liv · personal
Liv
A blue butterfly for life. Personal chief of staff, always around.
- Household and health
- Cadence and commitments
- Journal and memory
Open source · running daily · one operator
Every layer of computing has an operating system.
The top one never got one.
AI made producing work cheap. It did not make following it cheap. The bottleneck is what one person can hold in their head.
- 2
- named agents
- 1
- thread per session
- 30
- threads on a busy day
- 0
- vendors you cannot leave
01 — The layers
- L0Hardware
The machine.
Solved - L1Firmware
Operates the machine.
Solved - L2Software
Operates on the machine.
Solved - L3Humanware
You, and everything you built to run yourself.
Unsolved
Humanware is the layer. HumanwareOS is the project that operates it.
02 — The inversion
We used to work for the tools.
For most of computing the deal was that you learn the language, the terminal, the metaphor. Now the computers speak ours.
Which should have been the end of it. Instead, directing work turned out to be harder than doing it: more decisions, more context at once, more threads than one person can track.
Conductor and Superset handled parallel coding well, and nothing past it. Paperclip had the best mental model going — agents as skills, permissions, and context, routed like an org chart — but spun up too much too fast. A workflow in Notion was all friction.
The coordination layer already existed.
Slack spent a decade refining how people actually work. Threads. Notifications that mean something. Channels as context. Integrations with anything.
So the interface is not new. The agents move in behind it.
humanware
ThreadsLater
# building# life# works
# building🔄
Ariel
the noise logger needs a real mic driver. take it.
Liv
On it. Branch app/noise-logger, INMP441 over I²S.
Liv
Driver reads clean at 16 kHz. Calibration curve is in the PR.
noise-logger · PR #74 files · driver, calibration, test fixture
PushedOne thread. One session. The whole exchange is the interface.
03 — Why this is opinionated
One line in someone else's config made me stop and build my own.
“Never ask a human to do what an agent can do.”
Found in an agent config · the inverse of Agent Smith's line
That premiseThe human is overhead. Route around them.
This oneThe human is the point. Stay close to the decisions, the sequencing, the fidelity.
Both systems can run the same models on the same machine. Only one of them is trying to need you less every week.
04 — Sessions
Every thread is a session. Every session wears four things.
You read a thread's whole state off the reactions, without opening it.
StatusIn process
AgentLiv
ModelOpus
HarnessClaude Code
The status language — five states, nothing else
In process
On the agent, actively moving.
Clarify
On you. You owe an answer.
Act
On you. Only your hands can do it.
Scheduled
Comes back at a set time.
Done
You confirmed the outcome.
No generic "blocked." Dated means scheduled; undated means still the agent's problem.
05 — The cast
The agents are animals on purpose.
They are not humans, and they do not pretend to be. They are personified helpers with stable identities, real personalities, and different responsibilities. The animal makes each role recognizable at a glance; the system keeps the human in charge.
max · works
Max
An orange-red fox for work. CEO of the companies and projects.
- Strategy and products
- Content and priorities
- Tradeoffs, not accumulation
Together they form the HumanwareOS mark. One half holds life; the other holds work. They face each other because coordination matters more than autonomy. The space between them belongs to the human.
Two agents is the right number, because skills let each wear many hats without turning the system into a crowd.
06 — The complete stack
The complete stack. No lock-in.
| Slot | Running today |
|---|---|
| surface | Slacklaptop · phone · web |
| transport | Tailscaleno ports open to the internet |
| host | Mac minihardware you own, in your house |
| runtime | OpenClawone thread, one session |
| harness | Claude Code · Codex · OpenCode · PiSwappable |
| model | Anthropic · OpenAI · local · OpenRouterSwappable |
| memory | Markdown in gitreadable without this system running |
| secrets | Dopplernever on disk in the clear |
Every swappable row has already been swapped at least once.
07 — Getting started
Start with the agent. Let it build the rest with you.
Configure one durable agent locally, then give it the machine and the guide. It should drive every safe terminal, API, and UI step. You step in for identity, consent, and secrets.
-
1
Shape the agentAgent + human
Choose its remit, voice, boundaries, and first strategy.
-
2
Install the runtimeAgent drives
OpenClaw connects the model, files, tools, and background service.
-
3
Secure the keysHuman authenticates
Doppler holds credentials. Values never enter chat, files, logs, or git.
-
4
Connect SlackHuman approves
Start free with Socket Mode—no public URL or tunnel required.
-
5
Prove the handoffAgent verifies
Slack → runtime → agent → tools → reply in the same thread.
Core setup
Mac · model access · Slack Free · Doppler · GitHub
Full system
Follow the complete guide
Cloudflare · your domain · backups · local capture
Buzz is the intended bundled path once it is reliable enough. Slack is the pragmatic starting surface today.
08 — What it costs
FrameworkFree
Open source. Fork it, keep the spine, change the parts that are about you.
$0 / mo
ProductsAlready paying
Slack for the history, a model subscription. Nothing new is introduced here.
— / mo
Running itModel spend + a machine
Token cost, plus a Mac mini once. Local models move work out of this column.
— / mo
Two figures are blank because they have not been measured yet. They will be filled in, not estimated.
Humans care about humans.
We watch people run 26 miles slowly with a car right there. We cheer Magnus Carlsen when the engine beats every human alive. The struggle is the point — and agents doing the work does not change it.