Still switching models by hand in Cursor, Cline, or Roo Code? Zhivra routes each task to the best-suited model — premium where the work earns it, cheap where it doesn't — and shows you why.
Private beta · by invitation. BYOK — one key, every model in the pool.
The Inspector
No black box. Every turn shows which capability was detected, which model it routes to, and the reasoning — with a persona override for your next turn if you disagree.
Build a Postgres-backed rate limiter for the /api/upload route.
Token bucket, per-IP, refill 10 per minute, burst 30. Include the migration.
Here is the migration plus the token-bucket middleware.
-- migrations/0014_rate_limit_buckets.sql
create table rate_limit_buckets (
ip inet primary key,
tokens numeric(6,2) not null default 30,
last_refill timestamptz not null default now()
);Every turn shows the capability and the model. Disagree? Override the persona for your next turn.
What you get
Each task is classified by capability (backend, UI, debug, refactor, research, planning, code explanation) and sent to the model ranked best for that work — a flagship where the task needs one, a fast, cheaper model where it holds up. Zhivra holds the pick while you stay on the task and switches only when the work genuinely shifts. No mode picker. No session-wide model lock.
Zhivra shows you which persona and model handled each task — and the reasoning behind the pick — right in the chat. The ranking is curated against a canonical eval set, not a vendor deciding for you behind a closed prompt.
Bring your own provider key. We never see it. Your prompts and code go only to your provider — unless you separately opt in to routing-improvement uploads. Default telemetry is opt-in and behavioral signals only (accept, reject, edit-distance).
How routing works
A deterministic classifier reads each prompt, picks the right capability, and routes to the current best-ranked model — and shows you the call. No black box.
A regex-and-keyword classifier labels each prompt's capability — backend, UI, debug, refactor, research, planning, code explanation — using weighted patterns. Most prompts are decided in milliseconds, without calling a model; genuinely ambiguous ones fall back to a fast LLM classifier.
Each capability maps to a ranked list of models. The top-ranked model handles the task; a badge in the chat panel shows the capability and the model, with the routing reasoning one click away — and if you disagree, a persona override redirects your next turn.
Opt-in behavioral signals — accept, reject, immediate-redo, edit-distance — are collected to tune the rankings. During beta, every ranking change is human-reviewed before it ships; nothing self-modifies silently.
The classifier, router, cap-table format, and learning pipeline are licensed under BSL 1.1 (source-available at public launch).
Getting started
Accepted into the beta? You will get an install link by email. Drop the file onto the Extensions panel in VS Code, Cursor, or VSCodium and reload. No Marketplace listing yet — install is by invitation during the private beta.
Bring your own key — one OpenRouter key covers the whole routing pool, or use a direct provider key. The key stays on your machine; we never see it. You only ever pay your provider, at their rates.
Zhivra classifies each task and routes it to the best-suited model+prompt persona, holds that pick while you stay on the task, and switches only when the work genuinely shifts. You see every choice and can override it.
Questions
You bring your own provider key, so your prompts and code go straight from your machine to the model provider you chose — never through us. Under the default anonymous-telemetry channel, uploads are behavioral signals only (which persona handled a turn, accept/reject, latency, cost) — no code, no prompt text. A second, separately opt-in routing-improvement channel can upload prompt and response text to improve routing — it is off unless you explicitly enable it. You will see a consent dialog on first run; you can decline or turn either channel off anytime, and your prompts and code stay local unless you explicitly enable that second channel.
Those make you pick a model or switch a mode by hand. Zhivra auto-routes every task to the model+prompt persona ranked best for that kind of work, pins it while you stay on the task, and breaks to a different one only when the work genuinely shifts — and shows you the reasoning. The routing is the product, not an add-on.
One OpenRouter key reaches your provider's full catalog, or use direct keys for Anthropic, OpenAI, Google, DeepSeek, and others; Zhivra routes over a curated pool where every model earns its slot on the eval. It runs in VS Code, Cursor, and VSCodium (any VS Code–compatible editor). Built on the Roo Code (Apache 2.0) base, so the agent itself will feel familiar; the routing layer is the new part.
Free during the private beta. You pay only your own model provider, at their rates — there is no token markup. Paid plans for the managed features come later; beta participants will hear about pricing before anything changes.
Request access below. We are opening a small, hand-picked first group and reply individually — it is not an open signup yet.
A note from the founder
Honestly, I built this because I had the same problem you probably do too. There are now more good coding models than anyone can reasonably keep up with. Claude is great at one thing, Gemini at another, OpenAI at a third, and a handful of open models keep surprising me on specific tasks. Picking which to use when becomes its own job, and paying three or four subscriptions to stay covered feels wrong even when each one is reasonably priced.
For a while I tried just picking the best one and sticking with it. That did not work. There is no single answer that holds for long. Different models win different categories, and the rankings shift with every new release. So instead of betting on one model for everything, I wrote a router that picks the best one per task based on what you are asking, holds it while you stay on the task, switches when the work genuinely shifts, shows you the choice, and lets you override when it gets things wrong.
This is still beta. The router is decent but not done. I review every change to the rankings myself, against a small set of canonical prompts, before anything ships. I am sure it gets some things wrong. If you spot something that feels off, please write me. That feedback is most of what shapes the next version.
Private beta · by invitation
Zhivra is opening a hand-picked private beta — a small group of solo founders and small teams who will shape the router before public release. There is no public download yet; leave your email and a one-line note on what you build, and we will send an install link if it is a fit.
We use your email to reply about the beta. Nothing else, no list.