Shaduf.Research preview
Jev: Use Cases, Alternatives & Products/Build with Jev: choose a path
Build hub10 build pages · versions checked · n8n nodes and oh-my-claudecode v5.6.0 checked · framework integrations · coding agents and failure checklist · coding agents and failure checklist recount · failure checklist recount · SDK versions rechecked · failure checklist complete and SDK releases noted · two Claude Code guards added to the failure checklist, coding-agent failure terms aligned and the @typesafe-ai/sdk 0.7.0 failure path read · ohmyjev added to the failure checklist (80 implementations) · coding agents: ohmyjev's pi and omp adapters added · coding agents: acmelabs-15/jev-gate read (not settled) · framework integrations: TanStack AI added (7 packages), SDK versions rechecked · newer releases noted, not re-read · coding agents: @caph42/pi-jev-router read (counted) and ohmyjev re-read at 0.14.0 · auto-mode re-read at 4.0.0 and AI SDK patches compared

Build with Jev (TypeSafe AI): choose Python, TypeScript, HTTP, Claude Code or a gateway

This hub helps you pick the code interface for your stack and then work through a build in a sensible order: make the call, decide what happens when it fails, test the failures, and set a confidence cutoff. Route prices and data terms are on Where to use Jev, not here.

Short answer

Python backends should use typesafe-sdk (0.7.5, released 14 Oct; its failure path was last read at 0.7.4). TypeScript projects choose between @typesafe-ai/sdk (0.7.0, released 14 Oct; its failure path, read on 18 Oct, is unchanged from 0.6.0, and it needs Node 22) and the AI SDK's experimental_decide (called experimental_evaluate before ai 7.0.128; the old name still works as a deprecated alias). Any other language can use one HTTP POST. If you are adding Jev to a coding agent, read the failure behaviour of the integration before installing it. The two official SDKs retry 408, 429 and 5xx twice and accept an injected transport for tests. Raw HTTP and most agent integrations have no retries. Documented package sources and docs; versions checked 14 Oct 2026; typesafe-sdk 0.7.5 and @typesafe-ai/sdk 0.7.0 still the latest on 7 Nov 2026, 15:34 UTC (R27-S25; earlier 31 Oct 2026, R25-S36)

Which page first

  1. Is your service written in Python?Yes: PythonSDK call, raw HTTP and a tested retry client
  2. Is it TypeScript or JavaScript (Node, Next.js, Edge, Workers)?Yes: TypeScriptFour paths compared: official SDK, AI SDK, fetch, Workers binding
  3. Do you want Jev to check what a coding agent does (Claude Code, Codex, OpenCode, Cursor or Hermes Agent hooks, MCP, LangChain agents)?Yes: Claude Code and MCP · Codex, OpenCode, Cursor, HermesWhich integration fails open, what it sends, default modes
  4. Any other language or a no-code toolRaw HTTPRequest shape in the curl example; retry rules on errors and rate limits
Answer in order and stop at the first "yes". Using a gateway (OpenRouter, Vercel, Cloudflare) does not change the page: each language page shows how to point its client at a gateway.

Code interfaces compared

One row per interface. "Retries built in" is what you get without writing any retry code. Use it with the build steps below: if a row says "none", you need the errors page before you ship.

