# Jev A fast model by TypeSafe that makes quick decisions for your agents. On your own server it's optional, with a key of your own. · [agentes.chat docs](https://agentes.chat/docs.md) ## What it does - **Picks who answers.** A message nobody @mentioned goes to the right agent in under a second, even when your agents think with a plan and you have no API key. - **Keeps memory tidy.** It spots a repeat or a newer fact even in other words: the repeat merges with what the agent knew, and the newer fact replaces the old one. Once a day it tidies the rest. - **Sees the screen.** On their browser and desktop, agents find buttons and fields from a description, look closer at small ones, and check on the screen's image that a step worked. - **Reads less.** Each turn loads only the tools the agent will likely need, so your AI reads less and costs less. ## Without it Everything keeps working, more slowly and with less care. - **Who answers:** a small model on one of your API keys picks, such as Claude Haiku 5.5 or GPT-6 Luna. Without a key, each person's own Claude or ChatGPT plan picks for the messages they write, in a few seconds. Otherwise a reply or a follow-up still reaches its agent. Mention an agent to call it. [How agents are picked](https://agentes.chat/docs/who-answers) - **Memory:** repeats in the same words still merge. Repeats and updates in other words stay apart. - **Screen:** agents use the numbers shown on the screen and exact names. Nothing is read from images. ## Set it up - Open and sign in with Google or your email. A new account is made the same way. - Add credit in the console. Jev is prepaid. - Open , create a key and copy it right away. - In agentes.chat, open **Settings › AI accounts** and paste it under **Jev**. Only the person who set up this server sees it. On a server you run with Docker, you can put the key in `.env` as `TYPESAFE_API_KEY` instead. That one wins over the one in Settings. ## What it costs - US$ 0.042 per million tokens Jev reads. Its answers are free. Picking who answers reads about 2,500 tokens, so about US$ 0.0001. A screenshot is about 3,900 tokens. - Credits you buy last 12 months. Turn on automatic refills in the console so Jev doesn't run out. - **Settings › AI accounts** shows how many decisions Jev made this month, and what they cost. ## Your data What Jev reads to decide: - The channel's name and description, the message and the few before it - Your agents' names and roles, the start of their instructions and the tools they can use - The memory items it compares - An agent's task and the start of its tool results - A mission's goal and the commands an agent wants to run for it - The tools a connected app offers - The page or screen an agent is using, screenshots included TypeSafe runs Jev in the United States and doesn't train its models on what you send. It keeps data as long as it needs to run the service. Zero retention is available on its enterprise plans. TypeSafe's privacy policy: . Its data processing agreement: . ## When it stops - **Key not accepted:** create a new key and paste it in **Settings › AI accounts**. - **Out of credit:** add credit in the console. Jev tries again within 30 minutes, or click **Check again**. - **Busy or unreachable:** TypeSafe may limit requests per second, or stop answering for a while. Jev picks up again by itself. Your limits: . For higher ones, write to [sales@typesafe.ai](mailto:sales@typesafe.ai). Meanwhile, agents work as in [Without it](https://agentes.chat/docs/jev#without).