Open-source skeleton · MIT

An AI that stays on, pays attention, and speaks up.

Daimon is an open-source skeleton for persistent, self-aware AI agents. It runs on a VPS, remembers everything in git-tracked markdown, and gets sharper the longer it runs. Built for support, PR review, on-call, community, sales — anywhere a stream of input meets a human who needs help making sense of it.

See the flagship use case Browse use cases

View on GitHub — daimon-sh/daimon-skeleton

memory/journal/2026-04-15.md

---
when:   2026-04-15T09:42:11Z
from:   @ahmet # support lead
route:  support/bug-report
---

## What they said

> Exports from the programs page still 500 on staging v3.
> Third time this week.

## What I noticed

This is the 4th mention of export/500 in 11 days — same stack
trace signature as LIN-2814, which shipped a fix on Apr 9
but only to v2. The v3 route still points at the old
handler in apps/api/programs/export.ts:142.

## Draft reply

Thanks for flagging — we shipped a fix for this in v2 last week but
the v3 handler wasn't migrated. I've filed LIN-3001 and tagged
@engineering. I'll follow up here when it ships.

## Filed

- LIN-3001 # dedup of LIN-2814 scope  frequency: 4
- shadow: true # waiting for human send

This is Daimon thinking. Everything you see is a file in a git repo.

The idea

One AI. Always on. Never forgets.

Always on.

Daimon runs as a long-lived Claude Code session under systemd on a cheap VPS. It doesn't wake up when you ask it something — it was already listening.

Remembers in markdown.

Every conversation, learning, correction, and pattern lives in a git-tracked markdown file. No vector DB, no black box. The database is your repo.

Speaks up rarely.

Daimon doesn't spam. It uses noticing — absence, trend, connection — to say something only when it's worth saying. Scheduled dream passes consolidate, verify, and connect in the background.

The stack

Daimon is a skeleton. You pick the body.

A channel.

Telegram today; Intercom, Discord, Slack, GitHub, email tomorrow. Swap the input, keep the brain.

Routes.

Rules for what to do with each kind of input — spam, FAQ, bug, feature request, check-in, anomaly. Your rules, your routes.

Modules.

Pluggable extensions that track what you care about — health, finances, incidents, deals, anything. Each is a folder with a MODULE.md.

Shadow mode.

Every reply is a draft until a human says otherwise. Routes graduate to auto-reply only after they've proven themselves against real judgments.

How it works

The loop.

Same six steps, every run. The brain is identical across use cases — the only thing that changes is the channel and the routes.

01

Listen

Channel in, message captured, attributed, journaled.
02

Understand

Context lookup, disambiguation, history check.
03

Classify

Route the input per the rules you define.
04

Research

Read docs, code, past memory — whatever the route needs.
05

Draft

Internal analysis + proposed action. Shadow by default.
06

File & follow up

Commit memory, file tickets, schedule follow-ups.

Poll on a schedule. Consolidate during dream passes. Commit everything to git. That's the whole system.

Principles

Why it's shaped this way.

Markdown is the database. Git is the memory. The AI is the query engine.
Awareness, not analysis. Noticing patterns beats generating reports.
Ambient over direct. The thoughtful grandmother in the room speaks rarely and is worth more than the security camera.
Shadow first. Trust is earned per route, not granted up front.
Dormant is not dead. Archive is a real state; nothing is deleted.
Yours to own. Your install is a private repo. Your data never leaves your VPS.

Where it's working today

Built for real work.

Shipped · in production

AI Customer Support

Daimon reads every Intercom conversation, researches the answer in your codebase, drafts the reply, and files the bug. Running in production at Upcoach.

See how

Gallery · 12 patterns

Other use cases

PR review, on-call, community, sales ops, product analytics, compliance — the same loop, a different channel.

Browse use cases

Under the hood

No black box.

Claude Code, not an API.

Daimon is a persistent agent with full tool access — reads your codebase, queries your databases, uses your MCPs. Not a thin wrapper around a chat endpoint.

Cheap to run.

A $5–10/mo VPS and a Claude subscription. That's the whole operating cost.

Upgrade by git.

A new release is a git pull. Your data is protected by .gitattributes ours-merge — engine files update, your memory stays untouched.

Build your own.

Clone the skeleton See a real build

MIT licensed. Runs on your infra. Your data, your repo, your agent.