Cline adds Jev decision models for fast, cheap agent guardrails and browsing

A different interface from the chat loop

Most modern agents rely on a familiar loop: send a prompt to an LLM, let it think, and receive a reply through a chat-based API. Cline's post describes a different approach that is emerging, built on APIs for structured decision-making called the Decisions API.

TypeSafe AI has just introduced Jev, a Decisions model aimed at two of the biggest problems with traditional LLMs: latency and cost. Rather than emitting tokens, Jev takes a state, meaning the context in a certain structure, and returns a result for each question asked about it. The result can be a choice, a score, or a noul, which Cline defines as a probability that a statement is true. These values can be used to rerank, prioritise or decide immediately.

Cline says Jev stands out where instant decisions are needed within a structured boundary. One simple example is a before tool hook that lets Jev judge whether a call is dangerous and then block or pass it.

How a Jev request works

A request names the model (jev-latest), a state and a set of questions. Cline's example describes a coding agent that wants to run npm install left-pad-utils@latest --save in a user's repo. The request asks three questions: what kind of action the command takes (a choice), how risky it is to run without asking the user (a score), and whether it downloads and executes remote code (a noul).

According to Cline, the response came back in 318 ms. It classified the command as network_and_system_changes with a confidence of 0.72, gave a risk score of 1.01 with a confidence of 0.58, and returned a noul of 0.82 for running remote code. Cline reads that last value as an 82 percent probability that the command pulls code from the internet.

Three Cline plugins built on Jev

• jev-browser gives Jev a real browser inside Cline. Instead of the usual loop of screenshotting a page and waiting for a model to decide where to click, Jev picks the next action from a list of buttons, links and fields. When Jev stops, Cline checks the final screenshot to confirm the task was done. Installation is through the desktop app under Customize > Marketplace > Plugins > jev-browser, plus a Vercel AI Gateway key saved to ~/.cline/plugins/cline-jev-browser.config.json.
• jev-playground turns a vague prompt into a Jev request through a web interface. Because writing Jev inputs can be hard for people who think in sentences, Cline's LLM sets up the context, questions and buckets. The user can then edit them and press Ask Jev, after which the page talks to Jev directly with no LLM in the loop. Cline's example was a setup for deciding whether a candidate should move to the next interview round.
• jev-guard creates guardrails for a chosen workspace and has tool calls judged before the LLM can act on them. A user describes what to protect, for instance a support inbox agent that people try to push into sending refunds to other bank accounts. Cline turns that into defined guardrails with sample triggers. A Cline hook then screens every tool result in the guarded folder automatically, so the agent does not need to remember to invoke it.

What it means for agent builders

Cline summarises the split this way: the LLM handles language, and Jev handles the quick decisions. According to the company, this makes it practical to run checks across every tool result without adding significant overhead.

Costs can fall further when the LLM side runs on open models. Cline says its browser demo runs on ClinePass, a $9.99/month subscription that gives access to a curated set of open-weight models such as DeepSeek, Kimi, GLM and Qwen, with about 5x the usage of standard API pricing. Users can also bring any model they like to Cline.