KaizenEvery change remembered.
Every mistake reversible.
Every session continuous.
Terminal-first AI coding agent. Web UI coming soon.
You work at the Helm.
Single binary. SQLite-only. No cloud backend. Works with any LLM provider.
Effort levels: low · medium · high · ludicrous · auto
You Work at the Helm
Helm is Kaizen's coordination layer. It classifies every request, builds a plan, and delegates to specialist agents - but only after confirming scope for anything non-trivial. You stay in control.
- Helm plans before acting - confirms scope for large changes
- Delegates read tasks to Scout, write tasks to Cody, reviews to Sage, docs to Scribe
- Helm never edits files directly - pure coordination
- Confirmation gate scales by scope: trivial → proceed, large → plan first
See It in Action
Watch Kaizen orchestrate agents, capture checkpoints, and deliver results.
Your Agent Remembers Everything
Kaizen persists decisions, discoveries, and bugfixes in SQLite across every session. FTS5 full-text search with staleness detection keeps your context fresh - no re-learning, no cold starts.
- FTS5 full-text search across all sessions
- Staleness detection flags outdated memories
- Four memory types: decision, bugfix, discovery, architecture
- No vector database required
Every File Write, Checkpointed
Before Kaizen writes or edits any file, it CAS-snapshots the previous state.
Made a mistake? kaizen undo shows a numbered list, diff preview, and y/N confirmation.
No AI mistake is permanent.
- Pre-write CAS snapshots on every file operation
- Branch undo history like git
- Delta compression: bsdiff with chain depth cap
- Named savepoints for manual checkpoints
- Automatic verification - after edits, Kaizen detects and runs your build and test commands, feeding failures back for a fix pass
Parallel Agents, Not Parallel Promises
Helm plans the work and spawns Scout, Cody, Sage, and Scribe as separate OS processes by default, each with its own TUI tab and independent scrollback. An in-process mode runs them as goroutines that share the orchestrator's database and LLM client when startup latency matters.
- Helm orchestrator with
create_plan/update_plantools - Scout (search), Cody (write), Sage (review), Scribe (docs) specialists
- Up to 8 agents, 3 concurrent by default (configurable)
- Per-agent TUI tabs - click to reactivate, × to close
- Effort levels: low · medium · high · ludicrous · auto - adjusts turn budgets and thoroughness
- Sage returns structured PASS / FAIL / CONDITIONAL verdicts, triggering Cody retry loops
- Plan-step checkpoints: every completed plan step creates a named savepoint automatically
- Dependency-aware planning enforces one writer per file - edits to the same file are grouped into a single Cody spawn and dependent batches are sequenced
- Delegate to external CLI agents: Helm can spawn Claude Code, Aider or any command, translating its output into the shared event stream
Understands Your Code Structurally
Kaizen builds a tree-sitter symbol index across your entire codebase. The repo_map ranks files by cross-reference score so agents know what matters before reading a single line.
- Tree-sitter symbol index across 18 languages
- repo_map ranks files by cross-reference score - agents prioritise what matters
- Symbol lookups cost ~20 tokens vs ~4,000 for a full file read
- Incremental indexing: auto-reindex on every write, background watcher for external changes
Context That Never Disappears
Other agents summarise old context and discard it. Kaizen replaces tool output with CAS hash
references - the agent can recall any masked content on demand.
Three-level strategy: CAS masking first, LLM summarisation only as fallback.
- CAS-backed masking: tool output stored by SHA-256 hash
recalltool retrieves any masked content instantly- Triggered at 75% of 200K context window
- LLM summarisation only as last resort
"Memory + checkpoints + smart routing: a focused model with full project context can outperform a frontier model with none."
Works With Your Entire Stack
Point Kaizen at any LLM provider - or keep it fully local.
Olla
Run inference locally with Olla, TensorFoundry's open-source LLM router. Zero cloud, zero latency.
Learn more →
FoundryOS
Pair with FoundryOS for enterprise GPU fleet management and load-aware routing across your entire inference infrastructure.
Learn more →Any API
Connect directly to Anthropic, OpenAI, or any OpenAI-compatible endpoint. Kaizen works wherever your LLM is.
Any provider supportedBuilt to Run Anywhere
Frequently Asked Questions
What is Kaizen?
Kaizen is a terminal-first AI coding agent with persistent cross-session memory, CAS-backed undo on every file write and multi-agent orchestration. It ships as a single binary with no cloud backend and works with Anthropic, OpenAI and any compatible LLM endpoint. The Helm orchestrator coordinates specialist subagents: Scout for search, Cody for writing, Sage for review and Scribe for documentation.
Who is Kaizen designed for?
Kaizen is designed for software engineers and teams who want an AI coding agent that runs entirely on their own infrastructure, remembers context across sessions and gives them a reliable undo path for every AI-driven file change. It suits developers who prefer the terminal, want cost control via per-session budget caps and need to run on air-gapped or secure environments.
When will Kaizen be available?
Kaizen is currently in Early Access. General Availability is targeted for Q4 2026. Join the waitlist to get priority early access and a direct feedback channel with the engineering team.
What makes Kaizen different from other AI coding agents?
Three features set Kaizen apart. First, cross-session SQLite memory with FTS5 full-text search means the agent never forgets decisions, bugfixes or architectural discoveries across sessions. Second, every file write is CAS-checkpointed before it happens, so kaizen undo shows a diff preview and numbered history with branch support. Third, Helm coordinates up to eight parallel specialist agents as separate OS processes with independent TUI tabs, dependency-aware planning and Sage review loops.
What LLM providers does Kaizen support?
Kaizen works with Anthropic, OpenAI and any OpenAI-compatible endpoint. You can also point it at a local inference backend through Olla, TensorFoundry's open-source LLM proxy, for fully air-gapped operation with zero cloud dependency.
How do I get early access to Kaizen?
Join the Kaizen waitlist to register for early access and shape the product through direct engineering feedback. Join the waitlist.
Get Early Access to Kaizen
Join the waitlist. Shape the product with direct engineering feedback.