Introducing Jev Model Router for Claude Code This Claude Code Mod lets you use Jev through its direct TypeSafe AI API or Vercel AI Gateway With every request you send to Claude Code, Jev classifies the subagent model, main model (only at session start to avoid breaking the cache),... and effort level Install it with one command: npx claude-code-templates@latest --mod productivity/jev-model-router Full Mod details:show more

@typesafeai @vercel Holly uncached tokens, batman.

@typesafeai @vercel 上线没几天 就从新模型变成路由组件了 生态长出来的速度有点猛

@typesafeai @vercel We need Router to use Jev inside Claude code, what about Codex, bro?

@typesafeai @vercel Prompt caching 🪦🪦

@typesafeai @vercel Keeping main-model routing off by default is the right call. A cheaper tier stops being cheap if every switch throws away a long prompt cache.

@typesafeai @vercel and wave good bye to the cache and your tokens.

@typesafeai @vercel @dani_avila7 interesting approach with the Jev Model Router. curious, how does it handle model selection with varying workloads? handling this balance is often a tricky part in my builds too.

@typesafeai @vercel Log routes, or bills stay blind guesses.

@typesafeai @vercel wait Jev picks effort per request.

@typesafeai @vercel 每条请求自己分流 子任务走便宜的 主模型只在开局定一次 缓存也没被打断 今晚先把这条路由装上试试

@typesafeai @vercel 路由的关键不是省一次调用，而是把模型选择变成可观察策略：成本、延迟和质量都要有回放评估。

@typesafeai @vercel A router mod is the practical layer

@typesafeai @vercel I feel like this is going to be really good as long as you've benchmarked each model against the task that's going to happen and Jeff has data to work from when choosing which model to pick. Otherwise, it feels like it's just guessing without bringing true value to the decision.

@typesafeai @vercel This is awesome!

@typesafeai @vercel Does it expose why a model was picked? I'd want that context when debugging a bad result.

@typesafeai @vercel Caching at session start is the smart move.

@typesafeai @vercel Context window cost is a function of length squared. routing cheaper models for routine queries cuts that curve before it spikes. finally, ai that reads like it saves money instead of burning cash.

Routing is the honest answer to 'which model is best' — it depends on the step. The hard part nobody ships well is classifying the task before spending tokens: most routers guess from the prompt instead of the diff. Curious how it handles mid-task escalation — when the cheap model is 80% through a refactor and starts hallucinating.

@typesafeai @vercel the installation command looks super simple, just npx claude-code-templates@latest. nice!

@typesafeai @vercel Every request now has a middle manager deciding which model deserves the effort. Corporate structure is unavoidable.

@typesafeai @vercel or you can just to route to any model, includign Jev (when they will provide me access to debug it).

@typesafeai @vercel Model routing inside Claude Code feels like the adult version of “let’s just try another model”—especially if the route is visible in the logs. Otherwise the bill arrives with a plot twist.

@typesafeai @vercel You're a rockstar!!!

@typesafeai @vercel inb4 T3 start complaining

@typesafeai @vercel classifying per-request without breaking the prompt cache is the actual hard part, nice that it's called out explicitly

@typesafeai @vercel Smart caching strategy. Classifying the main model only at session start is a clever way to keep latency low.

@typesafeai @vercel classifying the subagent at the start could add unnecessary complexity

@typesafeai @vercel Amazing 🤩

@typesafeai @vercel the install is easy. if you never log why it picked that effort the next bill is a guess. i stamp the route next to the tool calls

@typesafeai @vercel The core challenge with model routing is managing state persistence across agent calls. If the subagents operate stateless, ensuring context history remains correctly scoped requires careful serialization and rehydration after each classification step.

@typesafeai @vercel Routing the main model only at session start is the interesting constraint. Does Jev still adapt when a subagent's task suddenly needs a much larger context?

@typesafeai @vercel Cool but why don’t just make a hook and use your own subscriptions haiku?

@typesafeai @vercel Pretty cool, will check it out, perhaps you can default it to your vLLM?

@typesafeai @vercel the cache detail is the whole game. next step is rerouting at explicit task boundaries, because long agent sessions change jobs without changing chats.

@typesafeai @vercel @grok does adaptive subagent routing win mainly on cost, latency, or fewer broken handoffs—and what metric would prove it survives a long coding session?