Routing by Difficulty: Jev Codex Router, JevRouter and hono-jev-router
Three open-source routers use Jev to decide where a request goes. One reports a 60% list-price saving from a replay of 237 turns. The fine print matters as much as the number.
Routing is the most natural job for a decision model. Something arrives, several destinations exist, and you need the right one fast and cheaply. Three open-source projects apply Jev to that problem at three different layers: which model handles a coding turn, which capability an agent should invoke, and which handler serves an HTTP request. They share an idea and diverge in almost everything else.
This article walks through each, quotes what their READMEs actually claim, and flags where the evidence is thin. The routers are all days old as of September 2026, so treat them as patterns to learn from rather than products to bet on.
Key Takeaways
- Jev Codex Router reports about 60% lower list-price cost on a 237-turn replay, from one person's workload.
- The fallback tier mattered more than the threshold. Its backtest found falling back to a middle model, not the frontier one, was what delivered the saving.
- JevRouter keeps policy outside the model. Permissions, availability and risk rules cannot be overridden by probabilities.
- hono-jev-router costs a model call per request and its README says not to use it for authentication or authorization.
- All three are early, with stars of 26, 4 and 19 respectively as of September 2026.
Jev Codex Router: choosing the model per turn
Jev Codex Router (26 stars) estimates how hard a turn is, with attention to technical complexity and context depth, before deciding whether a cheap or frontier model runs it. Its backtest notes describe 237 turns from a seven-day Codex session replay. Each prompt was classified once by Jev, and repeats reused earlier decisions. Costs use published OpenAI API rates as of September 2026.
Reported result: a full-frontier baseline of $871 fell to $349 under routing, or -59.9%, with 61 turns to the lowest tier, 165 to the middle tier and 11 to the top tier.
The most interesting finding is not the headline. When low-confidence cases fell back to the frontier model, savings were only -11.9%. Switching the fallback to the middle tier produced the full reduction. The author's takeaway is that the fallback tier matters far more than confidence-threshold tuning.
The caveats, in the repo's own terms: token volumes held constant across scenarios; prompt-cache invalidation from switching models mid-thread is not modelled; it is one operator's workload; and pricing includes promotional rates active through late 2026. In short, a list-price estimate from a replay, not a reduction in a paid bill. It is also listed on aiskill.market as Jev Codex Router.
JevRouter: models, tools, skills and subagents
JevRouter (4 stars) widens the scope. It turns models, tools (MCP tools, CLI commands, plugins), skills and subagents into one candidate set, asks Jev a single typed Choice question, and returns probabilities. Large candidate sets get a coarse top-K pass before the final Choice.
The design choice worth copying: "Jev owns the decision probabilities; JevRouter owns availability, permissions, risk and confirmation." Missing permissions or prohibited risk levels cannot be overridden by the model, and filtered candidates keep their original probabilities rather than being renormalized. It runs in decision-only mode: it recommends, and the host agent executes. Find it as JevRouter.
hono-jev-router: routing HTTP by meaning
hono-jev-router (19 stars) is a semantic router for the Hono web framework. You register descriptions; on each request Jev is asked one yes/no question per description and the first above the threshold wins.
The README is unusually direct about risk. Every semantically routed request triggers a model call, adding latency and cost, so "put a rate limit in front of it." It sends the method, URL, headers and (a bounded amount of) textual body to a third party. And: "Do not use it for authentication or authorization," because whoever sends the request can try to steer the answer. Matching is probabilistic and can misclassify. See hono-jev-router.
Comparing the three
| Project | Routes | Reported evidence |
|---|---|---|
| Jev Codex Router | Turn to model tier | Replay of 237 turns, list-price estimate |
| JevRouter | Model, tool, skill or subagent | Design docs; no published benchmark found |
| hono-jev-router | HTTP request to handler | Usage docs and warnings; no benchmark found |
What to do next
- Replay your own logs through a router in shadow mode before it controls anything.
- Choose the fallback tier deliberately; the Codex backtest suggests it is the lever.
- Keep permissions and risk in code, as JevRouter does.
- Learn how to set thresholds in calibrating Jev thresholds, and compare with typed vs free-text decisions.
- Browse the radar for more routers.