Choice
Select one label from a list: tool, model, route, browser action, or escalation path.
route: "human_review"Where Jev fits in your agent.
A practical guide to bounded decisions, typed outputs, and the code that turns a judgment into an action.
WHAT JEV ACTUALLY IS
A fuzzy if statement, not another talking agent
Jev is a decision model accessed over TypeSafe's network API. Your program sends text or structured state and defines the answers Jev is allowed to return. Jev responds with a choice, a score, or a yes/no probability instead of free-form prose.
That narrow job matters. Many agent loops repeatedly ask small questions: which tool should run, whether a browser task is complete, how risky a patch looks, or which queue should receive a ticket. A general language model can answer those questions, but it also brings open-ended generation, slower calls, and output parsing to a job that may only need a branch.
01Reality or agent statePage, trace, ticket, task, candidates
↓02Jev decidesChoice · Score · Noul probability
↓03Code applies policyThresholds, permissions, review, budgets
↓04Agent or tool actsCodex, Claude, Playwright, API, human
THREE OUTPUT SHAPES
Bound the answer before inference
Typed output does not make every answer correct. It keeps the answer inside the structure your code expects, while probabilities give policy a place to handle uncertainty.
Score
Rate urgency, quality, risk, relevance, or completion on an ordered scale.
risk: 3.2 / 4Noul
Return a yes/no probability so code can allow, block, or ask a human.
goal_met: 0.93Use Jev when the branch is known.
- Choose one option from a predefined set.
- Score or classify many similar items.
- Judge whether an agent run is complete.
- Select the next allowed action inside a loop.
- Send uncertainty to a review path.
Use something else when the answer must be invented.
- Write prose, code, emails, or search queries.
- Reason through an open-ended problem.
- Act when candidate actions cannot be defined safely.
- Interpret images, audio, or scanned PDFs directly.
- Automate without a low-confidence fallback.
Fast and cheap are claims to test.
TypeSafe publishes a 70–500 ms response range and a launch price of $0.042 per million input tokens, with output described as free. Those are vendor figures. Real agent latency also includes state construction, network distance, tools, retries, and the rest of the loop.
70–500 msVendor-published response range$0.042 / MTokVendor-published input price at launch
FAQ
Jev without the shorthand
Is Jev an AI agent or a replacement for Codex and Claude Code?
No. Jev is a hosted decision model. It can choose, score, or return a yes/no probability from bounded options. Codex, Claude Code, browser tools, and ordinary programs still generate content and perform actions.
Where does Jev run?
Jev runs on TypeSafe's servers and is called over an API. The Python SDK and official Agent Skill help your local code or coding agent build that integration; they do not install the Jev model on your computer.
What is the fastest way to try Jev?
Use the TypeSafe Playground without installing anything. Give it a state, add a Choice, Score, or Noul question, and inspect the returned answer and probabilities before deciding whether the task belongs in an agent workflow.
Are the price and latency numbers independently tested?
No. The 70–500 ms range and $0.042 per million input tokens are TypeSafe's launch figures. Community project metrics are labeled as reported by their authors. JevForAgents does not present either as a first-hand test.
SCENARIO DEMO
Resume facts → application fields
The supplied clip shows a resume beside an application form. A safe Jev workflow would first extract candidate facts, ask Jev which candidate best matches each field, and let browser code fill only the confirmed values. Sensitive or missing answers stay blank, and the user controls final submission.
Jev’s proposed role is semantic matching. Reading the PDF, pasting values, answering sensitive questions, and submitting the application remain outside that role. The original publication URL is still needed to verify that this recording actually uses Jev.
Explore video demos →