In staging — early access

The system that
improves the system.

A developer-first platform for AI Flows. Run multi-step workflows from your terminal, route LLM calls through the model that's right for the job, and keep your code on your machine — all with the same observability and cost discipline you'd expect from any serious tooling.

No credit card required · 1,000 free credits to start · ~3 min from sign-up to first chat

kaizen — develop-feature

# A real session — no marketing magic.

› Read main.rs and explain what it does

● file_read src/main.rs (327 lines, 14ms)

The entrypoint wires up tracing, loads config, runs migrations, registers the LLM providers, and starts the Axum router on $PORT...

› /flow run develop-feature jira_ticket=KC-170

● understand → implement → test → review

model: claude-sonnet-4-5 · spent: 142 CC · duration: 2m 18s

Built in RustOIDC via KeycloakMulti-provider LLM routingCapability-based file safetyCross-platform: Linux · macOS · Windows

Tooling that respects your codebase.

Three commitments. Everything else follows.

Multi-step Flows, defined in YAML

Compose agentic steps, scripted steps, and human approval gates into reusable workflows. Run them ad hoc from the TUI or from CI. Resume, retry, observe — none of the glue is your problem.

/flow open develop-feature
/flow inputs jira_ticket=KC-170
/flow run

Right model for the job

Backend routes LLM calls to the model that fits the task — Claude for nuance, GPT for breadth, Llama or DeepSeek for cost-sensitive bulk. Pick per-step, override per-Flow, or let the planner decide.

/models set foundry:claude-sonnet-4-5
/models list  # 30+ available

Your code stays on your machine

Tools execute locally. The backend brokers LLM requests but never sees your files. Capability-based safety means "read grants write" — nothing edits a file you didn’t already inspect.

/file structure src/main.rs
/file patch  src/main.rs --find ... --replace ...

From visit to first chat in three minutes.

We tested it. Average new account is chatting in under 180 seconds.

01
Sign up
Email + password. Email verification. 1,000 free credits in your account before you finish refilling your coffee.
02
Pick a plan
Free tier covers everyday hacking. Upgrade to Pro when you need more headroom — Stripe Checkout, your usual card, no awkward sales call.
03
Install the CLI and go
One command, OS-detected, checksum-verified. Authenticate with the same SSO you just signed up with — no API keys to copy-paste.

Compute Credits

One unit of cost. Across every model.

CC is our universal cost unit. Smaller models burn fewer credits per token; larger ones burn more. Switch models freely — the number on screen stays comparable, so you never have to relearn pricing.

Free

£0/month

no card required

1,000 CC / month

Plenty for hacking and trying things out.

Pro

Team

TBD/month

monthly per seat · roadmap

Pooled across seats

Shared Flows, shared cost visibility.

Express interest

Pricing for paid plans is finalising — Pro and Team will land in test mode first so we can dial in the right CC/£ ratio without surprises.

Install in seconds

One command. No package manager dance.

Auto-detects your OS and architecture. Verifies the checksum. Drops kaizen on your PATH.

curl -fsSL https://install.stage.code.oxfordkaizen.com | bash

Cautious about curl ... | bash? Read what the installer does first.

Questions we hear most.

Do my source files leave my machine?+

No. The CLI runs locally; tools (file reads, edits, shell commands) execute on your machine. The backend brokers LLM requests but never receives your project files. Only the messages you send to the model — and any code you choose to include in those messages — go over the wire.

Which models can I use?+

Several. The backend routes to whichever provider is configured for your account: a Claude family (Haiku, Sonnet, Opus), a GPT family, Llama, DeepSeek, Mistral, Kimi — all under one billing unit (CC). You pick per-Flow, per-step, or override at runtime.

How does pricing work?+

Free tier on signup gives you 1,000 CC — enough for hacking and small Flows. Beyond that, paid plans top up your monthly CC budget. CC is normalised across providers so switching models doesn't mean relearning prices.

Is the source code open?+

Not yet. The release binaries are publicly downloadable from our GitLab Generic Package Registry, but the source is currently private. We're evaluating an open-source pivot — if it matters to you, tell us.

Can I run it without an account?+

Yes — the CLI works against a local Ollama instance with zero account, zero billing. You lose multi-provider routing, telemetry, and the cloud-hosted Flow runner, but the core agentic loop still works.

What happens if I hit my CC limit?+

Currently: requests fail with a clear error message and the in-progress Flow stops gracefully. Coming soon: real-time CC tracking in the TUI plus optional auto-recharge for paid plans.

The system that improves the system.