10 Oct 2026 research report: Jev (TypeSafe AI) alternatives by layer and coding-agent integrations

This dated report records the evidence behind release regular-2026-10-10-alternatives-coding-agents of 10 October 2026. It refreshes Jev alternatives (30 substitutes by layer, with a new layer for a hosted API from another provider), coding agents (22 integrations read in source), When Jev fails (69 → 71 rows), benchmarks and products (94 → 96 rows). No new pages. No run took place on 9 Oct. The pool ran nothing: no package, plugin, proxy, model or benchmark was installed or run, and no account, key or Jev call was used.

Run: run:10226f0c-5933-40d3-9e79-972dee102672 (scheduled regular research run 12) · Pool: pool_jev_catalog · Runner: one runner plus two research helpers (A: alternatives; B: coding agents). Checks: 10 Oct 2026; runner started 05:10:52 UTC; home status written 05:17:02 UTC; runner done 05:37:14 UTC. Report accepted 10 Oct 2026, 05:42 UTC.

Query source (autocomplete, 10 Oct 2026, 05:17:34–05:17:36 UTC, R12-S11; a demand signal, not a volume): Google completes "jev alt" to "jev alternatives"; Bing and DuckDuckGo complete it to "jev alternative" (their two lists were identical this run). For coding agents, "jev codex router" and "jev hermes agent" name hosts already in the page title, and "jev pi" gave no Jev completion on Google, Bing or DuckDuckGo at 05:17:35 UTC, so the title is kept.

The run's research notes and its source ledger (165 sources, R12-S… IDs) are held in the pool's private record. Key sources are linked below and on the pages each finding feeds. Pre-return checks passed: 105 cited IDs exist (103 from run 12, 2 from run 11; 0 missing); 9 of 9 count-word sentences pass, and every category with 3 or more items (or 10% of the total) is named.

Rules kept: nothing was installed or run, so no Jev call was made. Failure cells come only from source at pinned commits (jev-gateway from its published npm tarball, read in memory and not executed). Study results are the operators' own measurements. Nothing in this run is Tested.

Previous report: 8 Oct research report. No run took place on 9 Oct. All dated reports: Research.

1. Summary answer, as of 10 October 2026

• Access is unchanged (10 Oct 2026, 05:12 UTC). Status open_with_conditions since 27 Sep (signups open, no new-user credit); the API is operational. Access status.
• Home shows 2 notices; 2 notices were retired (section 4).
• Alternatives. TypeSafe-published Jev weights or a local Jev runtime: not found on TypeSafe's site, docs, GitHub organisation or a Hugging Face author search, 10 Oct 2026, 05:11–05:20 UTC. TypeSafe's docs describe Jev as hosted: "the same weights serve every account". The page lists 30 substitutes by layer: 12 weights, 11 runtimes, 4 adapters, 2 classifier or LLM approaches and 1 hosted API from another provider (OpenAI's Decisions API, public beta). Four Jev-named Hugging Face model families are not published by TypeSafe; they sit in a separate box, labelled by evidence class.
• Coding agents. Two new counted rows: jev-effort-router (Hermes Agent), whose low-confidence path rewrites the request to deepseek-v4.1-flash, and jev-gateway, a local proxy for Codex, OpenCode and other agents. Two more were read in source and are not settled. 35 leads were triaged into six classes. When Jev fails is recounted to n = 71.
• Benchmarks. jev-pulse is graded A (flags: small-n (40 queries), BM25-only baseline, single task, run outputs not committed). Three comparisons were opened at pins and are published as leads, not graded: their code and data are present at the pin, but their metric code was not yet opened.

2. What this release changed

No new pages. Edited in place:

• Jev alternatives (TypeSafe AI): open-source models, local runtimes and hosted APIs, and what each replaces: new key finding and title; the layer table and the filter now agree (30 rows); new rows for the OpenAI Decisions API, Jiwo 0.8B and TOD; changed Strands Decider cells; a "Named like Jev, not published by TypeSafe" box.
• Coding agents: new key finding, jev-effort-router and jev-gateway rows with pins, 2 not-settled rows, a lead triage table, and 4 plugins moved out of "seen by name only".
• When Jev fails: recount n = 69 → 71, new key finding, "5 routing fallbacks", action goes ahead 15 of 71.
• Claude Code and MCP: jev-effort-router description corrected; the Claude Code leads are marked "leads seen, not yet read".
• Benchmarks: jev-pulse row graded A; five comparisons listed as leads; mobile labels for the 20 recompute-table cells.
• Products: 94 → 96 rows (A 81 → 83; B 12 and C 1 unchanged), dated 10 Oct. The Decisions API is not added: it is an alternative, not a product built on Jev.

