Animation: a planner node hands small steps to a pool of fast workers. Each finished step passes through a verification gate. Steps that fail retry, and the hardest climb an escalation ladder until the planner sends back a short fix brief.

The Kaliro engine · Early access

Yourplan,carriedfurther.

Kaliro is the engine that does the typing. Your Claude plans and reviews, Kaliro writes the code, and every step has to pass a verification gate before it reaches your project. Use it in the desktop app.

Early access is private. We'll write when it opens.

See how it works

Held-out benchmark: 5.8x less Claude usage, same pass rate. See the results.

Results

Same pass rate. Far less Claude.

The latest Kaliro engine with Claude planning, against Claude Code on Opus doing the whole task alone. 25 real bug-fix and feature tasks from open-source repos, graded by hidden tests.

5.8x

less Claude usage, with the same pass rate

25/25
passed, the same as Claude Code (25/25)
$2.83
of Claude usage, against $16.48 for Claude Code
6.6x
less on a typical task
25 of 25
tasks used less Claude than Claude Code

Claude usage per task, same tasks

  • dateutil-h201$0.75$0.09
  • immer-h001$1.00$0.10
  • ky-h002$0.65$0.09
  • p-map-h001$0.37$0.03
  • pathe-h101$0.92$0.07
  • qs-h001$0.55$0.05

Claude CodeKaliro

New projects from scratch

4 apps built from a prompt, graded by 183 hidden tests.

12.8x

less Claude usage

183/183
hidden tests passed (Claude Code: 4/4 tasks)
$0.38
of Claude usage, against $4.85
17.7 min
comparable build time, against 14.6 min for Claude Code

Steady gains, same 25 tasks

  • Opus-takeover build3.2x
  • Engine 1.34.1x
  • Engine 1.4.04.9x
  • Latest engine5.8x

Benchmarks: 25 held-out open-source tasks graded by hidden tests; Claude usage valued at API-equivalent prices. Latest engine.

Why it's cheaper

Your subscription goes further.

Most of a planner's cost is re-reading your project on every tool call. Kaliro takes the tool loop away from your Claude and keeps its context small. Same task, two ways.

Opus alone0 planner tokens
Kaliro0 planner tokens
  1. Add Invoice model
  2. Write migration
  3. Create invoice service
  4. Validate totals and lines
  5. API route and tests
  6. Update the docs
Illustrative numbers, not measurements.
Why the gap

Fewer calls.

The planner is called a handful of times per task, not on every tool call. The engine runs the tool loop, and hand-backs send Claude only the new information.

The lanes are an illustration, not data. Measured results are in the section above.

One way to use it

The desktop app.

The desktop app puts the engine in front of you. It works on your existing project folder, next to whatever editor you already use. You always see the running step and what's next, and checks and hand-backs stay small and quiet until something fails.

App previewDesign preview, not the shipped app. Figures are examples.
How it works

Big thinking at the head. Fast work in the trail.

Kaliro splits a coding task in two. The Claude you already pay for thinks. Kaliro's engine does the typing, and a gate checks every step.

  1. Your Claude

    Plan

    Your Claude, signed in with your own subscription, breaks the work into small steps and writes the acceptance tests for each one. It plans once, instead of re-reading your project on every tool call.

  2. Kaliro engine

    Execute

    The Kaliro engine writes the code for each step, fast and at low cost. Several attempts can run in parallel on harder steps.

  3. Verification gate

    Verify

    Type-check, build, lint, tests, static analysis and a secret scan run on every step. Nothing is merged on anyone's say-so.

  4. Hand-back

    Hand back

    If the engine can't finish a step, it hands it back to Claude with what it tried. The step climbs one rung at a time, and your Claude writes a patch only as a late fallback.

Try the escalation ladder

Step through what happens when a step keeps failing. The gate runs after every rung.

Engine only

Routed engine, adaptive N

A step is labelled by difficulty. The router picks how the engine runs the step and how many candidates to try: one for routine steps, more for hard ones. Every candidate goes through the same gate.

Gate: types · build · lint · tests · sast · secrets

Trust and privacy

Your code goes where you'd expect it to.

Planning stays in your Claude account, execution runs on US and Canadian infrastructure, and nothing is retained.

  • You review every change

    Each step has to pass its checks first, and the result arrives as a diff you review before it lands.

  • Planning runs on your own account

    Kaliro signs in through the official Claude client, with your subscription. We never extract, copy or reuse your login tokens.

  • Code goes to US and Canada providers only

    The engine runs on a provider with US and Canadian data centres. Nothing goes anywhere else.

  • Zero data retention

    We require zero data retention on the provider account and a data-processing agreement before any of your code is sent.

  • Scanned before it reaches you

    Static analysis and a secret scan run in the gate on every diff the engine writes, before it reaches your project.

TODO(legal): link the privacy policy and data-processing terms once they exist.

Coming soon

Also as an API and an MCP server.

Use the same plan, execute and verify loop from your own tools: point an agent, a script or another editor at Kaliro. Not available yet.

Illustration only. The configuration is not final.

# illustration only, not final
{ "mcpServers": { "kaliro": { "command": "kaliro", "args": ["mcp"] } } }
Pricing

Pricing is not final.

These tiers are placeholders. The goal is a price clearly below the subscription capacity Kaliro saves you. Join the waitlist and we'll share real numbers first.

  • Starter

    Placeholder

    $—/ month

    For trying Kaliro on a side project.

    • Placeholder: limited runs per month
    • Placeholder: one planner connection
    • Community support
    Join the waitlist
  • Pro

    Placeholder

    $—/ month

    For daily work on real codebases.

    • Placeholder: higher run limits
    • Placeholder: parallel candidates on hard steps
    • Email support
    Join the waitlist
  • Team

    Placeholder

    $—/ month

    For groups that share standards.

    • Placeholder: shared settings
    • Placeholder: usage reporting
    • Priority support
    Join the waitlist
FAQ

Straight answers.

What is Kaliro?

Kaliro is an engine for writing code. Your Claude plans, reviews and talks with you, and the Kaliro engine writes the code. Every step has to pass a verification gate before it reaches your project. You use it through the desktop app.

How do I get it?

Kaliro is in private early access. Join the waitlist and we'll send a download when it opens.

Is this a replacement for Claude?

No. Claude stays the planner and uses your own subscription. The engine takes over the repetitive writing, testing and fixing, so your Claude is spent on what it does best.

How much further does my subscription go?

On a held-out benchmark of 25 real bug-fix and feature tasks, the latest engine passed 25/25, the same as Claude Code on Opus, using 5.8x less Claude usage, and up to 6.6x less on a typical task. Building 4 new apps from a prompt used 12.8x less, with all 183 hidden tests passing and comparable build times. It used less Claude than Claude Code on every task.

What happens when the engine gets a step wrong?

If the engine can't finish a step, it hands it back to Claude with what it tried. The step climbs a ladder: retries with the error output, a diagnosis, then a short fix brief from Claude, and a patch later still. Nothing touches your files until a change passes its checks.

Where does my code go?

Planning goes through your own Claude account. Execution runs on US and Canadian infrastructure with zero data retention.

How can I use the engine?

The desktop app is one way to use the engine, and the way you get it today. An API and an MCP server are planned, and neither is available yet.

Which languages does it support?

TypeScript, JavaScript and Python today, tested on open-source repos in both. More languages later.

When can I try it?

Join the waitlist and we'll write when early access opens. We don't have a date yet.

Your plan, carried further.

Join the waitlist for early access. We'll write when it opens, and not before.

Early access is private. We'll write when it opens.