally costs
TypeSafe AI builds Jev.
The official route is TypeSafe's own API, and their quick start tells you to grab a key from the dashboard at console.typesafe.ai and send it as a Bearer token.
TypeSafe's homepage lists Jev at $42 per billion input tokens.
OpenCode Zen's pricing table lists the paid jev-1.13 model at $0.042 per million input tokens, with output marked free.
Vercel's model page for typesafe-ai/jev now lists $0.04 per million input tokens, with a 32K context window.
So even the paid route is very cheap.
But if you're reading a post about a free Jev API key, you want to test it before you hand over a card, and that's completely sensible.
The problem a free Jev route solves for operators
Here's the old way most of us run agent decisions.
Every tiny judgement goes on a full trip to a big chat model.
The model thinks out loud, writes a sentence, and then your code has to dig the actual answer out of that sentence.
That costs seconds per decision and you're never quite sure the parsing worked.
Jev cuts that whole step down to one small call that returns a typed answer your code can act on directly.
The catch is that you want to prove it works on your own data before you commit to a new provider.
That's the gap the keyless free route fills.
How to set up the free Jev route (no key, no signup)
This is the part you came for.
You only need a terminal and curl.
Step 1: Send your first keyless request
Copy this into your terminal exactly as it is.
curl https://opencode.ai/zen/v1/systemone \
-H "Content-Type: application/json" \
-d '{
"model": "jev-1.13-free",
"state": "Support note: the agent issued a full refund to the customer today.",
"questions": {
"refund_issued": {
"type": "noul",
"instructions": "Was a refund issued to the customer?"
}
}
}'
Notice there's no Authorization header anywhere in that request.
When I ran this on 5 October 2026, here's what came back.
{"model":"jev-1.13-free","answers":{"refund_issued":{"type":"noul","noul":0.98}},"usage":{"input_tokens":288,"output_tokens":23},"cost":"0"}
The noul value of 0.98 is the probability that the answer is yes.
The cost field reads "0", which is the proof you're on the free model.
The whole round trip took about 1.3 seconds from my desk, network lag included.
Step 2: Don't drop the "-free" suffix
This is the mistake I'd expect most people to make.
If you change the model name to plain jev-1.13 and leave the key out, Zen sends back HTTP 401.
The error message says rate-limited Zen models require a workspace.
So the free route lives entirely in that one model name.
OpenCode's own docs say it plainly: use jev-1.13-free instead of jev-1.13 to use the limited-time free model.
Step 3: Ask several questions in one call
Jev isn't limited to one yes-or-no question per request.
TypeSafe's docs describe three question types, and you can mix them in a single call.
| Question type | What it answers | What you get back |
|---|---|---|
| Noul | A yes-or-no question | The probability that the answer is yes |
| Choice | Pick one option from a set you define | The chosen option, a probability per option and a confidence |
| Score | Rate something against ordered levels | A score, a probability per level and a confidence |
TypeSafe's own cookbook on parallel questions reports that batching thirteen questions into one call was 12.2 times cheaper and 10 times faster with no change in answers.
So the operator habit is simple: pack your questions into one request rather than firing them one by one.
Here's a Choice and a Score in the same call.
curl https://opencode.ai/zen/v1/systemone \
-H "Content-Type: application/json" \
-d '{
"model": "jev-1.13-free",
"state": "I was charged twice this month and nobody has replied to my last two emails.",
"questions": {
"team": {
"type": "choice",
"instructions": "Which team should handle this?",
"criteria": {
"billing": "Payment or subscription issues",
"technical": "Bugs or integration problems",
"sales": "Pricing or account questions",
"other": "Anything that fits none of the above"
}
},
"frustration": {
"type": "score",
"instructions": "How frustrated is the customer?",
"criteria": ["Calm", "Frustrated", "Very angry"]
}
}
}'
I always include an "other" option in a Choice question, and I'll show you why in the gotchas section.
