One coding agent. Every model.
Routed per turn, automatically.

The right model for backend is not the right model for UI. Zhivra picks per turn, and shows you why.

For users of Cursor, Cline, Continue, or Roo Code who want routing they can see, audit, and override.

Install for VS Code View source on GitHub

Marketplace listing pending. BYOK — one key, hundreds of models.

The Inspector

Every routing decision is on the table.

No black box. Every turn shows which capability was detected, which model it routes to, and a one-click way to swap if you disagree.

ZHIVRA — capability: 1 of 13 categories

you

Build a Postgres-backed rate limiter for the /api/upload route. Token bucket, per-IP, refill 10 per minute, burst 30. Include the migration.

Capability: backend-api-build · Claude Opus 4.8 · override

zhivra

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. Override in one click.

What you get

Three commitments. Held to the line.

Routing per turn, not per session

Every prompt is classified by capability (backend, UI, debug, refactor, research, image, voice) and sent to the model ranked best for that work. No mode picker. No session-wide model lock.

No black-box ranking

Every routing decision happens against a versioned, signed cap-table you can read and audit. Updates ship weekly, gated by a canonical eval set. No vendor deciding for you in a closed loop.

Your keys, your prompts

Bring your own provider key. We never see it, and we never see your prompts or your code. Telemetry is opt-in and limited to behavioral signals (accept, reject, edit-distance).

How routing learns

Classify. Look up. Re-rank.

A deterministic classifier reads each prompt, picks the right capability, and looks up the current best-ranked model in a versioned cap-table. No black box.

Step 01
Classify the prompt

A regex-and-keyword classifier runs first. Backend verbs, UI nouns, debug phrasing, refactor patterns, image and voice intents each have weighted patterns. The classifier returns a capability label plus a confidence score. Over 80% of prompts are decided this way, in single-digit milliseconds, without calling a model.

technical Ambiguous prompts fall back to a small classifier model. The full feature vector and the candidate scores are logged on every turn.
Step 02
Look up the cap-table

The cap-table maps each capability to a ranked list of models with cost and context-window metadata. The chosen rank-1 model handles the turn. The badge in the chat panel shows the capability, the model, and the confidence. One click reroutes the same prompt through any other ranked model.

technical The active cap-table version is content-addressed (SHA-256) and the previous four versions are retained for rollback.
Step 03
Re-rank weekly from beta-cohort signals

Opt-in behavioral signals (accept, reject, immediate-redo, edit-distance, discard) feed a stratified quality score per (capability, model). A weekly job proposes rank updates, runs them against a canonical 50-prompt eval set, and rejects any version that regresses below baseline. During beta, the founder reviews and approves each promotion before publication.

technical A 48-hour rolling acceptance check at the edge auto-rolls back a published version if accept-rate drops by more than five percentage points across at least 100 turns.

The classifier, the router, the cap-table format, and the learning pipeline are open source under BSL 1.1. Read the code at github.com/Zhivra/zhivra.

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 per turn based on what you are asking, 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.

SD Suvajit Das · Founder

Join the beta cohort

Help shape the cap-table.

We are onboarding a hand-picked group of solo founders and small teams to shape the cap-table before public release. Leave your email and a one-line note on what you build, and we will reply within a week.

We use your email to reply about the beta. Nothing else, no list.

Or just install the beta now — grab the latest release .

One coding agent. Every model. Routed per turn, automatically.

Every routing decision is on the table.

Three commitments. Held to the line.

Routing per turn, not per session

No black-box ranking

Your keys, your prompts

Classify. Look up. Re-rank.

Classify the prompt

Look up the cap-table

Re-rank weekly from beta-cohort signals

Help shape the cap-table.

One coding agent. Every model.
Routed per turn, automatically.