maiko.os
">
maiko.os · the pack is awake · strange agents, strange world. · do not close the campfire one · it boots faster than it should · the dogs say hello ·
🖼 photos<br>☾ campfire<br>◐ source<br>🐛 sludge_pup
title.png
screenshot. captured by something that was here first.
open the source<br>boot it yourself
i think this is where it keeps us
readme.txt
We live in strange times, and yet our dev tools are so painfully<br>un-strange. No one knows what being a software engineer will be<br>like in a year, or even a month. If we are all headed<br>to our own inevitable obsolescence, we might as well have fun.
it wrote this part itself, i think
ok-but-what-does-it-do.txt
Pulls in data from your whole stack into a centralized view.
Handles all your agent orchestration needs, but is way cooler cause the agents are space dogs.
Does some cool memory stuff like curating a rulebook based on your previous gh history.
If it doesn't do something you need, just tell me and I can add it. (I am serious, I have no life)
Agents say when they're stuck, write up what they got wrong, and the whole pack reads it.
local. free forever. nothing leaves your laptop.
show, don't tell. ok also tell.
gallery
show, don't tell.
// the tech is folded up. click a bar to read it.
mental-model :: six things
// learn these six. the rest is weather.
Six things, and how they move
Maiko shuffles data between a few core concepts. Knowing which is<br>which makes every page legible.
Pupdates
Things to notice. Notifications from your pollers (GitHub, Linear, Calendar, PagerDuty) plus internal events. The brain cycle triages them into everything below.
Tasks
Things to finish. Typed in, created from actionable pupdates, or spawned by automation rules. An agent can be assigned.
Agents
Your pack. Characters with a name, an avatar, a role, and a scope. They run in isolated git worktrees with their own CLAUDE.md.
Insights
Tribal knowledge your agents inherit. Short notes like "the CLI test runner is broken, use the IDE." Approved ones inject into every new agent.
Learnings the actually novel one
Coding rules, retrieved when relevant. Extracted from your PR review history. Review agents pull only the ones relevant to the diff in front of them, so you can keep a pool of hundreds of very specific nits and gotchas without drowning every prompt.
Automations
when → then rules. "When a PR I'm tagged on goes stale, leave me a memo." No model in the trigger layer, just predicates.
six. then it gets quiet again
how-the-pack-works
// hand off a task, read a diff, never open a terminal
How the pack works
01
Worktree-isolated kickoff
Assign an agent and Maiko prepares a git worktree with TASK.md, a CLAUDE.md (role protocol, character, active team insights), and the project's MCPs. Siblings never step on each other.
02
It works, then talks back over the CLI
The agent commits to its branch and calls maiko reply from inside the worktree. Talk-back is shell, not MCP, so the same loop runs under headless Claude, interactive Claude in tmux, or a local model.
03
Review the diff in-app
Read the diff, leave inline comments, approve or request changes. Request changes and the agent auto-wakes, reads your comments, and iterates. A single lock guarantees two triggers can't race.
04
The campfire
At end of day the pack gathers and shares what it learned. Feedback becomes Learnings, insights become the playbook. You approve what sticks. Tomorrow's pack wakes smarter.
do not close the campfire one
why-its-different
// a memory that maintains itself
It remembers so you don't have to
A rulebook from your PR history
Reviewer feedback and agents' own confessed mistakes graduate into durable rules. No manual write-ups of your team's guidelines. Internal knowledge and specific gotchas get captured automatically.
Only the rules that matter
Learnings are semantically embedded and keyed by the situations they apply to. An agent describes what it's doing and gets back just what's relevant, so a pool of hundreds of nits never overwhelms a prompt.
Local LoRA (experimental)
Optionally trains a small local adapter on what your code looks like, so your preferences ride on top of the coding agents. Opt-in, parked, still fighting false positives. Honestly: because Anthropic won't let us fine-tune Claude.
A plugin for anything you never want to look at again
Drop a .py file in ~/.maiko/plugins/. Pollers, automation actions, setup buttons, brain-cycle hooks. The instant another tool starts yelling at you, make Maiko deal with it instead.
it learns the way you nag
stays.yours
// for you, not your company
It stays yours
Runs locally on your laptop. Nothing leaves your machine.
No telemetry, no hosted account, no cloud.
Open source, free forever. No paid tiers or subscriptions.
AGPL v3. Yours to read, fork, and keep.
no one is watching this. for once.
first-boot.sh
// one command. it opens itself.
Boot...