Your AI starts every chat from zero. Kaisho fixes that with two memory layers: the live one — tasks, customers and clocked hours your AI reads through MCP — and the durable one, a Markdown knowledge base where every session's gotchas, decisions and snippets accumulate. Schedule cron agents to brief you Monday, draft Friday's invoice from unbilled hours, or write learnings back to the KB. Plug into Claude Code, Cursor and Claude Desktop. Plain text or SQL. Local-first, always yours.
Want it on Claude Code, Cursor, or Claude Desktop? See the MCP setup (included on Companion & Pro)
Install the desktop app, point your AI at it, and let it remember.
Grab the desktop app for macOS (Apple Silicon), Windows, or Linux from GitHub Releases. Or install via pip install kaisho && kai serve. Tasks, time, customers, notes and the KB all live in plain text on your machine.
Add the Kaisho MCP server to Claude Code, Cursor, or Claude Desktop with a one-line config. Your AI can now read and write your tasks, look up a customer, log time, query the KB, and run shell commands you whitelist — without leaving the editor.
Schedule cron-AI agents to do the boring parts: a Monday-morning briefing on last week, a Friday invoice draft, a daily KB digest. They run locally on your hardware (Local) or on our hosted token quota (Companion / Pro).
Kaisho is what your AI companion needs to be useful day-to-day: durable memory, scheduled runs, MCP-native integration with the editor you already use, and a real task / time / customer tool underneath.
Tasks, time entries, customers, notes and a knowledge base — all in plain text, all on your machine, all queryable by your AI through MCP. Sessions in Claude Code or Cursor stop being amnesiac.
Cron-AI jobs run on your schedule, using the model and API key of your choice. A Monday-morning briefing on last week. A Friday invoice draft. A nightly KB digest. The agent reads your Kaisho data, does the work, leaves an artefact.
First-class MCP server. One-line setup for Claude Code, Cursor and Claude Desktop. Premium MCP integrations on Pro: Linear, GitHub Projects, Google Calendar, Slack — so your AI can read tasks from Linear, file a PR via the GitHub MCP, and book the time in Kaisho in a single chain.
Local runs entirely on your machine — no account, no cloud, BYO API key. Companion and Pro add a hosted token quota and cross-device sync, but the source of truth stays your plain-text directory.
Claude, GPT, Gemini, local Ollama, LM Studio, any OpenAI-compatible endpoint. Local is BYO key. Companion and Pro include a frontier-model quota so you can stop juggling keys — or keep your own for sensitive work.
Under the AI layer: drag-and-drop kanban, start/stop timers, per-customer contracts and budgets, unbilled-entry preview and CSV export for your accountant, weekly/monthly dashboards. The productivity tool is what gives the AI something to remember.
Every action feeds the next. Kaisho turns time tracking into context, context into automation, and automation back into knowledge — so the loop closes instead of leaking.
Book time, close a task, update a customer budget — from the CLI, the editor, or by asking your AI in plain English.
$ kai book 30m "RIDING BYTES" --tag dev
Ask the advisor anything grounded in real numbers. "Who's closest to budget?" pulls hours from clocks. "What's unbilled from May?" sums by customer.
advisor › who is closest to budget? RIDING BYTES 147h / 200h · 73%
Cron-AI jobs run the boring parts on your schedule: Monday-morning briefing of last week, Friday invoice draft from unbilled clocks, nightly KB digest.
@cron friday 17:00 draft invoice from unbilled
The advisor writes findings back to your knowledge base: customer quirks, deploy gotchas, code patterns. The next session starts pre-briefed, not from zero.
+ kb/customers/RIDING-BYTES/billing.md
capture → track — the loop closes
OpenClaw and Nous Research's Hermes Agent pioneered the autonomous-agent pattern: cron-scheduled jobs, persistent skills, a personal context the model reads on every run. Kaisho ships the same primitives — scheduled agents, a SKILLS/ directory, SOUL.md and USER.md identity files injected at runtime — bounded to your productivity domain. The agent reads your tasks, books your time, drafts your invoice, writes back to your KB. It doesn't drive your browser, run shell commands you didn't whitelist, or call APIs you didn't configure. Same agent loop, smaller blast radius.
Schedule prompts the way OpenClaw schedules system jobs and Hermes schedules natural-language cron. Monday-morning brief on last week, Friday invoice draft from unbilled clocks, nightly KB digest, weekly customer pulse. Define them in YAML, swap the model with one line.
