Task queue
Three workers pull from one queue. Every task keeps its own log, cost, retry count and cancel button, so a stuck job is visible instead of mysterious.
Self-hosted · v2026.11 · Launching Nov 10
A team of AI agents that runs on your machine.
A local dashboard for autonomous work: a task queue with live logs, nine specialist
agents, a paper-trading desk, and tooling that runs inside a sandbox. Download the
bundle, run two commands, open 127.0.0.1:8000. No cloud account — and
nothing leaves the box unless you ask it to.
TASK QUEUEWork in flight
3 workers
Queued12
Running3
Done128
Failed2
Summarize the research rungeneral · reading sources
Render the title sequencevideo · waits for approval
Rebalance the paper accountfinance · simulated
Check the store copywork · 3 sections
Ollama connected
0 requests left the box
What it does
The control room. The agents do the work.
Everything below runs on the machine you install it on. The dashboard is the control room; the agents work in the background.
Nine specialist agents
Design, art, audio, community, engineering, finance, study, video and classroom — each on its own port behind one shared sign-in, proxied through the dashboard.
Paper-trading desk
Real orders, fake money: the trading desk talks to Alpaca's sandbox endpoint, with position limits, a kill switch and backtests against your own history.
Sandboxed tools
Shell and Python run under bubblewrap with no network and a throwaway home. A command the model should not run is refused, not nudged with a string filter.
Local models first
Ollama on the machine is the default. Hosted models are optional, and their keys live in .secrets/ with owner-only permissions — never in the repo.
Built to be exposed carefully
Loopback binding by default, signed session cookies, CSRF origin checks, per-IP and global sign-in budgets, owner and collaborator roles.
Inside the box
One process. One journal. No broker.
One Python process, a SQLite journal, and a static dashboard. No container runtime, no message broker, no external queue.
- Sign-in and roles — the first owner is enrolled from the machine itself; invited accounts get one-time passwords.
- Permission queue — anything outward-facing (send, publish, push, run) waits for a click when approval is on.
- Analytics — throughput, durations and failure rates derived from the task history on read.
- Daylog and screencap — optional, bounded records of what ran and what was on screen.
- Discord bot — allowlisted user IDs only; an empty list means nobody.
- Modeler — Blender renders in a locked-down sandbox, if Blender is installed.
Install
# 1. unpack tar xzf ai-agent.tar.gz cd ai-agent # 2. create .venv + install deps ./setup.sh # 3. start the dashboard ./start_autonomous.sh start # open http://127.0.0.1:8000
The first sign-in asks for an email, the installation token from .secrets/dashboard_auth_token, and a password you may leave blank.
Windows: run the same commands inside WSL2. The sandbox and the service unit are Linux features.
Pricing
The app is free. The hosted seat is the product.
Download it, run it, keep it — no account, no expiry. Paid plans are the hosted workspace: we run it, update it and keep it online so you never touch a terminal.
Self-hosted
Your machine, your keys, no account.
$0
- Every agent and every tool
- MIT source, no telemetry
- Unlimited tasks, bounded only by your hardware
- Bring your own model keys
No card, no email, no trial clock
Pro
The same workspace, hosted and updated for you.
$10 / month
- Hosted workspace, always on
- One seat, all nine studio agents
- 500 tasks a month
- Automatic updates and backups
- Community support
Price locked while you stay subscribed
VIP
For people running this as a business.
$20 / month
- Everything in Pro, plus:
- Three seats and shared workspaces
- Priority queue and faster models
- API access and webhooks
- Advanced analytics and backtests
- Early access to new agents
- Private Discord, same-day support
Same 30-day refund, no questions
Founding members. Hosted workspaces open in waves from November 10. Reserving today holds your price for as long as you stay subscribed. Until your workspace is live you keep the full app, every update, and a full refund any time in the first 30 days.
Prices are in USD. Subscriptions renew monthly until cancelled. Cancel any time from the Stripe customer portal or by email — access continues to the end of the paid period. See Refunds, Terms, and Privacy.
Before you download
What you actually need.
This is a working local tool, not a hosted service. Being precise about what it is saves you a disappointing first ten minutes.
You need
Python 3.11 or newer, a POSIX shell, and about 200 MB of packages. Optional: Ollama for offline models, Chromium for the classroom browser, Blender for the modeler.
Paper money only
The trading desk ships pointed at Alpaca's sandbox. Live brokerage is not wired up by default, and nothing trades without your keys.
Yours to audit
Source included, MIT licensed, no telemetry. Checksums are in SHA256SUMS next to the archive.
Does it need a cloud account or an API key?
No. Ollama on your machine is the default model, and the dashboard works offline. If you add a hosted model later, its key stays in .secrets/ with owner-only permissions.
What does the hosted plan give me that the download does not?
We run the workspace, keep it updated, back it up and keep it online. The free download is the same software, on hardware you maintain. Hosted seats go out in waves from November 10.
Can I cancel or get a refund?
Yes. Cancel any time from the Stripe customer portal; access continues to the end of the period. Every charge is refundable for 30 days, no questions — see Refunds.
Does it work on Windows?
Inside WSL2, yes. The sandbox and the systemd unit are Linux features, so a native Windows install is not supported. macOS and Linux run it directly.
November 10
Founding access opens in
Reserve a seat before launch and the price stays yours while you remain subscribed.
0Days
00Hours
00Minutes
00Seconds
Reserve a founding seat
30-day refund · cancel any time