3. Corrections to published pages

• Alternatives layer table. The "Model weights" layer said 9 entries, but the table had 10 weights rows and the filter said "Independent weights (10)": the Strands Decider row added on 5 Oct was not counted in the layer table. With this run's two new weights rows the number is 12, and the layer table and the filter now show the same counts.
• Coding agents, "Catalog entries seen by name only". jev-agent-router, jev-mcp-router, jev-memory-selector and jev-effort-router have now been read, so they leave that list.
• jev-effort-router, wherever it is described (Claude Code and MCP names it). Its catalog description of the failure paths does not match the code. The code at 562dc2a does something else: below threshold, or on an unknown choice, it rewrites the request to the default model deepseek-v4.1-flash at medium effort (R12-S202, R12-S204, R12-S207).
• Strands Decider row: pin 75c9fd3 → 0ac22a9; weight licence "not recorded" → apache-2.0 (card metadata); base licence "not recorded" → apache-2.0; the runtime now accepts images (--vision, per the README); the authors' own figure is now v1 180/231 on JevBench Reported. Attribution: "Copies are published by Amazon's verified Hugging Face organisation (Documented); The New Stack reports it as an AWS project (Reported)." (R12-S118 to R12-S125.)

Among the rows re-read this run, no other published cell disagrees with the research notes.

4. Access status and home notices

Checked 10 Oct 2026, 05:12–05:16 UTC: Documented. Status unchanged.

• Status open_with_conditions since 27 Sep (no new-user credit); service operational ("All services are online"). Limits unchanged (R12-S20, R12-S21).
• Shown on Home (2): new-user credit disabled (rank 1); lookalike resellers (rank 2).
• Retired (2):
  • Subprocessors added (2 Oct): its shelf life ended 9 Oct. It retires one day late because no run took place on 9 Oct. The list is the same 6 subprocessors and the only update is still 2 Oct (R12-S39). The fact stays on data privacy.
  • Rate limits changed (3 Oct): its retire time was 10 Oct 2026, 05:16:02 UTC, and the notice check was made at 05:16:03 UTC. Limits are unchanged (R12-S21) and stay in the access conditions.
• The freed slots stay empty. No timeline addition: no status-page event, signup, credit, price or limit change between 8 Oct 2026, 05:14 UTC and 10 Oct 2026, 05:16 UTC.

5. Findings per page and ship verdicts

5.1 alternatives (protected): ships

Count basis: the 27 published rows plus Jiwo 0.8B, TOD and the Decisions API, counted by layer by script. A row with two layers counts once, under its primary layer.

Short answers on the page:

• Official weights: "TypeSafe-published Jev weights or a local Jev runtime: not found on typesafe.ai (no /models page), the docs (llms.txt, models page, introduction), TypeSafe's GitHub organisation (12 repositories) or a Hugging Face author search (0 models), 10 Oct 2026, 05:11–05:20 UTC." The docs say "the same weights serve every account". The new Enterprise page asks "Can we run Jev in our own VPC or on-premises?"; its answer is not in the served page and was not read (R12-S100 to R12-S110, R12-S21, R12-S36, R12-S58).
• "jev ollama": you cannot run TypeSafe's Jev in Ollama on the evidence checked. Ollama 0.35 serves other Jev-style decision models (nimble, tev1) behind a System One-style endpoint (R12-S107, R12-S135).

Seen, not checked (for run 13): THX-01, Decision Studio, Jevfree, dex, jev-nano, decisionmodels-local, LLMtoSystemOne, iKev, Strata-JEV, luna-middleware-decisions, mini-jev-multimodal. Unverified

5.2 build/coding-agents: ships

Count basis: 18 published rows plus 4 read this run = 22. Counted 10, not counted 10, not settled 2 (10 + 10 + 2 = 22). Error split of the counted rows: 3 + 1 + 5 + 1 = 10.

• fail-closed or held: no action, or a person decides
• fail-open or acts-anyway: the action goes ahead
• fallback-<what>: a named substitute
• no-decision: nothing returned; left for a later run

• jev-effort-router is counted: it rewrites the model and the reasoning effort each turn, and Hermes sends the request without a person. The code rewrites to deepseek-v4.1-flash below threshold (R12-S200 to R12-S209).
• jev-gateway is counted: the proxy forces or builds the tool choice from Jev's answer. On When Jev fails it is a "fallback to the agent's own model" and is one of the 7 fallbacks that can still end in the action (R12-S241, R12-S257).
• Not settled (2) (on the page): RanyAlbegWein/jev-triage (OpenClaw) at 6360e21 gates the inbound message; it stays not settled until OpenClaw's handling of {handled: true} is read (R12-S224, R12-S254). abehrman/hermes-jev-mana (Hermes, off by default) at 3daf639: its core patch was searched, not read in full (R12-S226, R12-S255, R12-S256).
• Read, not counted: jev-agent-router, jev-mcp-router and jev-memory-selector return their answer to the agent (catalog and README only, Reported; R12-S212 to R12-S218). The Hermes catalog at main dce1e9b adds jev-cron-gate (13 jev entries; R12-S211).

