Jev in Practice: Browser Use and Agent Workflows

A roundup of 12 open-source projects that use Jev for single-choice or scoring decisions, from browser control and context compaction to trading decisions, with a minimal path to get started.

A three-person team built a DM triage tool with Jev for a creator receiving over 4,000 DMs a day, surfacing only the messages that need her judgment, and reached the finals.

Persian tech outlet Digiato introduces Jev from TypeSafe AI, a model that outputs decision probabilities instead of text and is pitched as faster, cheaper and more controllable for automation.

An inbox triage setup splits sorting into five small questions answered by Jev while Grok Bot handles browser and terminal actions, cutting a $930 annual sorting bill to $34.

The author uses Jev to route support tickets to departments, has Python send low-confidence tickets to human review, and tracks routing accuracy alongside the share flagged for review.

The author experiments with Jev and WebMCP, where Jev selects the next tool and fills enumerable inputs, an LLM handles only free-text or numeric fields, and Jev's confidence score drives extra checks in the harness.

The author says the community has already built many things with Jev plus browser-use and computer-use, and that he will release his own version soon.

Using Jev to ask six typed questions of 3,412 social media candidates and pass only those worth a full read, at 15.7 seconds and $0.41 instead of 6 hours of manual reading.

An Ori Eval run of 30 classification tasks on 200 samples found Jev at 154ms median latency, 98.5% accuracy and $0.049 per 1,000 decisions, which the author uses to argue routing and classification should not call a full LLM every time.

The team wired Jev into production as a gatekeeper that decides within 500ms which agent takes a job, without generating any text.

The author walks through creating a Jev API key, setting TYPESAFE_API_KEY, installing the official Skill, then calling Jev in conversation so Claude Code or Codex executes while Jev makes decisions, claiming gains in speed and cost.

On the same synthetic cases, Jev routed with 100% accuracy at 323ms average latency versus 83.3% and 1821ms for GPT-5.6 Luna, prompting the author to build and open-source agent-tool-router.