jev-acp

# jev-acp

Use [Jev](https://typesafe.ai/) typed decisions from any
[Agent Client Protocol](https://agentclientprotocol.com/) client or IDE.

`jev-acp` is a decision agent, not a chat model. It evaluates explicit
`Choice`, `Score`, and `Noul` questions and returns typed answers,
probabilities, and confidence without involving an intermediate LLM.

## Screenshot

Jev in JetBrains: review a Choice request, then inspect the selected option,
reported confidence, and probability distribution.



*This screenshot illustrates the interface; the example output is not health advice.*

## Setup

Requires Node.js 22 or newer and a TypeSafe API key. You do not need to clone
this repository or build it to use a published package:

```bash
npx -y jev-acp
```

The agent communicates over ACP stdio, not an interactive terminal chat.
Configure your ACP client to launch it and pass `TYPESAFE_API_KEY` as shown below.
Without an explicit version, `npx` uses the stable `latest` channel.
For preview builds, use `npx -y jev-acp@next`. Before the first stable release,
use the preview channel (once bootstrapped) or [local development](#local-development).

Example [vscode-acp](https://github.com/formulahendry/vscode-acp) or
[acp-ui](https://github.com/formulahendry/acp-ui) configuration:

```json
{
  "agents": {
    "Jev": {
      "command": "npx",
      "args": ["-y", "jev-acp"],
      "env": {
        "TYPESAFE_API_KEY": "..."
      }
    }
  }
}
```

Use `"jev-acp@next"` in `args` to track previews, or an exact version to pin a
release. Some Windows clients require `command: "cmd"` and
`args: ["/c", "npx", "-y", "jev-acp"]` to launch `npx`.

## Start with a task

In your ACP client, open the **Task** selector. There is no command syntax to
learn. New sessions start quietly, without a greetin