Organization repository on GitHub
Experimental Hermes plugin: Jev-assisted model routing plans with budget and capability constraints. API access pending.
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Hermes Jev Router
Experimental Hermes plugin that asks Jev to label task complexity, then deterministic policy code picks the cheapest eligible allowed model and returns a routing plan. It does not dispatch the request.
- Stars
- 1
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- ussyverse
- Last verified
- Oct 2026
About this project
An opt-in Hermes plugin that uses TypeSafe Jev to assess task complexity and deterministic policy to select an allowed model within configured cost, token, context, latency-estimate and capability constraints.
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- The native jev_route_plan tool returns provider and model IDs plus an output cap; it does not switch models, run a completion or install a provider override.
- Complexity is one named Choice question; eligibility, cost and capability filtering are deterministic code, not the classifier.
- External calls need allow_external_jev, a key, and an approved_sha256 matching the payload; a hash only detects changed payloads.
- Python 3.11+ standard library only, no default network calls, and the example policy ships fictional placeholder prices, latencies and model IDs.
Watch out
No live Jev calls have been made because API access is pending, so every fixture in the repo is synthetic and is not evidence of quality or savings. The README is blunt about the scope: this is a working offline planner and plugin foundation, not a production router and not a catalog-approved plugin. Every price, latency and model id in policy.example.json is a fictional placeholder you must replace with measured values. Input token counts are caller-supplied estimates you have to produce with the right tokenizer; the planner is per-request, not a concurrent or global spend ledger, and reasoning-token charges, caching, multimodal charges and provider-specific billing are not modeled, so the README says not to advertise the estimate as an enforced cap. The HTTP socket timeout is not an end-to-end deadline, and over-budget results are discarded only after classification, which may already have cost money. The 0.8 confidence threshold is described as an uncalibrated starter value. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
It prints a plan. The tool returns provider and model IDs and an output cap; it does not switch the current Hermes model, execute a completion, install a provider override or collect conversation history. An OpenAI-compatible routing gateway with cancellation, streaming, tool calls and session affinity is listed on the roadmap as a separately reviewed integration, and the README notes that registering a tool does not transparently intercept requests.
No, and the README says so directly: the hash detects changed payloads but is not a substitute for human permission and is not a security boundary against an agent that can generate hashes or edit policy. Preview payloads contain your summary and may be logged by Hermes, so secrets and personal history must be omitted; the plugin makes no claim of automatic redaction. Nothing is persisted as routing history by the plugin itself.
No. The planner reserves router overhead and is per-request rather than a global ledger, so concurrent requests are not accounted against each other. Reasoning tokens, caching, multimodal and provider-specific billing are unmodeled, and the classifier itself can spend time and money even when no route is selected. The README explicitly says not to advertise the resulting estimate as an enforced billing cap; a global reservation ledger is on the roadmap.
It is a starter value, not a safety guarantee, and Jev classification is kept separate from deterministic routing on purpose. Unknown, low-confidence or failed classification uses only the configured fallback if it fits, otherwise it abstains; there is no unlimited escalation, no retry billing and no silent cheap downgrade for a complex task. The README cites RouteLLM on calibrating thresholds against your own query distribution and recommends validating confidence on local labeled tasks before enabling autonomous routing.
Install with hermes plugins install ussyverse/hermes-jev-router --no-enable, review the code, then run hermes plugins enable jev-router. Copy policy.example.json to a private operator-owned location and set JEV_ROUTER_POLICY to its absolute path in the environment that launches Hermes. Every price, latency and model id in the example is a fictional placeholder.
Only the summary and the fixed rubric go to TypeSafe, not the full request, model registry, prices, tokens or credentials of the target providers, and errors do not echo remote response bodies or secrets. Preview payloads contain your summary and may be logged by Hermes, and the plugin makes no claim of automatic redaction, so omit secrets and personal history. No routing history is persisted.
A foundation. No live Jev calls have been made because API access is pending, and all fixtures are synthetic, not evidence of quality or savings. Tests report 20 passed and the stock Hermes plugin doctor passes with one tool and zero hooks, but the README calls it a working offline planner and plugin foundation, not a production router or catalog-approved plugin.
More Jev resources like this
23GitHub stars
Coding-agent plugin that routes prompt classification, subagent tiering, edit rule checks, and fast Jev-backed compaction through small typed judgments.
Jev Agents & MCP#community#jevbest#catalog-import
Community
0x7067/claude-jevTypeScript
0x7067/claude-jevTypeScript14GitHub stars
Opt-in Hermes plugin that routes each turn to one skill: a wide Jev Choice plus request checks gate a narrow fits round, injecting a single ignorable hint line.
Jev Agents & MCP#community#jevbest#catalog-import
9GitHub stars
Hermes Agent plugin exposing Jev as four coding-workflow tools for parallel judgments, binary checks, option routing and rubric scoring, with a bundled question-design skill.
Jev Agents & MCP#community#jevbest#catalog-import
Community
ajensenwaud/hermes-jev-pluginPython
ajensenwaud/hermes-jev-pluginPython6GitHub stars
LangChain deepagents middleware that replaces SkillsMiddleware with per-turn judging, loading only the SKILL.md files a turn needs so large catalogs stay out of the prompt.
Jev Agents & MCP#community#jevbest#catalog-import
3GitHub stars
Python CLI that classifies a request into one of three tiers — local rule, structured decision or skill, or foundation model — using TypeSafe Jev, 20 bundled skill profiles with exclusion conditions, and local environment checks.
Jev Agents & MCP#community#jevbest#catalog-import
Community
maker-KK/todo-jevPython
maker-KK/todo-jevPython1GitHub stars
An OpenAI-compatible proxy that spends one Jev call per request on injection, harm and PII checks plus a tier label, then routes to a small or large backend from pure code.
Jev Agents & MCP#community#jevbest#catalog-import
Community
aniruddh-krovvidi/switchboardPython
aniruddh-krovvidi/switchboardPythonFrom the Jev community
Posts from builders shipping with Jev right now.