Omakiten

A checkpoint layer for AI-assisted development.

Your AI agents forget everything when the session ends. Omakiten is the checkpoint they read before acting and write before leaving — so the next agent picks up from real state, not a blank slate.

The Problem

AI agents lose context between sessions. Different tools see different parts of your work. One agent rediscovers a bug another already fixed. You repeat the same explanation. They repeat the same mistake.

Omakiten gives your agents a shared checkpoint: tasks, decisions, errors, solutions, and handoff notes — in one place every agent can read and update.

Before and After

Before: You explain the project from scratch to every agent. They suggest changes that break your team's process. The same error comes back because the previous fix lived only in a chat transcript.

After: Each agent reads the checkpoint before acting. Workflow rules are enforced at the shared state layer. Search crosses sessions and projects. Handoff notes survive. You spend less time repeating context and more time building.

How It Works

You keep working with whichever AI tool you prefer. Omakiten runs locally and connects to your agents through MCP — a standard protocol that lets AI tools read and write shared state.

flowchart TB
    user([You])

    subgraph projectA[Project A]
        claudeA[Claude Code]
        opencodeA[OpenCode]
    end

    subgraph projectB[Project B]
        claudeB[Claude Code]
    end

    user --> claudeA
    user --> opencodeA
    user --> claudeB

    claudeA <-->|MCP| omakiten
    opencodeA <-->|MCP| omakiten
    claudeB <-->|MCP| omakiten

    cli[okt CLI] <-->|same state| omakiten
    tui[okt tui] <-->|same state| omakiten

    omakiten[(Omakiten<br/>local checkpoint)]
    db[(SQLite<br/>tasks · errors · solutions<br/>decisions · plans · handoffs)]

    omakiten <--> db

Claude Code, OpenCode, or any supported AI tool in any project — they all read from the same local checkpoint. No agent invents its own memory.

Install

Linux / macOS / WSL:

curl -fsSL https://raw.githubusercontent.com/This-Is-NPC/omakiten/master/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/This-Is-NPC/omakiten/master/install.ps1 | iex

The installer walks you through language, workflow preset, and which AI tools to connect. For headless installs (CI, Docker, dotfiles):

OKT_CLI_LANG=en OKT_AGENT_LANG="English" OKT_PRESET=omakase OKT_HARNESSES=claude-code,opencode \
  bash <(curl -fsSL https://raw.githubusercontent.com/This-Is-NPC/omakiten/master/install.sh)

Supported tools: claude-code, claude-desktop, codex, crush, github-copilot, opencode.

Both installers verify the downloaded release archive's SHA-256 against the goreleaser-published checksums.txt before extracting or running it — the same gate every in-app okt update applies. By default, the checksum trust root is pinned to https://github.com/This-Is-NPC/omakiten/... even when an artifact download mirror is configured for tests or private release infrastructure. A mirror can provide checksums.txt only with the explicit opt-in OKT_ALLOW_MIRROR_CHECKSUM=1 plus the separate OKT_CHECKSUM_BASE=<mirror> override; enabling that means you trust the mirror for the checksum authority. A checksum mismatch or unavailable pinned checksum aborts the install non-zero and does not install or replace the binary in INSTALL_DIR (any binary already on your PATH from a prior install is left untouched, not removed). This model assumes TLS to the checksum origin; release signing / SLSA provenance is not implemented yet.

Your First Project

# Register the current project
okt init --name MyProject --slug my-project

# Inspect the current state
okt list

# Open the visual board
okt tui

Once registered, tell your agent to start a session with /okt-start. It reads the checkpoint, surfaces what's pending, and suggests the next move. At the end, /okt-pause writes a handoff note — so the next session resumes from real state, not a blank slate.

flowchart LR
    start([New session])
    read["/okt-start\nRead checkpoint\ntasks · handoffs · errors"]
    work[Work on tasks\nwith guardrails]
    write[Write back state\nprogress · solutions · notes]
    pause["/okt-pause\nSave handoff"]
    next([Next session])

    start --> read --> work --> write --> pause --> next
    next -->|picks up here| read

    guard[[Guardrail violated?\nMove rejected]]
    work --> guard
    guard -->|fix and retry| work

Talking to Your Agent

Once connected, agents understand natural language:

Agents also respond to structured slash commands for more precise control. See the full command surface.

Work Presets

Each preset is a work discipline — a set of rules and guardrails that apply to both you and your agents.

Guards enforce the rules. If an agent tries to skip a required review step, the move is rejected with an explicit error — not a silent state change.

TUI

okt tui opens a terminal board with three zones: Tasks (board, table, graph, plans), Stats (per-model usage, logs), and Settings (runtime info, entity browser).

Outside a project, it opens a multi-project home. Pick a project and the shell cds into it when you exit.

Read the TUI guide.

Update and Uninstall

okt update --check    # check for updates without changing anything
okt update --yes      # download and swap the binary atomically

okt uninstall --yes             # remove binary and wrapper, keep data
okt uninstall --yes --purge     # remove everything, including data and config

Documentation

Master index: .docs/README.md

• Changelog
• Contributing
• License