KKurtel
Early access · Onboarding design partners

Coding agents
that grow from
junior to senior.

Launch state-of-the-art coding agents asynchronously in the cloud, grounded in a persistent memory of your codebase — routes, call graph, and team conventions, injected exactly when they're needed. Every run makes the next one sharper.

2,000+ installs of the Kurtel CLI and counting.

Live runs
ActiveAll
3agents active
+12 PATTERNS LEARNED TODAY
MonWedFriSun
agent-7f3aSenior82%
agent-2c9dMid46%
agent-b41eJunior14%

Built on the most capable coding models and the strongest isolation primitives:

OpenAI Codex
Claude Code
Firecracker
Kubernetes
gVisor
OpenTelemetry
OpenAI Codex
Claude Code
Firecracker
Kubernetes
gVisor
OpenTelemetry
OpenAI Codex
Claude Code
Firecracker
Kubernetes
gVisor
OpenTelemetry
Measured · controlled A/B
−34%
Token cost

Net cost per task, caching included.

−42%
Agent turns

7.2 → 4.2 — less blind exploration.

Parallel agents

Fan out across repos — no local machine.

0 tok
Codebase index

Routes, imports & call graph, extracted locally.

01 · Platform

Launch agents in the cloud, come back to shipped code.

Kurtel runs SOTA coding agents asynchronously in isolated cloud sandboxes. Fire off a task, close your laptop, review the result.

$ kurtel run --repo billing
→ sandbox booted (1.2s)
agent working…

One command to launch

From CLI, dashboard, or API. Point an agent at a repo and a task — it provisions its own environment and gets to work.

Reviewable results

Agents return clean diffs and pull requests with the reasoning trail attached — review feels like reviewing a teammate.

Parallel runs

Fan out dozens of agents across repos at once, each in its own sandbox.

Codex + Claude Code

Built on the strongest coding models available today.

Stateless sandboxes

Fresh, isolated environments provisioned per run.

API-first

Drive everything from the CLI, SDK, and webhooks.

02 · Codebase memory

A living map of your codebase, injected per task.

Kurtel extracts your routes, import graph, and call graph locally — then injects only what each task needs. Deterministic, zero-token, and aware of what a change would break.

store.tsapi.tsindexercapsulesync
Routes · 9 inventoried
GET/clients/:id/invoicesbilling.ts:6
POST/clients/:id/chargebilling.ts:10
GET/users/:idusers.ts:5
God nodes
store.ts12 edges
api.ts16 edges
Impact
change payment.ts 3 routes in blast radius

Deterministic & free

AST extraction runs locally — same input, same map, zero tokens.

Cross-lingual

Prompt in any language; aligned vectors resolve it to the right code.

Blast radius

Reverse call-graph shows exactly what a change breaks, before it ships.

03 · Living memory

Conventions learned from every merged PR.

Kurtel reads how your team actually ships — merges, review comments, reverts — and distills it into living patterns. Each one earns or loses fitness over time, so the memory sharpens instead of drifting.

Constraint★ pinned

Stripe payments must use embedded checkout — never a redirect.

billing#stripe#payment#checkout
Fitness94%
12 PR evidence
Pattern

New API routes return a typed Result<T>, never raw JSON.

apiroutes#endpoint#route#handler
Fitness81%
7 PR evidence
Decision

Background work goes through the queue in jobs/, not inline in handlers.

jobsapi#async#job#worker
Fitness76%
5 PR evidence
Darwinian fitness
Born

A merged PR or review comment reveals a convention.

Confirmed

Later PRs that follow it raise its fitness.

Contradicted

A PR that breaks it lowers fitness — dying patterns go quiet.

Silence by default

Only what's relevant, ever injected

A pattern only reaches the agent when the task touches its zone or triggers. Nothing relevant? Nothing injected — no context tax for freeloading lines.

Manual vs automated

Stop hand-maintaining a skills system.

Teams write rulebooks — AGENTS.md, .cursor/rules, skill files — then spend forever keeping them current. Kurtel grows and prunes that knowledge for you, from how you actually ship.

Before Kurtel

A skills system you babysit

  • Write the rulebook by hand — AGENTS.md, cursor rules, skill docs.
  • Update it every time the codebase shifts, or it quietly goes stale.
  • Outdated rules mislead the agent; no one knows which still apply.
  • The whole rulebook gets dumped into context every prompt — noisy and costly.

Hours every week · knowledge that rots

After Kurtel

Memory that maintains itself

  • Conventions learned automatically from your commits, merges and reviews.
  • Fitness decays what's wrong — dying rules go quiet on their own.
  • Only the rules relevant to the task get injected. Nothing else.
  • Adapts as you ship. You never open a rulebook again.

Zero upkeep · sharpens as you ship

03 · Security

Your code stays yours. Always.

Agents work on your code, but it never leaves the isolated sandbox it runs in. We built Kurtel so we can't see what we're not supposed to.

Sandbox boundary

Code enters, work happens, results come back — the boundary is never crossed by us.

SOC 2 compliant

Isolated sandboxes

Every run executes in its own hardened, network-isolated, SOC 2-compliant cloud sandbox. No shared state between customers.

Zero code access

Your source stays inside the sandbox. Kurtel operators never read, copy, or retain your code.

Ephemeral by design

Sandboxes are torn down after each run. Nothing lingers once the work is shipped back.

Scoped credentials

Agents get the minimum access a task needs, revoked the moment the run ends.

04 · Get involved

We're looking for design partners.

Kurtel is early. We're partnering with a handful of teams who want self-improving coding agents and are willing to shape them with us.

Design Partner

Build Kurtel with us

A small cohort of teams shaping the product. Hands-on access, a direct line to the founders, and pricing locked in early.

Apply as a partner
What you get
  • Early access to async cloud agents
  • Agents tuned to your codebase & conventions
  • Direct Slack channel with the founding team
  • Influence over the roadmap
  • Founding-partner pricing
Waitlist

Stay in the loop

Not ready to go hands-on yet? Get notified as we open access and share progress along the way.

Join the waitlist
What you get
  • Product updates as we ship
  • Early invite when access expands
  • Occasional notes on what we're learning