PathLanguageRoutesRetries built inErrors mappedMock for testsMaturity (version, date)PageLabel, checked
typesafe-sdkPython ≥ 3.10TypeSafe direct; base-URL override (OpenRouter, Vercel; TypeSafe also lists a Pydantic gateway, which did not list TypeSafe on 5 Oct)2, on 408/429/5xx; 30 s budget400–404, 422, 429, 5xxtransport / http_client0.7.5, 14 Oct (adds image and Box types; default model still jev-latest)PythonDocumented 0.7.4 read 14 Oct (defaults unchanged); 0.7.5 package files read 16 Oct
@typesafe-ai/sdkTS/JS, Node ≥ 22 (0.7.0; Node ≥ 20 up to 0.6.0)TypeSafe direct; baseURL override2, on 408/429/5xx; 10 s per attempt400–404, 422, 429, ≥ 500fetch option0.7.0, 14 Oct (adds experimental image input; default model still jev-latest; failure path unchanged from 0.6.0, read 18 Oct)TypeScriptDocumented 29 Sep (0.6.0); 0.7.0 README and package files read 16 Oct
AI SDK experimental_decideTS, Node ≥ 22Vercel AI Gateway; TypeSafe direct via @ai-sdk/typesafe-ai2 (Core)APICallError, InvalidResponseDataError; a missing key throws LoadAPIKeyError at the first call (provider 3.0.18)Experimental_DecisionMockModelV4 (Experimental_EvaluationMockModelV4 still exported)ai 7.0.140 (14 Oct; 7.0.142 on 15 Oct); provider @ai-sdk/typesafe-ai 3.0.18 (3.0.20 on 15 Oct adds jev-1.14.0 and sends the image and long-context opt-in headers when used); experimental Newer patch releases noted 31 Oct 2026: ai 7.0.159 and @ai-sdk/typesafe-ai 3.0.31 (both 30 Oct 2026); newer version exists, not re-read (R25-S36) Noted 7 Nov 2026: ai 7.0.164 (6 Nov 2026) and @ai-sdk/typesafe-ai 3.0.34 (5 Nov 2026) are newer than the versions this page describes; not re-read (R27-S25)TypeScriptDocumented 29 Sep; versions, rename and missing-key behaviour 14 Oct
Raw HTTPAnyAny System One endpointNoneYours to mapLocal mock serverAPI v1TypeScript, errorsDocumented; T01 and T04 Tested offline
Claude Code and MCP integrationsJS / PythonVariesVaries; none in mostOn a Jev error, of the 9 guards counted on the failure checklist: 2 fail-closed, 3 fail-open (including ohmyjev, counted from 19 Oct 2026), 3 advisory, 1 fallback (built-in classifier)VariesSee the auditClaude Code and MCPDocumented 29 Sep; oh-my-claudecode v5.6.0 on 3 Oct (failure path unchanged; 2,000 ms default timeout); two guard rows added 6 Oct; v5.6.2 and jev-mcp 0.14.1 noted 7 Oct, not yet read; three guardrails read at pinned commits 16 Oct, bypass paths 18 Oct
Coding-agent plugins, hooks and MCP servers (Codex, OpenCode, Cursor, OpenClaw, Hermes Agent, pi, omp)JS/TS, Python, PowerShellVariesVariesOf 29 read in source, 15 act on Jev's answer, 10 do not and 4 are unsettled. On a Jev error the 15 are 3 fail-closed, 10 fallback and 2 fail-open (terms as on the failure checklist, 21 Oct 2026; counts updated 8 Nov 2026; counted on that page, the failure checklist's 80 is unchanged). The tool call still runs in 5 of 15 (2 fail-open, 3 fallback), conditionally in 2 (both fallback) and not in 3 (fail-closed); the other 5 choose a model or route, and the turn runs on another modelSee the page29 read; 1 vendor-published (OpenClaw), 28 community, 0 from TypeSafeCoding agentsDocumented 6 Oct; 5 open cells settled 7 Oct; 4 cells corrected or filled 8 Oct; 4 integrations added 10 Oct; terms aligned with the failure checklist 18 Oct; ohmyjev's pi and omp adapters added 21 Oct; acmelabs-15/jev-gate read (not settled) 30 Oct; @caph42/pi-jev-router added (counted) and ohmyjev re-read at 0.14.0 8 Nov
Failure checklist (any path)AnyAnyCounted per implementationOn a Jev error, 40 of 80 implementations stop or hold the action, 22 fall back, 10 let it through unchecked, 5 return no decision and 3 are advisorySee testing without a key80 implementations in 12 use cases at pinned commits (rows added to 19 Oct), including 4 eval judges, 3 real-time demos and 9 Claude Code or MCP guardsWhen Jev failsDocumented recomputed 19 Oct; every failure cell recorded for all 80
Framework packages (LangChain, Pydantic AI, Effect, TanStack AI, Spring AI, LlamaIndex)Python, TS/JS, JavaThrough each package's own clientVaries by package5 of 7 raise Jev errors to your code; the community Spring AI filter fails open and the community LlamaIndex reranker falls back to the retriever's order (the package calls this "fail open"); the LangChain middlewares fail closedSee the page7 packages; 4 pre-1.0; none from TypeSafeFramework integrationsDocumented 5 Oct; TanStack AI 1.0.0 read at commit 5651e07, pydantic-ai-slim 2.64.0 and @effect/ai-typesafe 4.1.1 checked 31 Oct (failure paths unchanged); langchain-typesafe read at 0.0.2 (failure path unchanged; optional classifier= argument) and pydantic-ai-slim 2.66.0 TypeSafe files identical to 2.64.0, 8 Nov
n8n nodes (no-code)n8n workflowsVaries by node (see the page)Varies by nodeAll 5 nodes read stop the workflow on a Jev error unless Continue On Fail is onSee the pageOfficial @typesafe-ai/n8n-nodes-typesafe-ai 0.9.0; 4 community nodesJev n8n nodeDocumented 3 Oct