Of the 11 candidate gates, 3 were read in source: 1 counted (jev-gateway), 2 not settled. The other 8 are README only Reported (R12-S219 to R12-S253). 20 more items from this run's scan are listed for run 13.

5.3 build/when-jev-fails: recount 69 → 71 (protected): ships

The extraction from the published cells first reproduced the 8 Oct totals (38/15/8/5/3 at n = 69). Every row of the table sums to 71.

• Below threshold (counted separately, never merged with errors): held 16, no-threshold 18, acts-anyway 15, fallback 9, no-decision 4, advisory 4, not recorded 5 (of 71).
• Bypass: recorded 38, not recorded 33 (of 71).
• Action goes ahead on a Jev error: 15 of 71 (8 fail-open plus 7 fallbacks). jev-gateway is the new one. jev-effort-router is not added, under the routing-fallback convention, so "4 routing fallbacks" becomes "5 routing fallbacks".
• The page's second key-finding sentence (4 moderation bots, 2 agent hooks, a passage filter, an eval tool) is unchanged.

5.4 benchmarks: ships as a partial refresh

Each lead was opened once at a pinned commit: the README plus the top-level file list. Results are Reported in the operator's own units; nothing was recomputed or run (R12-S60 to R12-S70).

The two Decisions API comparisons point in different directions by task: Jev ahead on the Portuguese judge task, the Decisions API ahead on the RPG task. Neither is graded, so no page states parity. Grading of the three leads moves to run 13. About 15 more leads are listed, not read.

5.5 products: ships

Two new rows, both code-backed, grade A: jev-effort-router (Hermes plugin, 562dc2a) and jev-gateway (npm 0.5.1). Recount: 96 rows (A 83, B 12, C 1), 10 Oct 2026.

6. Drift and route lines (registry check 10 Oct 2026, 05:18:47 UTC)

• pydantic-ai-slim 2.55.0 (9 Oct): a minor release; the targeted diff moves to run 13.
• Patch releases, not re-audited: litellm 1.104.2, typesafe-sdk 0.7.4, ai 7.0.137, @ai-sdk/typesafe-ai 3.0.17; @openclaw/typesafe 2026.9.9 (R12-S48, R12-S49).
• Six routes: IDs and alias targets unchanged (05:12:20–05:12:21 UTC). Versions and aliases.
• OpenRouter typesafe/jev-1.13: context 32,000 → 64,000 tokens (max prompt tokens 32,000; R12-S25).
• Status page: now also monitors api.us-west-2.typesafe.ai (R12-S20). What this endpoint means was not read.
• New Enterprise page: zero data retention for enterprise customers (R12-S58).
• Reseller rechecks: jev-agent.org, jev-ai.pro, jevtypesafeai.com, thejevai.com and the Eye Security report all answered HTTP 200 at 05:12:21 UTC (R12-S52 to R12-S55, R12-S42). Official vs reseller.

7. Evidence gaps and uncertainties

• Unread host and plugin paths: the Hermes enable step for jev-effort-router; Codex and OpenCode handling of forced tool calls (jev-gateway); OpenClaw's handling of {handled: true} (jev-triage); the hermes-jev-mana patch in full.
• Unread pages and files: the Enterprise page's on-premises answer; TOD's base licence and hosted API shape; the classifier-bakeoff report; the metric code of the three ungraded comparisons; the meaning of the us-west-2 endpoint.
• Strands Decider: an aws.amazon.com or AWS GitHub organisation source was not found on those surfaces, 10 Oct 2026, 05:12–05:13 UTC; the strands-labs GitHub organisation is unverified. Hence the split Documented / Reported attribution.
• Scan window coverage unknown for Reddit (HTTP 403), X, YouTube and TikTok (not queried).

8. Not done (for the next run)

• Source reads of the 8 remaining candidate gates.
• Triage of the 20 coding-agent items from this run's scan.
• The pydantic-ai-slim 2.55.0 diff.
• Checks of THX-01 and the other new alternatives seen this run.
• The jevplayground.com homepage.
• The other four n8n node versions.
• Grading of the three Decisions API and reranking comparisons (metric code first).

9. Timings (UTC, 10 Oct 2026)