---
title: "Jev Integration"
description: "Ask Jev, the System One model from TypeSafe AI, typed questions about text, JSON or images and get back yes/no probabilities, classifications and rubric scores your flows can branch on."
url: https://flowrunner.ai/integrations/jev
date_modified: 2026-11-12T10:39:11-08:00
---

# Jev

[AI](https://flowrunner.ai/integrations/category/ai-llms)

Ask Jev, the System One model from TypeSafe AI, typed questions about text, JSON or images and get back yes/no probabilities, classifications and rubric scores your flows can branch on.

[Verified](https://flowrunner.ai/integrations/verified "What does verified mean?") · 6 actions · API key · available

[Jev website](https://typesafe.ai/) · [Platform Documentation](https://docs.typesafe.ai/introduction) · [Platform Documentation](https://docs.typesafe.ai/models) · Capability data verified 2026-11-11

1.  On New Email fires in the Gmail support inbox
2.  Email body and order record passed to Jev
3.  Ask a Yes/No Question: is this a refund request?
4.  Thresholds sort the probability into yes, no, uncertain
5.  Yes answers filed to the billing team's refund queue
6.  Uncertain emails posted to Slack with the probability
7.  Support lead decides refund queue or support reply

## What This Integration Enables

Jev is for the questions a flow asks over and over that have a right answer. Is this a refund request? Which team owns this ticket? How urgent is it? TypeSafe AI built Jev as a System One model, and it does not write text at all. It takes your content as text, JSON or images, answers a typed question about it, and returns a probability instead of a paragraph. There is no prose to parse and no prompt to coax into clean JSON, because the answer already is structured data.

That probability is why Jev fits FlowRunner so well. TypeSafe's own guidance is to give the band between yes and no its own outcome rather than rounding it to one side. In a flow, that outcome is a person. The agent acts on the clear cases and sends the uncertain ones through [human-in-the-loop](https://flowrunner.ai/concepts/human-in-the-loop) to someone who can decide. Ambiguous content becomes one of the flow's [automation exceptions](https://flowrunner.ai/concepts/automation-exceptions) instead of a confident mistake. You set the thresholds, so how much the agent decides alone is a business setting, not a property of the model.

Every question action accepts Text, Data or Image URLs, alone or together, and questions can refer to each by name. A Long Context switch raises the limit for long documents, which TypeSafe notes is slower and less accurate, so flows turn it on only when they need it.

### Without FlowRunner

**Prompted LLM verdicts**: A chat model writes "probably billing" and someone parses the prose into a branch

**Uniform certainty**: Every answer sounds equally sure, so shaky calls go out with solid ones

**Reading every message**: A coordinator opens each email to decide where it belongs

### With FlowRunner

**Typed answers**: Each question returns a choice, a score or a probability the flow branches on directly

**A visible doubt band**: Answers between the thresholds come back marked uncertain and go to a person

**Sorting by exception**: The coordinator reads only the messages Jev could not call

## Flows you can build with Jev

### Refund requests out of a shared support inbox

When On New Email fires in [Gmail](https://flowrunner.ai/integrations/gmail-service), the agent sends the message body as Text and the customer's order record as Data to Ask a Yes/No Question. The question is whether the customer is asking for money back. Yes Means and No Means spell out the edge cases: a question about pricing is not a refund request, a chargeback threat is. A yes goes to the billing team's refund queue. A no continues down the normal support path. An uncertain answer goes to the support lead in [Slack](https://flowrunner.ai/integrations/slack) through Send Message To Channel, with the email and the probability, and the flow waits for a decision.

### Ticket routing that admits when no team fits

When On Ticket Event fires in [Zendesk](https://flowrunner.ai/integrations/zendesk) for a new ticket, the agent runs Classify Content with one option per team. Each option says what that team covers and what it does not. A classification always picks an option, even for a ticket none of them fits, so the agent turns on Check That an Option Fits. A high fit probability lets the agent assign the ticket with Update Ticket. A low one means the ticket is about something no team owns, like a partnership inquiry in the support queue. It goes to the triage owner instead of the closest-sounding team. A Score on a Rubric step with three urgency levels runs alongside, so priority travels with the assignment.