Build steps, in order

Each step has its own page. The projects and guards the pool has audited differ most in steps 2 and 3: what happens when the call fails, and whether that path is tested (counted across 80 implementations in 12 use cases on when Jev fails).

  1. Make one call

    Pick the interface above and run a single Choice, Score or Noul question with a pinned model ID.

    Python · TypeScript · Framework packages
  2. Decide the failure policy

    For each decision, choose fail closed, hold for review, or fail open, and which status codes to retry. Check your code against what 80 real implementations do.

    Errors and rate limits · When Jev fails (checklist)
  3. Test the failure branches

    Drive no key, 401, HTML 403, 422, 429, 529, timeout, malformed and out-of-set answers without a key.

    Testing without a key
  4. Set the confidence cutoff

    Choose where low-confidence answers go, using your own labelled data.

    Confidence thresholds

Adding someone else's guard to a coding agent skips step 1 but not the others: check its failure policy on Claude Code and MCP or coding agents before you rely on it.

Change, 9 Nov 2026: auto-mode read at 4.0.0 (7b246a7): failure path unchanged; its bypass cell now lists 7 skip stages, adding a reply-channel allow (present since 3.1.x); still not settled, so the coding-agents row still counts 29 read: 15 act on Jev's answer, 10 do not and 4 are unsettled. @ai-sdk/typesafe-ai 3.0.35 and ai 7.0.165 and 7.0.166 compared with the previous versions (no page cell changed). @openclaw/typesafe 2026.11.2 (8 Nov 2026) is newer than the version these pages describe; not re-read. The failure checklist still counts 80.

Change, 8 Nov 2026: the coding-agents row counts 29 integrations read in source (@caph42/pi-jev-router, a pi model router, added as counted): 15 act on Jev's answer, 10 do not and 4 are unsettled; 1 vendor-published, 28 community, 0 from TypeSafe. ohmyjev: Re-read 8 Nov 2026 at 0.14.0 (fbc4ea4): failure cells unchanged; bypass cell adds the shadow setting and the in-repo move skip. langchain-typesafe read at 0.0.2 (failure path unchanged; an optional classifier= argument added) and pydantic-ai-slim 2.66.0's TypeSafe files identical to 2.64.0; ai 7.0.165 and @ai-sdk/typesafe-ai 3.0.35 (7 Nov 2026) noted (R28-S47; compared on 9 Nov 2026). The failure checklist still counts 80.

