v0.38 — Self-hosted • Open source • MIT

You have the idea.
Now you have the team.

humux runs a squad of agents — a lead, a developer, a researcher, a marketer, a QA — that pick up your most ambitious ideas, do the work, and report back. One Docker container, on your hardware, under your own keys.

Telegram — Team Chat
You: @kindral explore the meal-prep subscription space and tell me if it's worth building
kindral (lead): On it. Market, competitors, revenue models. Back within the hour.
— 20 minutes later —
kindral: 3 viable competitors, $2B TAM, brutal churn. The wedge is local-first sourcing. Worth a prototype — approve?
You: approved. break it down and ship it.
kindral: Milestone created: MVP. Dispatching 4 issues.
seniordev (dev): Took #12 — landing page. PR incoming.
marketeer (marketing): Took #13 — positioning and launch copy.
qa (qa): Watching #12. I'll review the PR when it lands.
01 — the squad

One agent for your life.
A squad for your work.

Most ideas die because one person cannot be five people. A squad fixes that. Create specialized agents — a lead who coordinates, a developer who ships PRs, a researcher who reads everything, a marketer, a QA — each with their own skills, tools, model, voice, memory and identity.

They coordinate through GitHub issues, each appearing as their own GitHub App with a distinct name and avatar. Work is dispatched, executed, reviewed and shipped without you in the loop — unless a task is labeled for a human. You stay the one with taste and the final word.

Per-agent Skills, tools, LLM model, voice, GitHub identity, memory
Shared Cross-agent memory store, group chat context, issue tracker
Dispatch Lead assigns via issues • agents pick up, execute, report
Fan-out Any agent spawns subagents in parallel, under a shared budget
humux — agent tree
humux jobs tree
kindral (lead) — milestone: MVP
├─ seniordev #12 landing page PR #41 open
│  └─ subagent: scaffolding done · 41s
├─ marketeer #13 positioning running
│  ├─ subagent: competitor scan done · 2m
│  └─ subagent: pricing survey running
├─ qa #14 test plan queued
└─ seniordev #15 checkout flow queued
✔ 2 running · 2 queued · budget 38% used

Every agent gets its own Telegram bot, GitHub App identity and LLM configuration.

lead

Breaks it down

Turns a one-line ambition into a milestone, issues and owners. Escalates only what needs you.

researcher

Reads everything

Deep research over the live web, with sources cited and a report you can actually check.

developer

Ships the code

Reads, writes, edits and runs code in a confined workspace. Opens PRs under its own identity.

marketer

Tells the story

Positioning, launch copy, landing pages, images — published as hosted artifacts.

qa

Keeps it honest

Reviews PRs, runs the tests, files what's broken before your users find it.

02 — every other day

Not a chatbot. An assistant that acts.

The squad is the headline. The foundation is an assistant that already runs your day: it reads your email, checks your calendar, sends messages on your behalf and runs scheduled work — from Telegram, WhatsApp or the CLI.

"Triage my inbox"

Reads your email over IMAP, surfaces only what matters, drafts the replies. Per-agent mailboxes with vault-secured credentials — keys never enter the model context.

> "3 emails need attention. The lease renewal has a Friday deadline — want me to draft a reply?"

"Morning briefing at 7am"

Scheduled work runs on cron. Weather, unread mail, today's calendar, reminders, whatever your squad shipped overnight — delivered as text and as voice.

> Daily at 7am: weather + email digest + calendar + overnight PRs + the topics you follow

"Schedule a meeting with Alex"

CalDAV against Google Calendar, iCloud or Fastmail. Checks availability, creates the event and emails the other person — from a single message.

> "Tuesday 2pm is free. Created 'Coffee with Alex' and emailed her a confirmation."

"Make a photo collage from these"

Send photos or data over Telegram or email. Your agent builds shareable mini web apps — games, collages, dashboards — hosted under /artifacts/ with a link you can pass to anyone.

> "Done! Here's your collage: https://your-humux.dev/artifacts/summer-2026/"
Telegram — clio
You: morning briefing
clio: Good morning! Here's your briefing:
☀️ 24°C, partly cloudy. Rain expected tomorrow after 3pm.
📧 18 unread — 2 need your attention:
• Lease renewal — reply deadline Friday
• Dentist rescheduled to July 15
📅 2 events: Design review 10am, Dentist 3pm
🔔 Overnight: seniordev merged PR #41, qa filed 2 issues
You: reschedule the dentist to next week, any afternoon slot
clio: Checking your calendar... Wednesday 2pm is open. I'll email the clinic and update your calendar.
✔ Calendar updated · Email sent to Dr. Russo's office
03 — how it works

One container. Your data.

A Python orchestrator plus battle-tested CLI tools. Himalaya (Rust) for email, CalDAV for calendars, Playwright for browsing, faster-whisper or Kokoro for voice, SQLite for storage. No external databases. No Kubernetes. No account to sign up for.

Your API keys, your infrastructure. Bring your own LLM key (Anthropic, OpenAI, DeepSeek, Google, Grok, OpenRouter) — per agent if you want a cheap model doing the grunt work.
Encrypted secrets vault. Two-tier encryption — credentials never enter the model context. Bitwarden import.
No telemetry, no cloud callbacks. Fully air-gappable. MIT licensed, every line on GitHub.
Offline-capable voice. Kokoro 82M for TTS, faster-whisper for STT. No cloud voice API needed.
bash
docker run -d \
--name humux \
-p 8000:8000 \
-v ./data:/app/data \
ghcr.io/mattmezza/humux:latest
Pulling from mattmezza/humux... done
✔ Container started
> Admin UI at http://localhost:8000
> Setup wizard guides you through API key, Telegram, accounts
> Then hire your first agent and give it something hard
04 — built in

Everything native, nothing bolted on

Email, calendar, contacts, voice, code, search — integrated at the core, not through third-party plugins. New capabilities are taught by writing a markdown file.

Messaging

Telegram (per-agent bots, voice, reactions, inline approvals), WhatsApp, CLI. Live web transcripts

Multi-agent

Agents with distinct skills, tools, voice, model and account bindings. Fan out subtasks to several at once

Memory

Four tiers: lexical, semantic, importance-based forgetting, deduplication

Markdown skills

Teach new capabilities by writing files — no code, no plugin registry

Voice

STT + TTS, local or remote. Voice messages in Telegram, voice briefings

GitHub

Webhook-driven agent turns on issues, PRs and comments. Per-agent GitHub App identity

Browser & search

Headless Playwright, Tavily or self-hosted SearXNG. Deep research with cited reports. Web artifacts

Admin UI

PWA-installable. Setup wizard, agent editor, token dashboard, prompt inspector, log streams, memory inspector

06 — open source

Built in the open

MIT licensed. Every line of code is on GitHub. No telemetry, no usage tracking, no cloud callbacks. Your squad answers to you and nobody else.