🔥 Want the agent wiring I use for this? Inside the AI Profit Boardroom, I've got Jev and decision-layer walkthroughs that show how to plug a model like this into a real agent stack. Plus weekly coaching calls and 3,400+ members building real automations. → Get access here
How to use Jev inside an agent
Getting a number back is the easy bit.
The operator skill is deciding what your agent does with that number.
Gate actions on confidence, not just the answer
TypeSafe returns two different numbers, and they mean different things.
The probability tells you which answer is favoured.
The confidence tells you how sure the model is about the whole judgement.
TypeSafe's confidence docs suggest three lanes: act automatically when it's high, proceed with caution in the middle, and route to a human when it's low.
I set the line based on the cost of being wrong.
A misfiled email is cheap to fix, so I set a low bar and let it run.
An automatic refund moves real money and is hard to undo, so I set a high bar and make a person sign off.
Trim the state before you send it
Jev explains nothing when it gets something wrong.
You get a wrong answer and a number, and that's it.
So you debug the input rather than a chain of reasoning.
In my own testing for the free Jev guide on AgentOS.guide, I asked "is this task finished?" against three versions of the same job.
A thin line of log scored 0.22, the failed import scored 0.02, and the clean import scored 0.79.
The lesson is to cut the state down in your code first, so Jev only sees the part that matters.
Let Jev decide and another model write
Jev can't write the reply to a customer.
The pattern I'd use is a split.
Jev picks the queue, the urgency or the next step.
A writing model like Claude then does the words, only when words are actually needed.
That keeps the expensive model out of every small decision your agent makes.
If you're already running Hermes as your agent, my guide on how to set up Hermes Agent in one click gets the base running first.
Every free Jev route, side by side
The keyless Zen model is my pick for operators, but it isn't the only door.
| Route | Needs a key? | What I verified | Best for |
|---|---|---|---|
OpenCode Zen jev-1.13-free |
No | A keyless call returned a cost of "0" on 5 October 2026 | Testing Jev in scripts and agents today |
| jevplayground.com | No | The page says "Free to try · Usage limits apply" | Seeing Jev work in a browser with no code |
| Vercel AI Gateway | Yes, a Vercel key | Promo pricing ended 25 September 2026, and Jev now lists at $0.04 per million input tokens | Teams already on Vercel using the $5 monthly free credit |
| TypeSafe API | Yes, from console.typesafe.ai | The quick start uses a Bearer key and the jev-latest model |
Production use straight from the maker |
| OpenRouter | Yes, an OpenRouter key | Jev 1.13 was in beta there when I built my use-case demos | Operators who already route models through OpenRouter |
The playground is an independent test bench, not TypeSafe's own site, so treat it as a demo rather than something to build on.
What the Vercel free credit actually gives you
Vercel's pricing docs say every team gets a free tier with $5 of credit included each month.
That credit only covers models Vercel marks as free-tier eligible, and the free tier has lower rate limits and no bring-your-own-key option.
When I set it up in September, the free-tier key was allowed 30 requests a minute and 250,000 tokens a minute, and Jev answered on the free credit.
Vercel also blocked the request with a verification error until I added a card, even though I bought nothing.
Vercel's own docs warn that once you purchase credits, your account moves to the paid tier.
The promo price ended on 25 September 2026, so check the model page for current eligibility before you rely on it.
Gotchas and limits before you build on it
These are the things I'd want to know before I trusted a free Jev API key route in an agent.
- The free Zen model is labelled limited time, so build your code so the model name is one config value you can swap.
- Free models on Zen can come with different data terms, so read the Zen data policy before you send anything sensitive.
- "Zero hallucinations" from TypeSafe is a promise about the shape of the answer, meaning Jev always answers inside your options, not a promise that it's right.
- Jev reads text only, so if you want it to judge an image you have to describe the image in words first.
- TypeSafe publishes a "jaggedness" page listing known weak spots in jev-1.13, and it's worth reading before you go live.
The shape point catches people out.
In my testing for the AgentOS guide, I asked Jev to file "Can I bring my dog to the workshop?" with no "other" option on the list.
