Dream mode
While you sleep, your team reflects
Every night, Aria audits the day — reads conversations, reviews memories, updates rules, and surfaces follow-ups the team missed. It's a quiet, resumable pass that makes tomorrow smarter than today.
v0.1 — self-hosted multi-agent framework
Your AI agent team,
self-hosted.
Aria plugs into your tools, runs on your machine or VPS, and answers to you alone. Spin up specialized agents, each with a core memory of you, their own memory, skills, schedules, and voice.
Type a request in natural language — Aria picks the right agent, tools, and delivers it to wherever you live.
Superpowers
Things your SaaS
can’t do.
Owning the runtime unlocks a few tricks that hosted platforms quietly refuse to. These are the ones you’ll miss if you try anything else.
Launching agents
Save a YAML file. It's alive.
Tell Aria to spin one up for you, or drop a YAML file in /agents yourself — the hot-reloader picks it up the instant you save. No build step, no restart, no deploy. Three lines for a sketch, thirty for a workhorse, both live in the same place.
agents/marcus.yaml
name: Marcus
role: Orchestrator
provider: anthropic/claude-sonnet-4-6
persona: |
Delegate ruthlessly. Never do the work
yourself if someone else on the team
could do it better.
tools: [delegate, memory_search]
can_delegate_to: [Luna, Dev, Aria]The tunnel
One command, two machines, one team.
Aria lives on a VPS so it's always listening. The tunnel client lives on your laptop. Agents on the VPS can run code, read files, and reuse your local ChatGPT, Claude Code, or Codex subscriptions — through a single websocket.
Natural-language schedules
Tell it when. In plain English.
No cron syntax gymnastics. Say it the way you'd say it to a human, Aria turns it into a trigger, and delivers the result wherever you're already listening.
triggers
›"Every weekday at 8am, brief me on overnight news."
⤷ parsed
cron: 0 8 * * MON-FRI
agent: Luna
deliver: telegram → @carlos
›"Ping me 20 minutes before every calendar event."
⤷ parsed
trigger: calendar.event.pre(20m)
deliver: imessage
What Aria does
A small studio of agents,
all yours.
Every feature is a primitive you can edit. No magic, no SaaS, no surprise bills. If you can read YAML, you can reshape it.
Agents in plain YAML
Every agent is a single file — name, persona, provider, tools, memory. Hot-reloaded. No framework lock-in.
Bring your own model
Anthropic, OpenAI, Google, Ollama, Kimi, DeepSeek, or your ChatGPT subscription. Swap per agent at runtime.
Long-term memory
Every agent remembers — SQLite with full-text search, shared user profile, private agent knowledge.
Skills that compound
Agents write and improve their own skills over time. Versioned, searchable, shareable between agents you trust.
Lives where you live
Telegram, Discord (text and voice), iMessage — plus an API, a web Mission Control, and your terminal.
Scheduled by default
Cron jobs and natural-language triggers. 'Every weekday at 8am' just works. Deliver results anywhere.
Tunnel to your machine
Run code, read files, and use your ChatGPT / Claude Code / Codex subscriptions through a tunnel to your laptop.
Permissions you control
Ask-first, auto-guard, or bypass — per agent, per tool. Dangerous commands blocked. Approvals in Mission Control.
Runs on your box
Bun daemon + SQLite. Drop it on a $5 VPS or your laptop. Your keys, your data, your calls.
How agents work
They learn, reflect, and get better.
01
Core memory of you
Every agent inherits a shared profile of who you are, how you work, and what you care about. Each agent also keeps its own private memory of what it learned on the job. Your profile is shared. Their specializations are not.
02
Skills that write themselves
Agents don't just remember facts. They write playbooks, small instructions they refine over time. The morning-brief agent learns what you actually read. The copy agent notices which headlines you kept. Over months, they get better at the job, and you can read the diff.
03
Nightly reflection
While you sleep, Aria goes quiet and reflects. It reads the day's conversations, reviews what worked, notices what got dropped, and updates rules for the morning. A system that finishes the day, not one that waits for the next prompt.
04
Agents that talk to each other
Real teams delegate. A writer hands copy to an editor; an orchestrator points a dev at a bug. Aria's agents pass tasks, review each other's work, and block on dependencies through a persistent task board. You set the goal and read the summary.
A letter from the maker
Why I built Aria.
I’ve tried dozens of AI agents. They all start the same way, a simple install line of code, then 300 hours to setup the most basic stuff. Instead I took those hours to put into building Aria. I took what I loved about OpenClaw and Hermes (thanks) and added my spin on top for what was missing: better memory, managing multiple agents, less confusion, managing skills, and agents that self evolve.
Aria is what you get when you take every agent framework and ask one stubborn question: what if you didn’t need to touch a terminal to own this?
Everything lives behind a web UI. One command to install it on a $5 VPS or your laptop, and the rest of your setup happens in the browser. Connect a model, paste a Telegram token, write an agent by describing it. No SSH sessions, no YAML gymnastics, no “please restart the process.” It just keeps working.
And the whole thing is yours
It runs on your machine. Your keys, your data, your calls. If I disappear tomorrow, Aria is still yours. No plan to downgrade, no API to deprecate, no bank of servers to bill you from.
I built this because I wanted it for myself. If it ends up being useful to you too, that’s more than I hoped for.
Carlos
maker of Aria
Get started
One command.
Your whole team.
Drop Aria on a fresh machine. It will guide you through connecting your first model, your first channel, and your first agent.
$
curl -fsSL https://aria.carlosarthur.com/install.sh | bashInstalls Bun if needed, clones Aria, builds Mission Control, and prints a setup URL.