Last updated: September 2026

Jev for social media

Jev is TypeSafe AI first System One model. You hand it a piece of text and a set of typed questions, and it hands back typed answers with probabilities. TypeSafe describes those probabilities as calibrated and puts response time in the hundreds of milliseconds, and we have not verified either. It cannot write a single word, and that constraint is the whole point.

Which makes it very good at one job social media is full of: deciding things about text, at a volume no human and no language model budget can keep up with.

Three layers, and Jev is only one of them

Almost every confused take about Jev comes from expecting it to replace a language model. It does not. A working social pipeline has three distinct jobs, and they want three different tools.

Layer
Tool
What it owns
Judgment
Jev
Is this spam, how strong is this hook, which of these five buckets. Cheap, typed, and fast enough to run on every keystroke on TypeSafe's published figures.
Generation
A language model
Writing the draft, the reply, the rewrite. Jev cannot do any of this.
Action
OpenTweet
Auth, scheduling, rate limits, retries, and delivery to X, Bluesky and LinkedIn. The layer people underestimate until they build it.

The full build is written up in build an AI social media agent.

What a call looks like

One request, several questions, all answered in the same parallel pass. Asking for six things costs the same round trip as asking for one.

judge.ts
import { TypeSafeClient, score, noul } from '@typesafe-ai/sdk';

const client = new TypeSafeClient();

const { answers } = await client.systemOne({
  state: { post: draft, platform: 'X (Twitter)' },
  questions: {
    hook: score('How well does the opening line stop someone mid scroll?', [
      'No hook. Opens with context or a greeting.',
      'Functional. States the subject plainly.',
      'Strong. Creates tension or surprise in the first line.',
    ]),
    bait: noul('This post explicitly asks the reader for likes, reposts or follows.'),
  },
});

answers.hook.score;         // probability-weighted float across the rubric
answers.hook.confidence;    // how certain the model is
answers.bait.noul;          // a float 0 to 1. Nouls have no confidence field.

The detail most write-ups get wrong

A Noul returns only its float. There is no confidence field on a Noul, because the float already is the probability. Choice and Score do return confidence. The primitives reference has the exact response shape for all three.

This runs in production

OpenTweet free post scorer sends every draft you type to Jev as one request carrying seven questions: four Score rubrics for reach, hook, clarity and specificity, and three Nouls for AI-slop, engagement bait and toxicity. It runs one request per draft. TypeSafe publishes Jev's end-to-end response time, and we do not publish our own timings. No account needed, and you can watch the confidence band move as you edit.

Try the scorer, freeRead how it is built

Start here

Is Jev an LLM?
No, and the difference decides what you can build with it.
Choice, Score and Noul
The three primitives, their exact response shapes, and when each is right.
Can Jev hallucinate?
It always returns a valid answer. That is not the same as a correct one.
Limits and rate caps
TypeSafe's published context, request and token caps, and what they mean at volume.
Getting API access
The waitlist, plus three routes that skip it.
Cost calculator
Price your own workload and see what binds it, requests or tokens.

What to use it for

Ten jobs on social media that are decisions about text.

Score a post before publishing
A quality gate that runs before the thing goes out, not analytics after.
Classify social replies
One task solved three ways, so you learn the whole API.
Content moderation
Category design, thresholds, review queues, and the cost at scale.
Rage bait detection
Defining it precisely enough to be scoreable is the hard part.
Spam and bot detection
Jev judges the text. Account signals stay in your code.
Sentiment analysis
Sentiment is an ordered rubric, which makes it a Score and not a Choice.
Detect AI-written posts
The production rubric, and where it is unfair to real people.
Rank post variants
Generate several, score them all, publish the one that wins.
Brand crisis detection
Alert on the spike against a rolling baseline, not on one bad mention.
Community management
Triage mentions, skip the noise, surface what needs a person.

Build it

Wiring judgment to something that acts.

Build an AI social media agent
The full architecture. Jev judges, an LLM writes, OpenTweet ships.
Jev and MCP
Judgment layer plus action layer, wired to one agent.
Jev in n8n
A two-node workflow: classify, branch, publish.
Jev in Claude Code
Using Jev from an agentic coding tool, then publishing the result.
Setting a confidence threshold
Pick the bar from how reversible the action is, not from a gut number.
Classify replies, end to end
Pull, judge, route, respond.
Route DMs with AI
Sales, support, spam, partnership. Sorted on arrival.
Building a post scorer on Jev
Seven questions on one draft, and how the rubrics are written.

Compare it

Honest comparisons, including the cases where Jev is the wrong answer.

Jev alternatives
Seven ways to make the same decision, priced and timed.
Jev vs GPT
The benchmark table, with the caveats the press left out.
Jev vs Claude Haiku
Cheap LLM versus no LLM at all.
Jev vs Gemini Flash
Multimodal matters more than price if your feed has images.
Jev vs GPT-5 Nano
The closest competitor on price. The arithmetic, shown.
Jev vs structured outputs
Answering the "this is just JSON mode" objection properly.
Jev vs Instructor and Guardrails
Retrying until it parses versus being unable to emit anything else.
Classifying a million posts
At volume the rate limit binds before the bill does.

When not to reach for Jev

  • You need words back. Drafting, rewriting, summarising, replying. Jev emits no text at all.
  • Your content is images or video. Jev is text only, which rules out a lot of real social moderation.
  • You make a handful of decisions a day. At low volume the integration effort costs more than the inference it saves.
  • You need a defensible audit trail about a person. A probability is a filter, not evidence.

Frequently asked questions

What is Jev?

Jev is TypeSafe AI first System One model, released on 15 September 2026. You send it a state and a map of typed questions, and it returns typed answers with probabilities that TypeSafe describes as calibrated, in one parallel pass. It never generates text, so it is not a chatbot and not a replacement for a language model.

Is Jev an LLM?

No. A language model decodes text one token at a time. Jev answers a fixed schema of questions in parallel and returns numbers and labels only. That is why its output tokens are free, and TypeSafe puts its end-to-end response time in hundreds of milliseconds rather than seconds.

What is Jev good for on social media?

Judgment at volume. Scoring a draft before you publish it, classifying inbound replies and mentions, filtering a feed, flagging rage bait or spam, and ranking variants of a post against each other. All of these are decisions about text, which is exactly what Jev does.

Can Jev write my posts?

No. Jev cannot emit text at all. Writing is a job for a language model. A working pipeline uses both: a language model drafts, Jev judges, and something else publishes.

How much does Jev cost?

TypeSafe's published price is $0.042 per million input tokens, with free output and no free tier. On that pricing, a short social post scored on a handful of questions costs a few millionths of a dollar.

Is Jev used in a real product?

Yes. The free Will It Go Viral post scorer is built on Jev, where seven Jev questions grade every draft as you type. It is not used for generation anywhere, because Jev cannot generate.

Jev decides. OpenTweet publishes.

One API and an MCP server for X, Bluesky and LinkedIn, so the judgment your agent makes turns into something that actually posts. From $11.99 a month.

7-day free trial. Cancel anytime.