Agents: read the raw Markdown of this page, or start at llms.txt.
Tools: agent routing, context and skill selection
TL;DR Plugins, hooks, proxies and skills that use Jev to pick a model, a tool, a skill or a session, or to cut context, for Claude Code, Codex, Pi, Hermes and DeepSeek Harness. Unaudited community code; several authors measured no saving or a slowdown, so read the Note column before adopting one. Split from Tools and integrations hub: MCP servers, plugins and connectors on 2026-09-25 (nothing dropped); safety gates are on Tools: guardrails for agents (tool-call gates, permission hooks, injection screens, rule checks).
Pattern IDs: Decision patterns from the community (with fit verdicts) (P02 model routing, P07 context, P08 skill and tool selection: Patterns: agent internals, routing, gates, context and memory). Licence MIT unless stated. Numbers: the author's. "Our check" = our 2026-09-25 source and registry vetting (docs/sweep/2026-09-25-mrjev/partc.md), not a raw capture. Proxies that sit in front of your agent (jev-gateway, Jevonian, Yoshi) see every request; local decision services (jev-gateway's /router/decide, HiRoute's jev-decider) spend your key for any local caller.
Routers, context tools and skill selectors
| Tool | Does | Install | ★ · push | Note |
|---|---|---|---|---|
| typesafe-skill-router (Ritesh Patel) | Hermes hook before the model call: wide chunked Choice with none_of_these, gate Nouls, then a shortlist; silent on disagreement; lists what it sends |
hermes plugins install typesafe-skill-router (opt-in, off by default) |
14 · 2026-09-21 | port of TypeSafe's skill-suggestion cookbook (Cookbook: Skill suggestion). Author: ~0.6–1.2 s, ~$0.001 per turn (292 skills); second check p50 1.93 s, max 8.56 s; fits ~0.14 lower in Spanish; one miss and one off-target in the first live turns. "Budget for output tokens" contradicts docs (output is free, Models, aliases, pricing, rate limits, context). P08 |
| pi-jev-router (sugarforever) | Pi: four questions (route, on_topic, needs_history, one_off) decide continue, fork or new session before a prompt enters |
pi install npm:@sugarforever/pi-jev-router |
14 · 2026-09-23 | ~45% of prompts skip Jev; any error means continue. Author: ~0.5 s, ~1.4k input tokens, ~$0.00006; 12/13 hand-labelled pivots acceptable, needs_history beat the route Choice. Not philippdubach/pi-jev-router (Builds: coding agents, harnesses and orchestration). P07 |
| jev-blindspot (jsk4581) | one Jev call per prompt (worth_checking Noul + 5-level risk Score) decides whether to wake a headless Sonnet or Codex run that lists what the prompt left out, in a side panel |
npm install -g jev-blindspot |
20 · 2026-09-24 | never edits the session; daemon on 127.0.0.1:7461; --trust writes the Codex hook trust hash. No accuracy measured. P38 |
| jev-gateway (vinilana) | local proxy for Codex, Claude Code, OpenCode, Gemini, Devin: each turn a Jev Choice over the agent's tools forces or hints a tool when confident, else passes through | npm install -g jev-gateway |
224 · 2026-09-25 | fails open; per-turn overhead unverified (our 2026-09-25 sweep note says ~0.5–1 s; the README capture shows only a 712 ms key check). Author's bench, 120 sessions: bug-fix cheaper for every model; the feature task cost Opus 5 +61% input and +83% time, and GPT-5.6 Luna solved 3/5 vs 5/5 → Head-to-head: Jev inside agents, routers and tool gates. Routes include OpenCode Zen. P08 |
| JCR (NiazMorshed2007) | walks a tree of capability docs: Jev ranks children with a no-match option, keeps paths within 60% of the best; only chosen command docs reach the agent | clone | 20 · 2026-09-20 | author, 80 runs: agent input −85% (Opus 5), −23% (GPT-5.6-Sol); but Sol slower in 19 of 20 scenarios (median 23.2 → 45.4 s); per-run data not committed → Head-to-head: Jev inside agents, routers and tool gates. Bundles ~80 third-party skills without stated licences (our check). P08 |
| HiRoute (Higress group) | local routing engine for long agent runs; optional self-hosted jev-decider REST service: Choice picks the branch, optional Score rates the last segment |
pip install . (decider); core via checksum-verified script |
21 · 2026-09-25 · Apache-2.0 | defaults 0.80 / 0.50 are "not calibrated guarantees" (author); state capped at 24,000 bytes, 2.8 s budget. Keep the decider on loopback: it spends your key. No accuracy. P02 |
| codex-context-diet | Codex PostToolUse: 5 questions per bulky result; stale ones cut to head + note; local Laya option |
codex plugin marketplace add konstantinosbotonakis/codex-context-diet, then codex plugin add codex-context-diet@context-diet |
3 · 2026-09-23 · NOASSERTION | 20 of 65 results cut; ~$0.00003 each. P07 |
| Yoshi (CompozyOS) | proxy for Claude Code and Codex; Jev picks history to omit | Bun | 25 · 2026-09-18 | both current trials slower, no validated saving. P07 |
| Switchloom (Instructa) | Codex skill routing tasks across models, Jev mode | cargo install |
1 · 2026-09-20 | author advises against: no Jev gain in two pilots. P02 |
| Hunch (Kelbie) | plain-English repo search, diff review rules, GitHub Action | npx skills add Kelbie/hunch |
