User repository on GitHub
TypeSafe decisions for Pi: batched evaluation tool, terminal playground, and typed API for extension authors
- Forks
- 13
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Repos & SDKs
Pi TypeSafe — DevMortimer
A Pi extension and library that gives the agent and other extensions one consented, key-managed Jev client, with a batched typesafe_evaluate tool for classify, triage, compare, and score judgments.
- Stars
- 48
- Language
- TypeScript
- License
- MIT
- Reading time
- 3 min
- Published by
- Community
- Author
- DevMortimer
- Last verified
- Sep 2026
About this project
Jev inside Pi. Jev is TypeSafe's judgment model: send it some state and typed questions and it returns probabilities instead of prose, in well under a second, for a fraction of a cent. This package gives Pi three things built on it: A tool for the agent. typesafeevaluate hands small structured judgments (classify, triage, compare, score) to Jev and returns calibrated numbers in one batched call.
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
Highlights
- typesafe_evaluate takes JSON state plus 1 to 32 questions and returns answers, model, token usage, and elapsed time.
- The tool is registered at startup but disabled until /typesafe enable; automated or headless runs set PI_TYPESAFE_ENABLED=1.
- Per request: 32 questions and 64 KiB of JSON; per session: 20 attempts, a 15-second timeout, and no automatic retries.
- Daily counters live in ~/.pi/agent/pi-typesafe/usage.json and roll over at local midnight; a cap error names the cap, amount used, and day.
- Independent project, not affiliated with TypeSafe AI or the Pi authors.
Quickstart
json
{
"state": { "message": "I was charged twice. Please help today." },
"questions": {
"team": { "type": "choice", "instructions": "Which team should handle this?",
"criteria": { "billing": "Charges and payments", "technical": "Software failures", "other": "None of these" } },
"urgent": { "type": "noul", "instructions": "Does the sender request help today?" }
}
}Watch out
MIT-licensed. Requires Pi 0.85 or newer and Node.js 22.19 or newer; usage is billed to your TypeSafe account, an enabled extension with no usable key looks exactly like a working one, and the SDK's TYPESAFE_BASE_URL and TYPESAFE_LOG_LEVEL overrides are ignored.
Questions about this entry
Short answers drawn from the linked project or page.
Three caps, all per local day: requests, input tokens and estimated dollars. The environment may lower an explicit cap but never raise it, and a reached cap raises a budget error naming the cap, the amount used and the day. Counters persist in the extension's usage file and roll over at local midnight, so a restart does not reset the spend.
Not until you opt in. The evaluate tool is registered at startup but stays disabled until you run the enable command in the session, which comes with a data notice to confirm. For automated or headless runs you set the environment flag explicitly. A status command reports key state, session and today's counters, and which cap is currently reached.
Only the state and questions you or the agent submit, sent to the TypeSafe API and nowhere else. No files, no conversation history, no telemetry. Error messages never include upstream response bodies, headers, keys or your submitted state. Playground and test output stays in the terminal and never enters the model's context.
The stated reasons are speed, cost and independence: the main model is slower, costs more per call and grades its own work, while this returns a calibrated number from a separate model in a fraction of a second. The README argues that matters most when the same question is asked many times, across every tool call, file or issue in a list.
Per request, 32 questions and 64 KiB of JSON. Per session, 20 attempts, a 15-second timeout, and no automatic retries. Per day there is no cap unless you set one. Session limits reset when a session starts or reloads, and daily counters live in ~/.pi/agent/pi-typesafe/usage.json.
/typesafe test sends one built-in sample request and /typesafe playground edits request JSON in Pi's editor; both show results in the terminal only and do not enter the model's context. Neither requires the agent tool to be enabled.
Yes. It exposes a typed API with one client, one key store, and one login prompt so extensions such as pi-warden do not each ask for a key. createTypeSafe takes maxRequests and maxUsdPerDay, and the library has no dependency on Pi, so it is safe in tests.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Tool calling as classify plus action, back to twelve-factor agents
X: A chat bot with no LLM
Other listings of this project
Forks and copies of this project in the directory. This entry is the canonical one.
DevMortimer's Pi extension hands batched Choice, Score, and Noul judgments to Jev: a typesafe_evaluate agent tool, a terminal playground, and a typed API other extensions reuse, with daily request, token, and dollar caps.
Jev Agents & MCP#community#pi#typescript
Community
A Pi extension that runs four hazard checks before bash, write and edit and four diff-quality checks afterwards, sharing one System One request per phase and applying its own thresholds.
Jev Agents & MCP#community#jevbest#catalog-import
Community
twilwa/pi-typesafeTypeScript
twilwa/pi-typesafeTypeScriptMore Jev resources like this
A filesystem-first framework for durable AI agents whose auto model router defaults to Jev through Vercel AI Gateway and whose evaluate helper asks typed Choice, Score, and Boolean questions inside tools.
Jev Repos & SDKs#community#typescript#agents
Community
vercel/eveTypeScript
vercel/eveTypeScriptA Juspay TypeScript SDK that treats Jev as a third inference type beside generate and stream: providers declare which kinds they serve, and tryDecide returns typed boolean, choice, or score answers, null when no decision provider is set.
Jev Repos & SDKs#community#typescript#sdk
Community
juspay/neurolinkTypeScript
juspay/neurolinkTypeScriptAn ACP and MCP adapter that bridges TypeSafe Jev with any LLM: native Jev when TYPESAFE_API_KEY is set, or Codex, Claude, Grok, OpenCode, and OpenAI-compatible endpoints standing in as System One when it is not.
Jev Repos & SDKs#community#typescript#mcp
Community
tacticocc/JevbridgeTypeScript
tacticocc/JevbridgeTypeScriptThe AI Toolkit for TypeScript: Vercel's free, open-source AI SDK for building AI-powered applications and agents, with GitHub topics covering OpenAI, Anthropic, Gemini, React, Svelte, Vue, and Next.js.
Jev Repos & SDKs#community#catalog-import#logicrw
Communityvercel/aiTypeScript
vercel/aiTypeScriptA small TypeScript client for asking typed questions about your data: one ask call mixes yes/no, choice, and score questions and returns probabilities, selected options, and confidences.
Jev Repos & SDKs#community#typescript#sdk
Community
pithings/advocaatTypeScript
pithings/advocaatTypeScriptA DSPy fork whose @typesafeify decorator routes decision-shaped signature fields to Jev, so bool, Literal, and scored outputs come back typed while freeform fields still use the LM.
Jev Repos & SDKs#community#python#sdk
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.