The TypeSafe agent skill
The TypeSafe agent skill is an official drop-in so coding agents stop emitting one HTTP call per question. System One wants many atomic questions in one request to POST https://api.typesafe.ai/v1/systemone. The skill is instructions for tools that call Jev. It is not Jev, not an extra HTTP route, and not a key issuer.
This page is for builders wiring TypeSafe Jev (System One) into agents, routing, or evals. jev.pro is an independent unofficial field guide — not affiliated with TypeSafe AI. We do not sell or issue keys. Confirm install commands on docs.typesafe.ai and the API reference for the underlying POST.
Install (as TypeSafe documents)
Claude Code:
claude plugin marketplace add typesafe-ai/skills
claude plugin install typesafe@typesafe-aiOther agents:
npx skills add typesafe-ai/skills --skill typesafe-aiYou can read SKILL.md at github.com/typesafe-ai/skills. Install the skill in the repo that will call Jev; do not assume jev.pro hosts a copy. The runtime still needs TYPESAFE_API_KEY (Python ≥ 3.10 typesafe-sdk or Node 20+ @typesafe-ai/sdk) or a documented gateway secret.
Why it exists
Agents default to serial, chat-shaped tool use. That wastes the request budget and re-sends state. Pair the skill with speculative fan-out and atomic questions. Choice (max 255), Score (2–10 levels), and Noul (P(yes), no confidence) belong in one questions map. Question ids are not sent to the model; they only key answers.
There is no documented /v1/batch for multi-row files. The skill’s “batch” is many judgments about one state. Multi-row CSV work is your worker.
Production notes
- Pin
jev-1.13.0once the agent’s generated questions feed a live policy.jev-latest/jev-previewcurrently resolve to that id — confirm. - Official SDKs retry 429/529. An agent that retries 422 by rewording blindly can burn input tokens (vendor claim: $0.042 / million input, output free).
- Do not let the skill invent extra first-party URLs (
/v1/explain,/v1/search). Those titles on jev.pro are application patterns. - Schema-safe tool output is not factual correctness. Keep a human hold on irreversible actions.
- Vendor-claimed context (64k state+all questions; 32k state+longest question) still applies when an agent dumps a whole repo into
state. Filter.
When not to use the skill
Skip it if humans write the questions map and the agent only ships glue code. Skip it if you need LangChain’s wrapper — that is a LangChain document, not TypeSafe. The skill will not make System One generate essays or tool-call JSON blobs.
See also
- API hub — the POST the skill is supposed to emit
- Speculative fan-out — many questions, one state
- Atomic questions — one judgment per question
- Python SDK —
TypeSafeClient.system_onefrom generated code
FAQ
Does the skill give the agent a TypeSafe key?
No. The runtime still needs TYPESAFE_API_KEY or a documented gateway secret. This site does not issue keys.
Is this skill Jev itself?
No. It is instructions for tools that call Jev.
LangChain instead?
See LangChain harness — a LangChain document, not TypeSafe.
Will extra questions change other answers?
Questions are independent. Unrelated state can still hurt accuracy on jev-1.13.
What this page does not claim
- Official TypeSafe status, logos, or support.
- That jev.pro sells, issues, or proxies API keys.
- Extra first-party HTTP routes beyond
POST /v1/systemoneandGET /v1/modelsunless the API reference lists them. - Invented latency SLOs, search volume, or unofficial prices.
- That a schema-constrained answer is automatically factually correct.
Disclaimer
This is an independent unofficial site and is not affiliated with TypeSafe AI; official documentation is available at https://docs.typesafe.ai. Never treat jev.pro as TypeSafe official documentation. We do not sell, issue, or proxy API keys.