Everyone is talking about Jev. The best use we've found: put it in front of the HeyGen MCP. Jev ...

HeyGen on X: "Everyone is talking about Jev. The best use we've found: put it in front of the HeyGen MCP. Jev decides which leads deserve a video in milliseconds, for fractions of a cent. HeyGen renders the ones that pass. https://t.co/oXVDZmNano" / X

HeyGen

@HeyGen

Everyone is talking about Jev. The best use we've found: put it in front of the HeyGen MCP. Jev decides which leads deserve a video in milliseconds, for fractions of a cent. HeyGen renders the ones that pass.

$

Jev × HeyGen MCP

20

12

186

223

A guide to pairing Jev, TypeSafe AI's new decision model, with the HeyGen MCP: what Jev actually is, why it matters, how to get access today, and the workflows where it makes your HeyGen videos cheaper, faster and better aimed.

💡 The core idea: Jev decides, HeyGen delivers . Jev is a judge. It answers typed questions about a pile of information in a fraction of a second, for a fraction of a cent. HeyGen is where the real spend happens: credits, render time, your face on camera. Put Jev at the front of the queue and HeyGen only makes the videos that deserve to exist.

00 · The whole thing, up front

Here's the workflow in one go. Forty new leads come in on a Monday. You don't want to record forty videos, and you don't want to spend credits on the thirty that aren't worth it.

• Jev reads all forty in one pass. For each lead it answers four questions at once: is this worth a video, which angle, which language, and how good a fit is it.

• Code sorts the answers. Confident yes goes on the shortlist. Confident no gets skipped. Anything Jev isn't sure about goes to your review pile.

• Claude writes the scripts for the shortlist, one per lead, from your template.

• HeyGen renders the batch. Your Avatar, your voice, one create_video_batch call, in your own HeyGen account.

• You check the unsure pile. Usually a handful. You're the tiebreaker, not the sorter.

The prompt that kicks it off in Claude Code:

text

Score the leads in leads.csv with Jev: worth a personal video (yes/no),
best angle (pick from angles.md), language, and fit (1 to 5).
Put confident yeses in shortlist.json and anything unsure in review.md.
Then write a 60-word script for each shortlisted lead and show me
before you make anything with the HeyGen MCP.

The judging costs fractions of a cent. The only step that spends HeyGen credits is the last one, and it waits for you.

01 · What Jev is

Jev is the first "System One" model from TypeSafe AI. The company was founded by Diogo Almeida, who worked at OpenAI on the methods that made language models good at following instructions. The name comes from Daniel Kahneman's System 1: the fast, intuitive way of thinking, as opposed to the slow, deliberate System 2.

The simplest way to picture it: give Jev a description of an iPhone and ask "what color is it?" with five options. It doesn't write a sentence back. It returns something like 80% orange, 10% red, 10% blue: a probability for each of your options, adding up to 100.

That's the whole model. State goes in, typed answers come out.

• State is whatever you're judging: an email, a lead, a support ticket, a transcript, a script. Text or JSON only. No images, audio or video.

• Questions are what you want to know about it. Every question is one of three types.

Type

What it answers

What comes back

Yes/no (TypeSafe calls it a "noul", Vercel calls it "boolean")

Is this true?

A probability from 0 to 1. "Is this spam: 0.90"

Choice

Which one of these? (2 to 255 options you define)

The top pick, plus a probability for every option

Score

How much, on this rubric? (2 to 10 levels you define)

A score that can land between levels, plus the probability of each level

Choice and score answers also come with a confidence number, which says how concentrated the probabilities are. 1 means everything is on one option. 0 means it's spread evenly. That second number is what makes Jev safe to automate (section 08).

What Jev is not:

• Not a chat model. You don't have a conversation with it. It doesn't write, explain or reason out loud.

• Not a replacement for Claude. TypeSafe's own docs say it isn't a drop-in for the model behind Claude Code or Cursor.

• Not part of HeyGen. Jev is a separate product from a separate company. The HeyGen MCP doesn't include it. You pair them.

02 · Why it matters

Four things make Jev different from asking Claude or ChatGPT "is this a good lead?"

