Jev AI

Jev for coding agents · Claude Code · Codex · OpenCode · Cursor

Jev for Coding Agents

Give your agent a model that answers yes, no, which one and how much, with a probability, instead of another paragraph.

Coding agents make hundreds of small decisions: which file matters, whether a diff follows the rules, which tool to call next, whether the task is really done. Jev answers exactly that kind of question in about half a second, with a calibrated probability the agent can branch on. This page is the setup: one prompt that wires the Jev AI API into your project, the jevgrep skill for repository search, and four patterns to start from.

Copy the setup prompt Get a Jev API key jevgrep CLI

1 · Paste one prompt into your agent

It carries the connection details itself, so the agent does not need to browse: base URL, endpoint, header, default model, and the rule that the key stays server-side. The agent reads the docs, configures the client and makes one small test call on your balance.

Integrate Jev AI into this project. Inspect the existing code and read https://jev-ai.pro/docs first, especially #official-sdk for setup and #errors for failure handling.

Connection details: base URL https://jev-ai.pro/api; decision endpoint POST https://jev-ai.pro/api/v1/systemone; header "Authorization: Bearer $JEV_AI_API_KEY" with a key created at https://jev-ai.pro/jev-api; default model jev-latest.

If using TypeSafe's official SDK (@typesafe-ai/sdk), reuse the existing client and explicitly set baseURL to https://jev-ai.pro/api with a Jev AI key. Changing only the key is not enough. TypeSafe publishes the SDK; Jev AI provides the compatible endpoint.

Keep JEV_AI_API_KEY server-side; never put it in browser code, logs, commits or chat. Tell me where to configure it locally and in deployment.

Follow the docs for request/response formats, retries and model limits. Start with jev-latest, verify the actual destination, and check GET https://jev-ai.pro/api/v1/models without inference. Then show me how to make one small decision call using my balance.
  • Claude CodePaste the prompt as a message in the project, or save it under .claude/ as a skill. The jevgrep skill installs into the same folder.
  • CodexPaste the prompt into the Codex session, or add it to AGENTS.md so every session starts with the connection details.
  • OpenCodePaste the prompt, or install the jevgrep skill, which OpenCode reads from the project.
  • Cursor and othersAny agent that can read a project file and run curl works: the prompt tells it the base URL, the header and the model.

2 · Let the agent search the repository with Jev

jevgrep answers “where is X handled?” by having Jev score files and excerpts for relevance, then hands the agent the evidence. Its built-in skill teaches Claude Code, Codex and OpenCode when to use it.

  1. 01

    Install with the Jev AI connection

    One command on the jevgrep page installs the CLI and points it at your Jev AI key. Node.js 22+, macOS or Linux.

  2. 02

    Run jg skill

    Select your agent and the skill is written into the project. Use jg skill --global for every project.

  3. 03

    Ask a question, not a symbol name

    jg "Where is authentication checked before a request reaches a handler?" . returns files, excerpts and line references the agent can act on.

The upstream project reports a 28.6% reduction in coding-agent cost on ten tuned SWE-bench tasks, with Jev inference excluded. Measure it on your own tasks.

3 · Four decisions agents call Jev for

Each one is a ready judge in the playground: open it, edit the criteria, then export the request.

Check a diff against project rules

Send each CLAUDE.md or AGENTS.md rule as a choice question over the diff: complies, violates or insufficient evidence. The agent gets a per-rule verdict instead of a paragraph.

Open the judge

Pick the next tool

Give Jev the task and the tool list; it returns one choice with a probability for every tool, including no tool at all and the ones that need human confirmation.

Open the judge

Verify a completion claim

Before an agent reports done, Jev checks the claim against tool results, task requirements and allowed actions. Jev as a judge for agent evals, in one call.

Open the judge

Guard what the agent reads

Web pages, files and tool output can carry instructions. A yes/no question over the text flags overrides and exfiltration attempts before the agent acts on them.

Open the judge

Jev decides, the LLM writes

The pairing that works in practice.

Jev does not generate code or prose, and that is the point. Claude, GPT or Codex write the change; Jev tells the agent, with a probability, whether the change follows the rules, which file to open, which tool to call, and whether the claim of completion holds up. Because the answer is a number, the agent can set thresholds: act above 0.9, ask a person below 0.6.

The pattern is the same one the LLM-as-a-judge and LLM router use cases use outside agents: a cheap, fast, calibrated decision in front of an expensive generative step.

With TypeSafe’s official SDK

If the agent prefers the SDK to raw HTTP, this is the client setup it should produce.

// Official TypeSafe SDK, published by TypeSafe — @typesafe-ai/sdk 0.6.0
// Server-side JavaScript / TypeScript using the Jev AI compatible endpoint
import { TypeSafeClient, choice } from '@typesafe-ai/sdk'

const apiKey = process.env.JEV_AI_API_KEY
if (!apiKey) throw new Error('Set JEV_AI_API_KEY on your server')

const client = new TypeSafeClient({
  apiKey,
  baseURL: 'https://jev-ai.pro/api', // Required, including for existing SDK users
  retry: { maxRetries: 0 }, // Avoid replaying a request with an uncertain outcome
})

// The SDK appends /v1/systemone to baseURL.
const response = await client.systemOne({
  model: 'jev-latest',
  state: { document: 'I was charged twice. Please fix this ASAP.' },
  questions: {
    category: choice('What is this ticket about?', { billing: null, technical: null, other: null }),
  },
})
console.log(response.answers.category.choice)

Setting baseURL is required; changing only the key sends the request to TypeSafe with a key it does not recognise. The documentation covers retries, errors and model limits.

Coding agents FAQ

How do I use Jev with Claude Code?

Copy the setup prompt on this page into a Claude Code session in your project. It tells the agent the Jev AI base URL, the decision endpoint, the Authorization header and the default model, and asks it to read the documentation, keep the key server-side and make one small test call. For repository search, install jevgrep and run jg skill to add its Claude Code skill.

Does the same setup work for Codex and OpenCode?

Yes. The prompt is plain text with the connection details inside it, so any coding agent can follow it. The jevgrep skill installer supports Claude Code, Codex, OpenCode and other agents, and installs the skill into the project you run it in.

Is there a Jev MCP server?

Jev AI does not publish an MCP server. Agents call Jev over HTTP or through TypeSafe’s SDK, which is what the setup prompt configures. The MCP tool router use case shows how to use Jev itself to choose among MCP tools.

Does Jev replace Claude, GPT or Codex in an agent?

No. Jev does not write code or text; it answers typed questions with calibrated probabilities. The useful split is Jev for decisions the agent branches on, such as routing, rule checks and verification, and the generative model for the writing.

Where does the API key go?

In the environment of the process that calls Jev, as JEV_AI_API_KEY, never in browser code, logs, commits or chat. The prompt asks the agent to tell you where to configure it locally and in deployment. Create and revoke keys on the Jev API page.

What does an agent’s call cost?

The same as any API call: metered on input tokens, with credits as the fallback, from the balance on your Jev AI account. jevgrep searches can make several calls per question, so start with a small folder.

Can the agent use TypeSafe’s official SDK?

Yes. @typesafe-ai/sdk works against Jev AI when its baseURL is set to the Jev AI API and a Jev AI key is supplied; changing only the key is not enough. The example on this page shows the exact client setup.

Claude Code, Codex, OpenCode and Cursor are products of their respective owners and are not affiliated with Jev AI. jevgrep is an open-source project by its author; Jev AI distributes an adapted installer.