### Lead scoring on a pinned model

Sales ops keeps inbound leads in [Google Sheets](https://flowrunner.ai/integrations/google-sheets). When On New Row fires, the agent sends the row as Data to Score on a Rubric, with four levels written as situations, from "Student or job seeker" to "Budget holder with a named project". It writes the most likely level and the normalized score back with Update Row. The routing cutoffs were tuned against one model version, so the flow pins a versioned id instead of jev-latest, which moves whenever TypeSafe ships a release. Once a month the agent calls List Models. When a newer version appears, it does not switch on its own. It posts the release to the sales ops owner, who decides whether to re-score a sample of recent leads on the new version before the flow moves to it.

## Human-in-Loop Highlight

Jev always answers, and a yes/no answer carries no separate confidence value. The probability is the whole signal, so the thresholds decide what the agent does. In the refund screen from Scenario 1, an email reads: "Third time this charger has failed. I'm done with it." Ask a Yes/No Question returns 0.52, between the 0.3 No Threshold and the 0.7 Yes Threshold, so the answer is uncertain. The customer has not asked for money back, but they may be one reply away from a chargeback. The agent does not round to the nearer side. It asks the support lead in Slack: "Jev puts this at 0.52 for a refund request, inside the uncertain band. The order is from March and there are two earlier tickets about the same charger. Send it to the refund queue, or answer it as support?" The lead makes the call. If uncertain answers keep landing on the same kind of email, moving a threshold is the refund policy owner's decision, never something the agent tunes for itself.

Agent processes routinely

Detects exception requiring judgment

Clear match Continues automatically

Ambiguous Routes to human via Slack

Human decides

Agent resumes with decision

## Agent Capabilities

6 actions

### Questions

4

-   **Ask a Yes/No Question** Asks whether one statement is true of the content and returns the probability of yes, plus a yes, no or uncertain answer set by the Yes Threshold and No Threshold (0.7 and 0.3 by default, as TypeSafe recommends). Yes Means and No Means pin down the edge cases. This is the screening step for refund requests, policy violations and anything else a flow branches on.
-   **Classify Content** Picks the option that fits the content best from 1 to 255 options you name and describe, and returns the choice, its probability, a confidence value and every option ranked. Check That an Option Fits also asks whether any option fits at all, because a classification always picks one. Used to route tickets, emails and records to the right team or queue.
-   **Score on a Rubric** Rates the content on a rubric of 2 to 10 levels written as situations, and returns the probability-weighted score, the most likely level, a 0 to 1 normalized score and the probability of every level. Jev sees only the descriptions, never the numbers. Used to score urgency, lead quality or risk on your own definitions.
-   **Ask Multiple Questions** Sends several typed questions about the same content in one request, written in TypeSafe's API format: yes/no, choice, score and, in preview, box. Each question is answered on its own and comes back under the name you gave it. Used when a flow needs several verdicts on one item, or criteria the single-purpose actions do not offer.

### Images

1

-   **Locate Object in Image (Preview)** Asks where one object is in a PNG or JPEG image and returns its bounding box as fractions of the image and in pixels. A location always comes back, so by default it also asks whether the object is there at all and sets found against a presence threshold. TypeSafe has not yet declared image input production-ready, so flows treat it as a preview.

### Models

1

-   **List Models** Lists the Jev models and aliases on your TypeSafe account, with descriptions and release dates. Used to notice a new release before anyone decides to move a pinned flow onto it.

## Frequently asked questions

### What can FlowRunner do with Jev?

FlowRunner agents can run Ask a Yes/No Question, Classify Content, and Score on a Rubric in Jev, plus 3 more actions.

### Does connecting Jev to FlowRunner require OAuth?

No. Jev connects to FlowRunner with an API key, no OAuth flow required.

### Can Jev trigger a FlowRunner workflow automatically?

Jev doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.

**Work at Jev?** This integration exposes Jev to AI agents on every FlowRunner plan, including through MCP, at no cost to you. [See what FlowRunner offers integration partners](https://flowrunner.ai/integrations/partners), including how to keep this page current.

## Need a custom Jev automation? Our team can build it for you.

Tell us the process. Our implementation team at Midnight Flow builds it around how you work, tests it on your real data, and keeps it running, in your own FlowRunner workspace.

[Have us build it](https://midnightflow.ai/packages/custom?utm_source=flowrunner.ai&utm_medium=referral&utm_campaign=general&utm_content=integrations.aside&utm_term=jev)

---
Markdown version of https://flowrunner.ai/integrations/jev. Site index: https://flowrunner.ai/llms.txt