4 · 2026-09-23 | TypeSafe, gateway or local SemIf. Not steven-shoemaker/hunch or carldaws/hunch (Community SDKs and libraries). P10, P17 |
| Jevify (altryne) | offload bulk judgments; "Jevify this" audits a product | npx skills add altryne/jevify --skill jevify |
25 · 2026-09-21 | not ryana/jevify or fidecastro/jevify (PyPI jevify, Open replicas and Jev-compatible servers) |
| Jev Architect (Karan Bansal) | find, design, test decision loops; may say "keep it in code" | npx skills add or plugin |
6 · 2026-09-20 | cf. Consult guide: could Jev help this project? |
| Augustus (24601) | skill: find placements, build, evaluate and hill-climb decision systems; Jev the default hosted example | claude plugin marketplace add 24601/Augustus@v0.7.2, then claude plugin install augustus@augustus |
10 · 2026-09-25 | nearest community analogue to Consult guide: could Jev help this project?. P42 |
| prompt2jev (sumleo) | prompt or LLM code → linted request.json + script |
plugin or npx skills add |
3 · 2026-09-21 | P01 |
| clay-jev-people-ranker (PromptGTM) | Clay filters, Choice + Noul qualify founders | install_skill.py |
0 · 2026-09-21 | 577 people in 20.1 s; 490,395 input tokens, author's estimate $0.0206 (= 490,395 × $0.042/Mtok, our arithmetic). P19 |
| SkillRanker (Dicklesworthstone) | Rust CLI ranking agent skills for the next step from live session context; Claude Code hooks, abstention, offline replay | install script or cargo install |
120 · 2026-09-25 · MIT with an OpenAI/Anthropic rider (GitHub: NOASSERTION) | a failed hook recommendation falls back quietly. P08 |
| JevRouter (BillionsBobby) | one Choice over models, subagents, skills, MCP tools, CLIs; its code enforces availability, permissions, risk, confirmation | npx --yes github:BillionsBobby/JevRouter agent start ... (unpinned: runs the default branch) |
222 · 2026-09-21 | Toolathlon comparison vs DeepSeek V4.1 Flash is 10 tasks, author-run. P02, P05 |
| pi-typesafe (DevMortimer) | Pi: typesafe_evaluate tool (1–32 questions), /typesafe playground, one client and key store other extensions share |
pi install npm:pi-typesafe |
46 · 2026-09-24 | tool off until consent; key file owner-only; 32 questions is its cap, not TypeSafe's. P01 |
| jev-use (shitianfang) | Claude Code, Codex, pi: steps needing no text go to Jev, typed escalation back to the LLM | npx -y jev-use install |
27 · 2026-09-22 | via Vercel Gateway 2026-09-19: single call p50 223 / p95 364 ms; 12 questions in one call 224 ms vs 12 calls 2,662 ms. P01, P07 |
| dsh-jev-tools (HorusJiang) | DeepSeek Harness: ranks tool-output segments to prune, screens fetched pages for injection, picks a skill, jev_gate before "done" |
dsh plugin --profile web add dsh-jev-tools |
6 · 2026-09-24 | ranks, never thresholds; pruning fails open, the gate fails to escalate; author: picking well is unmeasured. Beside buberlo/dsh-jev (Tools: guardrails for agents (tool-call gates, permission hooks, injection screens, rule checks)). P07, P08, P16 |
| Jevonian (xinyao27) | local proxy: one Jev call picks model and thinking level per turn | npm install --global jevonian |
4 · 2026-09-23 · AGPL-3.0 | flag: reuses your Claude Pro/Max and ChatGPT sign-ins through a third-party client; author: "at your own risk". P02 |
Names only (our check): pi-jev (TheoOliveira; unscoped npm pi-jev; Pi skill suggestions at Noul ≥ 0.65, opt-in routing, guard, compaction, pi-jev-gate CLI; npx pi-jev-gate outside a pi-jev project fetches another publisher's package); jev-pruner (tamaratran; trims Bash output over 10,000 estimated tokens, one Noul per chunk, original passes through on failure; needs CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1; same author as fast-jev-compaction on Repos: coding agents, orchestration, memory, browser and computer use, integrations); jev-dsh-decision (Devin-AXIS, no licence; recommends tools, skills or agents in DeepSeek Harness, never runs them; duplicates dsh-jev-tools); Eval Genius (Alex Greenshpun, Apache-2.0; judge-cascade skill that works on exported answers, no live Jev calls).
Elsewhere: building-with-jev-skill, ryana/jevify, hermes-jev-skills, fast-jev-compaction, winnow, Canny, pi-warden, jev-codex-router, jev-router, compact-adviser, jev-rules → Repos: coding agents, orchestration, memory, browser and computer use, integrations.
Related
- Tools and integrations hub: MCP servers, plugins and connectors — hub: MCP servers, connectors, key safety, name collisions
- Tools: guardrails for agents (tool-call gates, permission hooks, injection screens, rule checks) — tool-call gates and rule checks
- Head-to-head: Jev inside agents, routers and tool gates — router, gateway and skill-picker measurements (jev-gateway, JCR, JevRouter and others); Head-to-head: Jev against other models and methods — jev-use's latency and cost run
- Repos: coding agents, orchestration, memory, browser and computer use, integrations, Builds: coding agents, harnesses and orchestration — agent repos and builds in depth
Sources
Links inline; raw captures (2026-09-23 to 25) in frontmatter. Names-only lines and "our check" notes rest on our 2026-09-25 vetting at source (docs/sweep/2026-09-25-mrjev/partc.md, parta.md) where no capture exists.