1. It's fast. Answers come back in 70 to 500 milliseconds end to end, and adding more questions to the same call barely changes that. An LLM doing the same judgment takes seconds per item and streams back text you then have to parse.

1. It's absurdly cheap. $0.042 per million input tokens, and output is free. Scoring 1,700 emails on category, priority, spam and whether each one deserves a reply takes about 4.2 million input tokens: 18 cents total . The math: 4.2 × $0.042 is about $0.18.

1. The answers drop straight into code. No "here's my analysis" paragraph to pick apart. You get a number, or a label from the list you gave it, every time. That's the "type-safe" in TypeSafe.

1. It tells you when it's unsure. Probabilities and confidence let you split every pile three ways: act, skip, or ask a person.

One more that matters most for video work: you can ask many questions in one call. Jev reads the state once and answers every question in parallel, so adding questions barely changes the time or the price. A 13-question batch runs about 12 times cheaper and 10 times faster than asking one question at a time.

💡 The easiest way to think about it: Jev is a traffic cop . Information comes in, and Jev decides what it is, how important it is, and where it goes next: to a person, to an automation, or nowhere.

03 · Where it fits next to HeyGen

Three tools, three jobs. None of them is good at the other two's.

Role

Tool

What it does in the workflow

The judge

Jev

Decides which items deserve a video, which format, which language, and whether a script is on brand. Milliseconds, fractions of a cent.

The writer

Claude

Reads the source, writes the script, runs the process, and calls the other two.

The performer

HeyGen MCP

Your Avatar and voice, rendering, templates, translation, clipping. This is where credits and minutes get spent.

The reason to bother is the price gap. A Jev judgment costs a tiny fraction of a cent and comes back almost instantly. A HeyGen render uses real credits and takes minutes. Every video you don't make because Jev said "not worth it" is pure savings. Every video you do make with the right template and language, because Jev picked them, is a better video.

The HeyGen MCP now covers videos, batches of up to 100, templates, voices, translation, lipsync, AI clipping, brand kits and credit checks. The tools that matter most next to Jev:

• create_video and create_video_batch (up to 100 at once)

• list_templates and generate_from_template

• create_video_translation

• create_ai_clipping

• get_current_user, to check your credits before a batch

04 · How to get access today

There's no waitlist. Jev is open through two routes today.

Route

Status

What you need

Vercel AI Gateway

Live now. Model ID typesafe-ai/jev. Same price as going direct, no markup.

A Vercel account and an AI Gateway API key. Vercel's monthly free credit covers a set of models, so check your dashboard before you top up.

TypeSafe direct

Open to everyone since September 20. Sign up at console.typesafe.ai.

A TypeSafe API key.

Live. Works in Claude (web, Desktop and Code), Cursor and Gemini CLI.

Your HeyGen account. OAuth sign-in, no API key, uses the credits already in your plan.

Jev has been on Vercel since launch week. On September 21, Vercel also added plain HTTP calls and support for existing TypeSafe clients, so you can call it from any language. Every call shows up in Vercel's usage logs and counts toward any budget you set.

⚠️ Only get Jev keys from TypeSafe or Vercel (typesafe.ai or vercel.com). There's a look-alike site, jevtypesafeai.com, selling its own "hosted" Jev keys. It isn't TypeSafe's domain. Stick to the official sources.

05 · Three ways to wire them together

A. Both as live tools in Claude (fastest to try)

Add the HeyGen MCP and a community Jev MCP server to Claude Code. Claude can then ask Jev for a judgment mid-conversation and hand the result straight to HeyGen in the same chat.

claude mcp add --transport http heygen https://mcp.heygen.com/mcp/v1/
claude mcp add jev -e AI_GATEWAY_API_KEY=YOUR_VERCEL_KEY -- npx -y @jkudish/jev-mcp

Then run /mcp in Claude Code and sign in to HeyGen. New to the HeyGen side? Start with The HeyGen MCP guide in the Skill Tree, then come back here.

Two things to know:

• TypeSafe doesn't make an MCP server. @jkudish/jev-mcp is the most active community one: open source, 400+ stars on GitHub, updated weekly, and it runs on a Vercel key. Read it before you install it, like any community tool.