It put it under sales at 98%, with a confidence of 0.96.
The answer was inside the box and still wrong, which is why I always add an "other" option now.
Real examples I ran on the free model
These are runs from my own testing on jev-1.13-free, published in the free Jev guide.
- A 48-email sample inbox went through the free model, and 45 of the 48 landed in the right folder in 8.4 seconds.
- Five of those emails came back under 0.60 sure, and two of the three misses sat under that line, which is exactly what a confidence gate catches.
- Twenty-four contact-form messages were graded hot or cold, and all twenty-four landed on the right side.
That's the kind of job an agent does hundreds of times a day.
If you'd like a second agent in the loop doing the heavy work, my walkthrough on the Hermes MCP server for Codex shows how two agents share a project.
🚀 Want this running inside a full agent system? The AI Profit Boardroom is where I show the Agent OS, with a decision layer like Jev underneath that routes leads, grades enquiries and picks what each agent does next. You get the walkthroughs, the coaching calls and 3,400+ members to compare setups with. → Join the Boardroom here
Who the free Jev route is for
It's for operators who want to test a decision layer this afternoon without opening a new account.
It's for anyone running agents who's tired of paying a big chat model to answer yes or no.
It isn't for anything you'd let run unsupervised with money attached, at least not until you've set confidence gates and swapped to a keyed route.
My advice is to prove it on your own data on the free model, then move to a keyed route the day it matters.
Related reading
📺 Video notes + links to the tools 👉
🎥 Learn how I make these videos 👉
🆓 Get a FREE AI Course + Community + 1,000 AI Agents 👉
Also On Our Network
- 🌐 How to turn a free Jev API key into a lead-sorting service
- 🌐 My honest verdict on every free Jev route
- 🌐 Using free Jev as the decision layer in an Agent OS
- 🌐 The full visual guide to the three free Jev doors
About Julian
I'm Julian Goldie, an AI entrepreneur, SEO expert, and founder of the AI Profit Boardroom, which has 3,400+ members.
I help business owners scale with AI agents, automation, and SEO.
- I have 400,000+ YouTube subscribers who watch my AI tool tests every week.
- I built a 7-figure agency from the ground up.
- I run daily AI training inside the Boardroom.
- I wrote two Amazon best-sellers on SEO and agency growth.
→ Get my best AI training inside the AI Profit Boardroom
Free Jev API Key FAQ
Is there a free Jev API key?
You don't need one for the free route.
OpenCode Zen lists jev-1.13-free at https://opencode.ai/zen/v1/systemone, and a keyless POST request returned an answer with a cost of "0" when I re-checked it on 5 October 2026.
OpenCode labels it a limited-time free model.
What is the Jev free tier?
There isn't one single Jev free tier.
The free options are OpenCode Zen's jev-1.13-free model, the jevplayground.com test bench, which says free to try with usage limits, and Vercel AI Gateway's $5 monthly free credit for free-tier eligible models.
What happens if I call the paid jev-1.13 model without a key?
It fails.
When I sent a keyless request to jev-1.13 on OpenCode Zen, it returned HTTP 401 with a message saying rate-limited Zen models require a workspace.
Only the jev-1.13-free model name works without a key.
Where do I get an official Jev API key?
TypeSafe's quick start says to get your API key from the dashboard at console.typesafe.ai and send it as a Bearer token to https://api.typesafe.ai/v1/systemone.
Jev is also available through Vercel AI Gateway, OpenRouter and the paid jev-1.13 model on OpenCode Zen.
Can Jev write text for my agent?
No, because Jev is a decision model rather than a writing model.
It answers typed questions with values and probabilities, so the common pattern is to let Jev pick the route or next step and let a writing model produce any words.
The bottom line
The fastest way to get a working Jev call today is to skip the sign-up entirely and use the jev-1.13-free model name on OpenCode Zen.
Set confidence gates, trim your state, keep the model name in one config value, and you'll be ready to switch when the free window closes.
That's everything you need to know about getting a free Jev API key working inside your agents.
