TypeSafe AI (Jev) for Pi: Plugins & Comparison

Overview

TypeSafe AI is the company behind Jev, its first public System One model, announced on September 15, 2026 in early access. Jev returns typed decisions and probabilities for questions defined by software. These community Pi extensions use it as a judging service; your selected coding model still writes the code.

This selection was reviewed on September 22, 2026. Start with the decision you want Jev to make, then check the package owner and data boundary. All five entries are source-reviewed; we did not run paid API evaluations.

TypeSafe AI ↗Introducing Jev ↗

Choose by the decision you need

Choose y0usaf for tool-call risk judgments and output advice. Warden covers broader session supervision. Theo combines inactive-tool discovery, skill recommendations and explicit judgments. Skill-picker focuses on keeping a large skill catalog out of the system prompt. win4r selects the coding model at task boundaries.

Try one integration first and inspect its decisions before enabling more automation. The two pi-jev packages have different owners and both register /jev. Use the owner-qualified names and verified installation commands on the cards.

API access and installation names

The TypeSafe console and documentation are linked below. Each featured integration accepts TYPESAFE_API_KEY. Set it locally before starting Pi; a coding-model subscription does not supply TypeSafe access. Warden also requires explicit consent before its cloud judge runs. win4r requires classifier.enabled to enable Jev requests.

The unscoped npm package pi-jev is Theo’s; y0usaf publishes @y0usaf/pi-jev. Use the Git installation for win4r: npm:pi-jev-router belongs to a different author. Skill-picker also uses its author’s Git installation.

export TYPESAFE_API_KEY='YOUR_TYPESAFE_API_KEY'

y0usaf: observe before enforcing

After installation, /jev reports the key source and mode; /jev last and /jev output show recent judgments. The default gate is shadow. /jev mode enforce enables confirmation for flagged actions in interactive sessions; /jev off disables both judges for the session.

Optional settings belong in ~/.pi/agent/pi-jev.json or project .pi/pi-jev.json, whose values take precedence. Shadow avoids blocking but still calls Jev. The output judge examines a command after it runs, so it cannot prevent execution or undo a leak.

{
  "gate": { "mode": "shadow" }
}

Warden: local checks and online consent

Installation enables local checks. Run /warden enable to review the data notice and provide or reuse a TypeSafe key. TYPESAFE_API_KEY takes priority over the shared pi-typesafe keystore. /warden status shows the active setup; /warden test can send a billable synthetic request when an online judge is available.

The default mode is steer. User settings live in ~/.pi/agent/pi-warden/config.json; project .pi/pi-warden.json can adjust or disable guards but cannot grant cloud consent or change the global backend and mode. Put explicit project rules in pi-warden.md. Review the data inputs before enabling cloud judgment, since rule lookup can include other repository documents.

Theo: tool discovery and skill recommendations

Use /jev status to inspect configuration. Start with explicit jev_find_tools, jev_find_skill or jev_evaluate calls; /jev auto on enables tool and skill matching before user turns, and /jev auto off disables it. Auto mode is off by default.

Tool and skill matching each has its own Jev request when candidates exist. The source can make two requests in one automatic pass, despite the README’s one-request wording. Keep auto-model separate in your cost estimate: that feature classifies locally. Optional orchestration requires pi-subagents; enable guards or compaction only after reviewing their additional data inputs.

Skill-picker: search the catalog when needed

After installing, use /reload or restart Pi. Ask the agent to use skill_search with a concrete task; it loads matching instructions. skill_load accepts known skill names and makes no Jev request. Manual /skill:name commands remain available.

Optional settings go in ~/.pi/agent/skill-jev.json, or the directory selected by PI_CODING_AGENT_DIR. The example keeps the defaults. Search evaluates every enabled skill across parallel batches. Locally loaded instructions still enter the main model’s context, so compare total provider usage rather than treating local loading as zero-token work.

{
  "minScore": 1.4,
  "maxSkills": 3,
  "shardSize": 50
}

win4r: declare models, then inspect routing

After Git installation, restart Pi and run /route init. It creates a starting configuration with the current model; declare your actual model tiers and capabilities before /route reload and /route doctor. The default shadow mode only proposes changes, and the Jev classifier starts disabled.

With TYPESAFE_API_KEY set, /route jev on enables Jev requests. Inspect a proposal with /route preview before choosing /route auto. The first task or /route new can trigger classification; continuation turns keep the model. This router does not create subagents or remove conversation history.

Why these five, as of September 22, 2026

This is an editorial selection based on identifiable Pi packages, distinct uses, recent repository activity and attributable discussion. Cumulative GitHub stars measure attention over time, not search volume or two-day growth. These five are not a ranking of the five hottest projects under one shared metric.

AISuperDomain’s September 20 post links specifically to win4r/pi-jev-router. Theo’s post announces v0.4.0; this page reviews v0.5.0. The popular Ricker and zhouluobo posts discuss Jev more broadly, so their engagement is not credited to a single Pi plugin. Post bookmarks, views and repost counts were not independently verified.

The additional implementations below retain context pruning, focused code checks and another model-routing policy. Their narrower use cases still matter even when they are not in the featured selection.