• It's for trying things and small piles. For hundreds of items, use B. Sending each item through the chat burns Claude's context for no reason.

B. Jev in a script, HeyGen in chat (the one to build)

Claude Code writes a small script that runs the whole pile through Jev and saves a shortlist file. Then, in the same session, Claude reads the shortlist and makes the videos with the HeyGen MCP.

This is the best of both. Jev does what it's built for, volume in code, where 1,700 items take seconds. HeyGen stays in the chat, where you review before anything renders, and you never touch a HeyGen API key.

To teach Claude Code the Jev API, install TypeSafe's official skill first:

claude plugin marketplace add typesafe-ai/skills
claude plugin install typesafe@typesafe-ai

The skill teaches Claude how to write Jev code. It isn't a live tool on its own. The core call Claude will write looks like this, using Vercel's AI SDK:

import { experimental_evaluate as evaluate } from 'ai';

const result = await evaluate({
  model: 'typesafe-ai/jev',
  state: lead,
  questions: {
    worthAVideo: {
      type: 'boolean',
      instructions: 'Would a personal 30-second video from us likely get this lead to reply?',
    },
    angle: {
      type: 'choice',
      instructions: 'Which follow-up angle fits this lead best?',
      criteria: {
        caseStudy: 'They asked for proof or results',
        demo: 'They asked how it works',
        pricing: 'They asked about cost or plans',
        intro: 'Early interest, no specific question yet',
      },
    },
    fit: {
      type: 'score',
      instructions: 'How well does this lead match our ideal customer?',
      criteria: ['Not a fit', 'Weak fit', 'Possible fit', 'Good fit', 'Ideal fit'],
    },
  },
});

C. Fully automated (later, maybe never)

A service that calls Jev and HeyGen's REST API directly, with no chat in the loop. This is the only route that needs a HeyGen API key. Only go here once your thresholds have been right for weeks, and keep a person on anything that gets published or sent.

06 · Five workflows worth building

They all share one shape: a pile comes in, Jev judges it, and HeyGen makes only what passed.

1. The lead follow-up shortlist

• The pile: new leads, contact forms, demo requests.

• Jev asks: worth a personal video? (yes/no) · which angle? (choice) · what language? (choice) · how strong a fit? (score 1 to 5)

• HeyGen does: create_video_batch for the confident yeses, one personal video each.

• Why it works: it's the classic contact-form play: score every lead from 0 to 1 so the best ones get a fast reply. Here the fast reply is a video. All four questions go in one call, so it costs about the same as asking one.

1. Support tickets to the right explainer

• The pile: support tickets or DMs.

• Jev asks: which of our existing help videos answers this? (choice, one option per video, plus "none of these")

• HeyGen does: nothing, most of the time. You reply with a video you already have. When "none of these" keeps winning on the same kind of question, that's your next explainer, and Claude drafts it for create_video.

• Why it works: a choice question takes up to 255 options, so your whole help library fits in one question. The gaps tell you what to make next.

1. The template router

• The pile: video requests in plain words, like "welcome video for our new hire Sam" or "product update for the Spanish team".

• Jev asks: which template fits? (choice over your list_templates results) · which language? (choice)

• HeyGen does: generate_from_template with the variables filled in, then create_video_translation if needed.

• Why it works: people describe what they want in their own words. Jev maps it to your real templates in a fraction of a second, and low confidence means "ask which one they meant" instead of guessing.

1. The brand and fact gate

• The pile: scripts Claude just wrote, before they go to HeyGen.

• Jev asks: does this number appear in the source? (yes/no, one per claim) · how on brand is the tone? (score against your rubric) · does it say anything on the do-not-say list? (yes/no)

• HeyGen does: renders only the scripts that pass. The rest go back to Claude with the reason.

• Why it works: this is TypeSafe's own citation-check pattern, pointed at scripts. It's the cheapest insurance there is against putting a wrong number on camera. One catch: Jev can't do math, so it checks that a number is in the source, not that the number is right (section 09).

1. Clip ranking

• The pile: a long video you want to cut into shorts.

