Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@jev-kit Run evidence checks on these claims against their sources"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
English | 繁體中文
Batch judgments for coding agents and experimental Ego Lite browser automation, powered by TypeSafe Jev.
Built on jev-use and jev-mcp patterns, with local source/quote validation. Pass existing text or tool output directly; no extra LLM summary is needed.
Tool | Capability | Example use |
| Check up to 256 claim/source pairs per input. Verify exact quotes locally; return support, contradiction, insufficient evidence or review | Check report citations, release claims or answers against logs and documents |
| Classify up to 64 records against 2–32 custom classes, with an explicit manual-review option | Triage issues, group feedback or label tool outputs |
| Use regex to find candidates, then select exact source values for up to 8 fields | Find the current version, date or identifier among several mentions |
| Compare 2–6 supplied alternatives against priorities and up to 3 requirements; flag missing or conflicting evidence | Choose an implementation or processing route using known tradeoffs |
| Reorder up to 30 supplied search results; select an inspection prefix while retaining all remaining IDs | Prioritize code snippets or document passages from existing search results |
These are advisory judgments. Sources and alternatives come from the caller; the tools do not search for missing evidence or execute the selected decision. Input schemas and examples.
For best results, send related records together using fixed classes/questions. Version 0.4.1 keeps at most 8 records per request and overlaps up to 4 requests per invocation, splits long evidence without truncation, and sends identical source strings once per batch. If inputs already exist in local JSON, use the file CLI instead of having a model retype them. For a small set already in the main model's context, direct reasoning was faster in two-host trials. High-confidence errors occurred, so reviewing only flagged items is not a reliable shortcut.
jev_evidence, jev_classify and jev_decide items (evidence items, classify items, decide candidates) accept images: [{"path": "/absolute/file.png"}], up to 4 per item. Classify text / evidence source_text may be omitted when images are present. PNG and JPEG only, detected by file content; each image ≤8 MiB and 280–1568 px per side; ≤20 MiB and ≤16 images per request. Images stay with their item when batching. jev-kit sends the required TypeSafe-Danger-Zone-Images header automatically. Receipts hold only image counts and SHA-256 hashes, never image data. --validate-only checks the files offline. TypeSafe image input is experimental; formats, limits and accuracy may change. Requests use jev-1.14.0 (override with JEV_KIT_MODEL).
Run a browser job with a goal, starting URL and expected outcome. The pinned jev-ultrafast policy chooses actions; Ego Lite executes them.
Click controls, fill fields, select options and navigate pages for searches, forms and article lookup. Generated field text uses a configured text model or Claude CLI helper.
Check target identity, visibility and page state before acting. Re-observe supported stale-target failures instead of replaying the old action.
Verify the expected URL and/or page text independently of Jev's completion signal.
Set allowed origins, a step limit and a time budget; stop for caller handoff on observed popups or dialogs. Record actions and outcomes in a private receipt.
~/.local/share/jev-kit/jev browser --input job.jsonRequires separate Ego Lite/upstream setup. Time limits are checked between operations; login, transaction approvals, frames, shadow DOM and popup continuation are outside the tested scope. Setup and job example.
Native integration, MCP and Skill: expose the same five judgment tools across supported hosts; the included Skill explains when to use them and how to handle uncertain results.
CLI batch processing: read JSON from a file or stdin, return structured JSON, and validate inputs offline with --validate-only.
Shared installation manager: detect hosts, select integrations, update managed hosts together, check configuration drift and remove individual integrations while preserving unrelated settings.
Installation makes tools available for the agent to call. It does not add automatic permission hooks, context compaction or model routing.
Low-confidence, incomplete or invalid judgments remain marked for review. Reranking failures retain the original candidate order. Local quote checks, empty extraction candidates and single-candidate reranking can skip model calls.
Results include review flags, timing, resolved model and usage when a call returns them; source-based tools also retain IDs/references and hashes for comparison with the original input. CLI output can be saved as a new private receipt without overwriting an existing file.
Related MCP server: jev-agent-mcp
Requires Git, Node.js 22+, Python 3.11+ and a POSIX shell. Host installation tested on macOS; Windows unsupported.
curl -fsSL https://raw.githubusercontent.com/WaynezProg/jev-kit/main/install.sh | shThe installer detects configured hosts: Codex, Claude Code, OpenCode, Muse, Grok, Gemini CLI, Cursor, VS Code and Pi. It prefers native integration, with MCP + Skill fallback where supported.
Save your TypeSafe API key to ~/.config/jev-benchmark/typesafe-api-key with file permission 0600, or configure TYPESAFE_API_KEY / TYPESAFE_API_KEY_FILE in the host environment. Restart the host after installation; start a new Codex task.
# Status, update, remove
~/.local/share/jev-kit/jev status
~/.local/share/jev-kit/jev update
~/.local/share/jev-kit/jev uninstallHost configuration and verification details. Cursor/VS Code editor UI discovery remains unverified; OpenCode V2 has contract tests only.
Measured workflow results; browser timings compare Jev against Fable 5.1 low with up to five actions per response. Times are medians.
Workload | Observed result |
Batch scheduling: 32 bilingual records, 3 repeats | Evidence 1.106 → 0.315 s, classification 1.072 → 0.331 s. Paired labels unchanged; larger 32-record requests reduced accuracy. This measures the judgment stage, not whole-agent time |
Shared-source evidence: 8 claims, 3 paired CLI runs | Input tokens 9,855 → 4,633 (53% fewer); both bundles answered 8/8 correctly in every run. This saving depends on repeated sources |
New browser tasks: 3 tasks × 2 runs | Jev 4/6, both Fable policies 6/6. Jev was faster on visible semantic selection (4.17 vs 8.29 s); Fable batch was faster on text entry, and Jev blocked on both dependent-field trials |
Two-host evidence workflow: 18 items × 2 repeats | Direct Codex/Luna and Claude/Fable each 36/36. Jev assistance was slower; reviewing only flags scored 33/36 and 35/36, respectively. This comparison used the frozen pre-optimization runtime |
Ego Lite: 3 form tasks × 2 runs | Both passed 6/6; Jev 6.35 s vs Fable 7.30 s (13.0% less time) |
Ego Lite: 2 Wikipedia tasks × 2 runs, after fix | Both passed 4/4; Jev 5.42 s vs Fable 7.23 s (25.1% less time) |
Code search: 64 queries | Paired target ranked first: BM25 46.9%, Jev 70.3%, Fable 93.8%. In the separate repair test, Jev was slower than BM25: 5.25 s vs 4.42 s |
Model assistance | Haiku accuracy did not improve; Fable routing was faster but less accurate. No demonstrated reduction in required thinking level |
The earlier browser samples below the new trials use familiar tasks; Wikipedia initially passed only 3/4 before the fix. The new dependent-field failures narrow the useful scope to observed controls and limited text generation. These small samples do not establish a general coding speedup, cost saving or reliable reduction in reasoning effort.
Browser measurements · Search and repair measurements · Model comparisons — links pin the tested historical revision.
Local 0.4.1 engineering checks on 2026-09-21: 51 JavaScript + 36 Python tests passed, including concurrent failure isolation, stable ordering, source binding and oversized-input handling. The last published macOS/Linux CI run covers the earlier revision, not these local changes. Engineering checks are separate from the task results above.
This server cannot be deployed
Typed AI decisions for agents: classify, route, screen evidence, score and batch-label text.
Bounded authorization and independently verifiable execution evidence for agents.
Source-backed prompting and coding-agent guidance with provenance and freshness checks.
Pre-execution governance for AI agents. Deterministic PASS/FAIL/REVIEW verdicts, replayable proof.
