Jev Studio

One-stop kit for playing with TypeSafe's Jev: MCP tools for Choice/Noul/Score, ready-made prompt libraries, and slash commands for every cookbook.

Badges

Install

pip install jev-studio

Two commands are installed:

• jev — the CLI for TypeSafe's Jev decisions (verify, screen, classify, extract, match, route, ask, find, rerank, compact, batch).
• jev-studio — the MCP server that serves the Jev ruleset over stdio.

The CLI: jev

Jev turns natural-language state and a set of typed questions into typed answers (yes/no, choice, or score) with probabilities. The CLI wraps it in one-command judgments over text you pipe in.

Authenticate

Any one of these is enough (first found wins):

jev auth login               # store TYPESAFE_API_KEY in the OS keychain (or 0600 file)
export TYPESAFE_API_KEY=...  # from https://console.typesafe.ai/settings/keys
export OPENROUTER_API_KEY=sk-or-...
export CLOUDFLARE_API_TOKEN=... CLOUDFLARE_ACCOUNT_ID=...

Inspect what's live: jev auth status.

Commands

Every command speaks JSON (--json), Markdown (--md), JSONL, CSV, TSV, or a colored text table. Use --pluck to lift one field. Use --fail-on <list> to make the process exit 2 when a judgment condition trips (great for CI gates).

Judgments

• jev verify — check claims against evidence. Verdict + probabilities per claim, with a supporting_evidence back-reference when there is more than one evidence item.
• jev screen — flag prompt injection, empty/boilerplate content, and irrelevance before an agent reads text. Recommends pass / review / block / skip.
• jev classify — one label (single), several (--multi), or a hierarchy (--taxonomy). Confidence-gated auto vs review.
• jev extract — pull typed values out of text. Regex proposes candidates, Jev picks, code normalizes. Builtins: email, phone, url, amount, date, percent, number. Custom regexes via name=/regex/:description.
• jev match — decide if record pairs are the same, different, or unclear. Feed pairs, one list, or two lists (cross product).
• jev route — pick a handler for a request and fill its closed-set arguments in one call.
• jev ask — raw System One passthrough for any state and any questions (repeated --noul, --choice, --score, or --questions-json).

Ranking

• jev find — rank up to 250 candidates against a plain-language query. Returns top-K and a document-level answered / partial / absent verdict.
• jev rerank — score each candidate independently and sort. Several can be relevant, or none.

Pipelines

• jev compact — shrink an agent transcript by dropping stale tool calls verbatim; recent turns are always kept.
• jev batch — run a per-row command (classify, screen, verify) over JSONL or newline-delimited input with a bounded worker pool. Emits JSONL, one line per row.

Account

• jev models — list the models available to your account.
• jev auth — login, logout, status. Keychain-first, file fallback.
• jev config — show, path, keys, set, unset. Defaults < file < env < flags.
• jev update — check PyPI for a newer release.

Examples

# Verify a claim against a file
jev verify "Helmets are optional for adults" --evidence @ordinance.txt

# Screen scraped HTML for prompt injection before letting an agent see it
curl -s https://example.com | jev screen --purpose "extract pricing" --fail-on block,review

# Classify a ticket in three labels, JSON output
jev classify "The invoice total is wrong" --labels billing,bug,feature --json

# Extract an email and a date from an invoice
jev extract @invoice.txt --want email --want date --context "an invoice"

# Rank documentation files for a question
jev find "how do I rotate API keys" --files 'docs/*.md' -k 3

# Batch-classify tickets with concurrency 8
jev batch -i @tickets.txt --concurrency 8 -- classify --labels billing,technical,sales

# Route a support request to a handler
echo "cancel my subscription" | jev route --handlers "cancel:cancel plan,upgrade,help"

# Inspect and edit configuration
jev config set verify.autoAccept 0.9

Exit codes

Input references

Anywhere the CLI takes a value it accepts text, @path/to/file, or - (stdin). Escape a literal leading @ as @@.

Global flags

-m, --model NAME       Jev model, e.g. jev-latest or jev-1.13.0
-P, --provider NAME    auto | typesafe | openrouter | cloudflare
--timeout MS           per-request timeout
--dry-run              print the request that would be sent; don't call the API
--json, --md, --format text|json|jsonl|csv|tsv|md
--pluck PATH           print one field (e.g. label, results[].verdict)
-q, --quiet            omit the usage/model footer in text output
--no-color             disable ANSI colors

Configuration

Config file: $XDG_CONFIG_HOME/jev/config.json (or ~/.config/jev/config.json). Env overrides file, flags override env. See jev config keys for every settable path.

Environment overrides:

JEV_PROVIDER   JEV_MODEL   JEV_TIMEOUT_MS   JEV_FORMAT
JEV_CONFIG     JEV_CREDENTIALS   JEV_CREDENTIAL_STORE   JEV_NO_STORED_CREDENTIALS
JEV_DEBUG=1    # include stack traces on error

Scriptability

Every jev command is safe to run unattended. None of them prompt for confirmation, and none ever will for a non-destructive change.

Read-only — never change stored credentials or config, safe to loop: verify, screen, classify, extract, match, route, ask, find, rerank, compact, batch, models, update (only checks PyPI), version, auth status, config show|path|keys.

Mutating — write local state, still prompt-less:

Side effect on any command: the daily update check caches its result at update-check.json beside the config file. Disable with JEV_NO_UPDATE_CHECK=1 or -q.

In automation, prefer env vars (TYPESAFE_API_KEY, JEV_*) over auth login / config set so runs never touch shared state. Set JEV_NO_STORED_CREDENTIALS=1 to ignore stored keys entirely.

Convention: any future destructive command (e.g. a hypothetical auth purge or config reset) will require --yes to run non-interactively. Non-destructive mutations stay prompt-less.

The MCP server: jev-studio

jev-studio            # speaks MCP over stdio

Point an MCP host at it:

{
  "mcpServers": {
    "jev": { "command": "jev-studio" }
  }
}

Exposes:

• Prompt jev — the Jev ruleset as a user message. Optional mode: lite, full, ultra. Omit for full.
• Tool jev_instructions — same text plus structured output ({mode, instructions}) for hosts that pull context via tools. Read-only.

Develop

pip install -e ".[dev]"
pytest

Ship to PyPI:

python -m build
twine upload dist/*

Claude Code

/plugin marketplace add utk2103/jev-Studio
/plugin install prompt-studio@jev-studio

Two separate prompts. Start a new session; the ruleset lands in system context on SessionStart.

Local clone:

/plugin marketplace add /path/to/jev-Studio
/plugin install prompt-studio@jev-studio

Codex

codex plugin marketplace add utk2103/jev-Studio
codex plugin add prompt-studio@jev-studio

Try something fun

A curated collection of 726 Jev use cases plus 50 runnable evals — every build linked to its project and source post. Search it, copy a brief, hand it to your agent.

Browse: jev-directory.netlify.app

Visual examples in this repo: examples/ — jewellery catalog triage and product auto-crop from photos, using Jev's image input and Box questions.

Credits

The CLI is a Python port of the TypeScript jev-cli, rebuilt on the standard library with the same command surface and question shapes.

Contributing

See CONTRIBUTING.md.

License

MIT License — see LICENSE.