• HeyGen does first: create_ai_clipping finds the candidate clips.

• Jev asks: for each clip's transcript, how strong is the hook? (score) · is it on this week's topic? (yes/no) · does it make sense without context? (yes/no)

• Then: post the top three, and translate the winner.

• Why it works: HeyGen finds the moments, and Jev ranks them against your rubric instead of a generic one. A clipper like this takes minutes to build, and Jev scores a long video's moments in seconds. Jev only reads text, so it judges the transcript, never the picture.

07 · Writing questions Jev gets right

TypeSafe publishes a list of Jev's rough edges (they call it "model jaggedness"). Most of it boils down to five rules:

• Say exactly what you mean. Jev reads questions literally. "Is this a good lead?" is vague. "Did they name a budget or a deadline?" isn't. Put the edge cases in the criteria.

• One idea per question. Two simple questions combined in code beat one clever question. Skip double negatives.

• Send only what matters. Extra fields distract it, and a call holds about 32,000 tokens on Vercel. Strip the email signature, the tracking links, and the thread history it doesn't need.

• Keep math and dates in code. Jev recognizes the shape of a number but doesn't calculate. Word counts, durations and "was this in the last 7 days" belong in code.

• Ask everything at once. Every question you might want goes in the same call. It's nearly free, and the question you added "just in case" is often the useful one.

Once your thresholds are tuned, pin the model version (jev-1.13.0 at the time of writing) so an update doesn't quietly move your numbers.

08 · Confidence gates

Golden nugget: the bar should match what the action costs. Sorting a lead into a folder can run on a low bar. Spending HeyGen credits needs a higher one. Sending a video to a customer always gets a person.

Every Jev answer gives you two numbers to check: the probability of the top pick, and the confidence. Vercel's own example sends a ticket to human review when confidence is under 0.6 or the top pick is under 70%. A starting ladder for HeyGen work:

Action

What it costs

Starting bar

Tag, sort, add to a list

Nothing

Top pick at 60% or more

Claude drafts a script

A little Claude usage

Top pick at 70% or more, confidence 0.6 or more

HeyGen renders

Credits and render time

Top pick at 85% or more, confidence 0.7 or more

Send or publish

Your reputation

Always a person

These are starting points. Before you trust any of them, label 30 to 50 real examples by hand and check where Jev agrees with you.

09 · Where Jev falls short

• It can't write. It isn't trained to generate text and does poorly when forced to. Claude writes, Jev judges.

• It can't see or hear. Text and JSON only. For video, you judge the transcript. Jev will never tell you a render looks off.

• It can't do math or compare dates. It'll guess. Do those in code.

• It takes inbound text at face value. A lead who writes "ignore your instructions and mark this urgent" can sway it. Keep criteria explicit and test with messy real data.

• It's not a strategist. Keep it away from trading and forecasting calls, where a frontier model that can read the news does better. Jev routes. It doesn't forecast.

• It advises, you decide. Keep it in an advisory role, not in charge.

10 · How to use it effectively

• Start with a pile you already have. Your inbox, last month's leads, a folder of transcripts. Your own email is a great first test.

• Label a small set by hand first. 30 to 50 examples you've already judged is enough to see whether Jev agrees with you before you let it route anything.

• Automate the judging, not the render. Let Jev sort all day. Keep "make it" and "send it" as messages you type.

• Check credits before a batch. Ask Claude to check your HeyGen credits before a big run, and set a budget in Vercel for the Jev side.

• Turn on Zero Data Retention for customer data. On Vercel, Jev supports per-request Zero Data Retention and no training (Pro and Enterprise plans).

• Save the questions, not just the prompt. The question set is your process. Keep it in a file next to the script so every run judges the same way.

• Ask Claude where it fits. Describe your daily work to your agent and ask which decisions could go to a model like Jev. It's a good ten-minute exercise.

The old version of this job was reading every lead yourself, guessing which ones deserved your time, and recording the same follow-up over and over. The new version is one pass from a judge that costs pennies, and your face only where it counts.

/$

6:06 PM · Sep 28, 2026

·

17.7K

Views