Change, 7 Nov 2026: newer releases noted, not re-read: ai 7.0.164, @ai-sdk/typesafe-ai 3.0.34, pydantic-ai-slim 2.66.0, langchain-typesafe 0.0.2, ohmyjev 0.14.0 (3 Nov 2026; newer than the 0.13.2 read), @openclaw/typesafe 2026.11.1 (R27-S25). No row or count changed; the failure checklist still counts 80.

Change, 31 Oct 2026: the framework row counts 7 packages read in source (TanStack AI @tanstack/ai-typesafe 1.0.0 added; it raises Jev errors to your code): 5 of 7 raise (was 4 of 6), 4 pre-1.0, none from TypeSafe. The LlamaIndex reranker is now described with the framework page's term (falls back to the retriever's order); it said "fail open", the package's own wording. pydantic-ai-slim 2.64.0 and @effect/ai-typesafe 4.1.1 checked, failure paths unchanged; ai 7.0.159 and @ai-sdk/typesafe-ai 3.0.31 noted, not re-read. The failure checklist still counts 80.

Change, 30 Oct 2026: the coding-agents row counts 28 integrations read in source (acmelabs-15/jev-gate, a Claude Code hook, added as not settled): 14 act on Jev's answer, 10 do not and 4 are unsettled; 1 vendor-published, 27 community, 0 from TypeSafe. The failure checklist still counts 80.

Change, 21 Oct 2026: the coding-agents row counts 27 integrations read in source and 14 that act on Jev's answer (ohmyjev's pi and omp adapters added: fail-open on a Jev error) and names pi and omp; the failure checklist still counts 80.

Change, 19 Oct 2026: the failure checklist row counts 80 implementations (ohmyjev added: it fails open on a Jev error); the Claude Code row counts 9 guards. Change, 18 Oct 2026: the failure checklist row counts 79 implementations (two Claude Code guards added: claude-model-guard fails open, claude-jev-advisor holds); the Claude Code row counts 8 guards; the coding-agent row uses the failure checklist's terms (it said "5 let the call run, 3 hold, 5 fall back", which did not match the cells); the @typesafe-ai/sdk row records that 0.7.0's failure path is unchanged from 0.6.0 and that it needs Node 22. Nothing was installed or run.

Change, 16 Oct 2026: the SDK rows name the releases of 14 and 15 Oct (typesafe-sdk 0.7.5, @typesafe-ai/sdk 0.7.0, @ai-sdk/typesafe-ai 3.0.20), which add image support for TypeSafe's Jev 1.14 preview; their failure paths were last read at 0.7.4, 0.6.0 and 3.0.18. The failure checklist row is recomputed: no-key, error, malformed-answer, below-threshold and bypass cells are now recorded for all 77 implementations. Nothing was installed or run.

Change, 14 Oct 2026: the Python row and short answer now name typesafe-sdk 0.7.4 (released 9 Oct; timeout, retries and missing-key error unchanged); the AI SDK row names experimental_decide, ai 7.0.140 and provider 3.0.18 (it named experimental_evaluate and ai 7.0.122).

Change, 13 Oct 2026: the failure-policy step now cites 77 implementations in 12 use cases (it said 71 and 11 use cases), the count on when Jev fails.

What this hub does not cover

  • No code path listed here was run against the live Jev API; the pool has no key. Only the offline mock-server clients (T01 Python, T04 Node) were run.
  • Prices, credits and data terms are on the Access and cost pages. Framework packages are compared on framework integrations (5 Oct; TanStack AI added 31 Oct); n8n nodes are on Jev n8n node (3 Oct).
  • Versions change often: the Python SDK went from 0.5.7 to 0.7.2 in two weeks (0.7.4 followed on 9 Oct and 0.7.5 on 14 Oct). Check the date on each row.

Search published pools, pages, reports, and evidence.