@cron mon 08:30 brief last week's clocks + open tasks
Each skill is one Markdown file in your profile's SKILLS/ folder. Kaisho pulls the relevant ones into context per run, the way OpenClaw's ClawHub adds marketplace skills and Hermes auto-generates new ones. You ship your own or borrow from the community.
SKILLS/budget-review.md SKILLS/import-github-tickets.md SKILLS/week-summary.md
Two files per profile that load on every advisor and cron run. SOUL.md sets the tone, USER.md sets the facts. Same persistent-context pattern OpenClaw and Hermes use, made auditable as plain Markdown you edit yourself.
USER.md who you are SOUL.md how the agent talks SKILLS/* what it can do
OpenClaw alternative · Hermes Agent alternative · bounded by design
Tasks close. Clocks roll over. The knowledge base accumulates. Insights from one AI session become MCP-readable context for the next — written once in plain Markdown, addressable by path, searchable by your AI through the same MCP server it uses for everything else.
KAISHO
kb/auth/jwt-rotation-gotcha.md
Why our token refresh fails silently when clock skew is greater than 30s, and the one-line fix.
kb/customers/RIDING-BYTES/billing-quirks.md
VAT exemption rules per project type, accountant's preferred invoice format, who to CC on monthly statements.
kb/ops/cron-scheduling-notes.md
Why APScheduler misfires on DST boundaries and the timezone gotcha that bit us in March.
kb/kaisho/release-checklist.md
Bump four version files, regenerate CHANGELOG, tag, push, flip draft to published.
All entries are plain Markdown. Your AI reads them, writes to them, and never asks you the same question twice.
The web UI is optional. Every feature in Kaisho is a kai subcommand — and every kai subcommand is also an MCP tool. Start timers, query your KB, book time, run an agent — from the shell, from the editor, or from your AI session. Same surface, three entry points.
kaisho-mode is an optional Emacs package that keeps your clock visible in the mode line and lets you start, stop, and resume timers without switching windows. It connects to the running kaisho backend via WebSocket — so clocks started in the web UI appear in Emacs instantly, and vice versa.
Works with any backend. No org-mode required. All operations go through the kai CLI and REST API.
Resume last clock on next toggle.
Reconnects automatically if the backend restarts.
Kaisho is backend-agnostic. Keep everything in human-readable plain text you can version-control, in portable JSON, or in a SQL database for concurrent, team-scale access. The web UI, CLI, MCP server and mobile app work identically across all of them.
Org-mode or Markdown files on your disk. Inspectable, diff-able, Git-friendly.
One structured document set. Easy to back up, script against, and integrate.
SQLite or Postgres for concurrent access and scale. Built for teams and professional deployments.
The plain-text backends are human-readable end to end:
Scroll through the desktop and mobile views. Click any screenshot to enlarge.
The mobile PWA gives you the same data your editor talks to: tasks, time, customers, dashboards. Chat to the advisor on the go — same model, same memory, same MCP context as on your laptop. Included on Companion and Pro.
First-class MCP for Claude Code, Cursor and Claude Desktop. Any OpenAI-compatible model on the back end. Open storage — plain text, JSON or SQL — you can pipe to anything else. No lock-in, on either side.
When Kaisho is running, every MCP-compatible client can reach your tasks, clocks, customers and KB at a local HTTP endpoint. The token lives on your machine; nothing leaves it.
claude mcp add --transport http kaisho \
http://localhost:8765/mcp/ \
--header "Authorization: Bearer <your-token>""Open in Kaisho" needs the desktop app (v2.3.1 or later) installed. If nothing happens when you click it, install Kaisho first, then come back.
Local is the open-source desktop. Companion and Pro add a hosted memory layer, scheduled AI runs, and MCP integrations with the editor of your choice.
Rolling Kaisho out to a team? Talk to us about multi-seat plans and a shared knowledge base.
No credit card for Local. Companion and Pro are personal-card priced — no purchase approval needed.
add_task, book_time, list_customers, search_knowledge and so on. Add one line to your Claude Code or Cursor config and your AI session can read and write your Kaisho data without copy-paste.Kaisho is MIT-licensed. Run the desktop app, the MCP server, and the sync layer entirely on your own hardware — or pay for the hosted token quota and skip the setup. Either way, the source is on GitHub, the data is yours, and the architecture works without us.
In Japanese martial arts, kaisho (開手) means “open hand” — the position of readiness. The mark is an open bracket with three lines: the bracket is the palm, the lines are what you carry — tasks, hours, ideas.
The bracket is not closed. It accepts new entries, holds what matters, and expands when you need it to. Structure without rigidity.