typesafe-mcp
Give coding assistants structured decisions
Implementation and evidence →JEV GUIDE
This walkthrough uses synthetic feedback about an empty exported file. The example follows the documented interface; it has not been sent to the API and contains no invented model response.
Obtain an API credential from the TypeSafe console and provide TYPESAFE_API_KEY through your environment or server secret store. Keep it out of public HTML, browser scripts and Git.
The example uses curl; a Python SDK path is also documented. Check your account’s access and current billing before execution.
Source: TypeSafe · Quick start ↗
Save this JSON as request.json. The state contains the feedback, reproducible_bug is the response key used by your application, and instructions contains the full judgment to make.
{
"model": "jev-latest",
"state": {
"report": "The preview loads, but clicking Export produces a blank file."
},
"questions": {
"reproducible_bug": {
"type": "noul",
"instructions": "Does `report` describe a specific software action and an unexpected result?"
}
}
}Source: TypeSafe · API reference ↗
Run this command only after configuring the environment variable. It has a 30-second timeout and no automatic retry loop. Executing it contacts the TypeSafe API and may consume paid usage or your account quota.
# TYPESAFE_API_KEY must already be set in your shell.
# Save the JSON example below as request.json.
curl --fail-with-body --max-time 30 \
https://api.typesafe.ai/v1/systemone \
-H "Authorization: Bearer $TYPESAFE_API_KEY" \
-H "Content-Type: application/json" \
--data-binary @request.jsonSource: TypeSafe · API reference ↗
Check that answers contains reproducible_bug, its type is noul, and its value is a number between zero and one. Record the returned model and usage; the latest alias is not a pinned version.
Display the decision in a development view before connecting it to a production action. A valid response can still disagree with your expected label.
Add feature requests, incomplete reports and successful exports as contrasting inputs. Define expected handling and rerun comparisons when instructions or model versions change. Then consider a backend wrapper or one of the projects below.
An experimental image and Box path is now documented. Box requires 1.14 and the image flag; older text examples are not image-localization examples. Follow the version update below for aliases, dependencies and formats. Existing text examples retain their original purpose.
Source: TypeSafe · Box ↗
Download request.json fixture ↓
Give coding assistants structured decisions
Implementation and evidence →Package decisions as versioned services
Implementation and evidence →01 / JEV GUIDE
An October 15, 2026 review of Jev 1.14, preview versus latest, experimental images and Box, long context and SDK migration checks.
Read guide →04 / JEV GUIDE
Write focused Jev questions with original feedback-routing examples, explicit criteria and a practical edge-case checklist.
Read guide →07 / JEV GUIDE
Design Jev Choice classification with distinct labels, unknown inputs and review paths, then explore model-routing and semantic-filtering projects.
Read guide →08 / JEV GUIDE
Place Jev after retrieval: define relevance, preserve provenance, distinguish snippets from full text and evaluate rankings with labeled queries.
Read guide →11 / JEV GUIDE
Understand the TypeSafe AI community node, self-hosting, State and Choice outputs, then design email routing with review and bounded retries.
Read guide →14 / JEV GUIDE
Explore NPC decisions through Jev Lab and HEIST//ONE: local state, finite actions, stale responses, rule-based fallbacks and live-model versus replay checks.
Read guide →