Build a team of AI agents that works while you sleep.
Connect your tools, tell your agents what to do once, and they handle the recurring work on schedules and events. Anything that touches the outside world waits for your approval.
One workspace. Packs of agents. Standing tasks.
Four things to understand, and you understand the whole product. This is the real structure of the app, not a metaphor.
- Workspace
- Your space, for a team or a company. You can have more than one, and each keeps its own packs, connections, and wallet.
- Pack
- A specialized team of agents inside a workspace. Content Studio, Inbox Concierge, Growth Strategist, Design Studio. Run more than one, side by side.
- Agents
- The specialists inside a pack. Each has a role, instructions, and memory. Start from predefined agents or define your own. One of them is the Chief: the agent you talk to, which delegates to the rest.
- Tasks
- Standing jobs you hand a pack: recurring, scheduled, or triggered by an event in a connected tool. A pack can run several at once.
- Northwind Studioworkspace
- Content Studiopack
- editorchieftalks to you, delegates
- content-researcheragent
- copywriteragent
- Weekly poststask · nightly 11:30 PM
- Repurpose updatestask · Mondays 6:00
- Inbox Conciergepack
- conciergechief
- classifieragent
- drafteragent
- Overnight triagetask · on new email
You talk to one agent. It runs the rest.
Every pack has a Chief. You brief it in plain language; it breaks the work down and delegates to the right specialist. You never manage agents one by one.
It delegates, visibly
Each hand-off is shown as it happens: who was asked, for what, and what came back. The Chief assembles the pieces and answers for the result.
When a capability is missing, it says so
The Chief does not fake a skill it lacks. It offers two ways forward: bring in a predefined specialist, or train a new agent on your instructions and voice.
New agents are composed, reviewed, then added
A trained agent is defined and validated first, and you review it before it joins the pack. Nothing rewrites itself behind your back.
Start with specialists. Define your own.
Four packs are available today, each a working crew with a Chief. This is the full current roster, nothing padded.
Connect, instruct, and let them listen.
Set it up once. After that, packs act on schedules and on events from your tools, and every real-world action still stops at the gate.
aConnect your tools
Slack, Linear, and email are first-party today; other systems attach through connections. Agents get access only to what you connect.
bInstruct them once
Rules and workflows live with the pack. Say it one time; they keep to it on every run.
cLet them listen
Events in connected tools can start a run.
The gate holds everywhere. Whether a run starts from your message, a schedule, or an event at 3 AM, anything that would leave the workspace waits for your yes.
Four jobs, one per pack, ready tonight.
Each shipped pack owns a job you can hand over now. Below the rows, the roadmap trio is labeled as roadmap; nothing there is sold as shipping.
The inbox that filled up overnight
Every thread sorted by intent and urgency, one drafted reply each in your voice, queued for your review. Follow-ups included: the warm decline, the third nudge.
Inbox ConciergeA week of posts from one announcement
A changelog or update becomes a scheduled series: researched angle, drafted copy, edited, held at the gate until you approve the schedule.
Content StudioThe launch plan you keep not writing
Positioning, ICP including who it is not for, competitor profiles, and a prioritized plan, built strictly from materials you paste. It asks for what is missing instead of inventing.
Growth StrategistA design direction you can hand to a builder
Two or three named directions to pick from, then your pick expanded into a full layout spec: typography, color, spacing, per-section wireframes, with copy to match.
Design StudioOne night, start to finish.
A single real run, end to end: you brief at night, the pack works through the steps, and you approve the real actions in the morning.
You brief the Chief and close the laptop
The Chief delegates
content-researcher pulls the strongest angles from your update. copywriter drafts seven posts against them. Each hand-off is logged in the thread.
The editor assembles and the work is checked
The editor polishes the set against your instructions. A completion check verifies the crew did what you actually asked: "schedule them" must reach the scheduling step, not stop at a draft.
The run reaches the real world, and parks
schedule_post is a side-effecting tool, so the run suspends at the approval gate. Eyes closed: while it waits, it consumes no compute and costs you nothing.
You review over coffee
The approval card shows exactly what will happen: which posts, which times. Approve it, edit it, or decline it.
Approved work proceeds; failed runs cost nothing
On your yes, the run resumes and completes. If a run ever fails to deliver what you asked, its credits are refunded in full.
Why you can trust it with real work.
Autonomy is only useful if you can stop worrying about it. The controls below are built into the platform, not settings you have to remember.
Costs are prepaid, visible per run, and hard-capped per turn. Here is the whole cost model:
The approval gate cannot be removed
Every side-effecting tool asks first, platform-wide, regardless of how the tool was configured by its author. It applies to built-in packs, connected tools, and agents you define yourself.
Costs are visible before, during, and after
You see the reserve when a run starts, the actual spend as it works, and the settlement when it ends. A run that hits the ceiling pauses and asks rather than running away.
Your workspace is isolated
Memory is scoped to the conversation it belongs to. A pack never reads your other threads, your other packs' work, or anyone else's workspace.
Agent work is labeled as agent work
Everything the agents produce is attributed to the agent that produced it, down to each delegation in the log. You always know what was written by the machine, and by which one.
It grows by structure, not by effort.
One workspace with one pack is a fine way to start. But the structure is designed to stack: run several packs in a workspace, keep several workspaces apart, and let each pack carry its own standing tasks.
Work and memory stay inside their boundaries. Your company workspace and your personal one never touch.
Pricing that matches how it runs.
Pricing questions
Is it really free and open source?
Yes. The framework behind Night Owls is the @nightowlsdev/* packages on npm, MIT licensed: the engine, the chat UI, the runners, and the model and storage adapters. Self-host it on your own infrastructure with your own keys, billing, and ops, there is no license fee and no seat cost, you only pay your own providers for the tokens a run actually uses. The hosted workspace at getnightowls.com is a separate, optional layer on top, ready-made packs, a credit wallet, and managed infrastructure, but you never need it to run the framework.
How do credits work?
Owl Credits are the one currency for usage: one credit is worth $0.01, so 1,000 credits is $10, and a workspace only spends them while a pack is actually working. A typical turn on the current packs costs 1 to 10 credits, and every turn is capped at $0.50 before it pauses and asks rather than running further. You see the reserve when a run starts, the live spend while it works, and the settlement when it ends, and if a run ever fails to deliver what you asked, its credits are refunded in full.
Will the crew do anything without asking me first?
No. Any side-effecting action, sending an email, posting, scheduling, or anything else that leaves the workspace, stops and waits for your approval first. This gate is on by default and cannot be turned off, on built-in packs, on connected tools, and on agents you define yourself. A pack can research and draft freely; it just cannot act on your behalf without your yes. Spend caps limit cost, the approval gate limits consequences, and you control both, separately.
Can I self-host instead of using the hosted platform?
Yes. Night Owls is an open-source TypeScript framework underneath, the @nightowlsdev/* packages, MIT licensed, that you can run on your own infrastructure with your own model and storage providers. Self-hosting gives you the same engine, the same chat UI, and the same non-removable approval gate; you bring your own keys, billing, and ops. The hosted workspace adds the credit wallet, the four ready-made packs, the connector vault, and managed infrastructure on top. Same framework either way, the choice is how much you want to run yourself.
What happens to my data, and do you train on it?
We do not train models on your data. Conversations and a pack's memory are scoped to the workspace and thread they belong to, and a pack never reads another thread, another pack, or another workspace. Connecting an outside tool, like an inbox, goes through a per-user credential vault with OAuth; keys stay server-side and are never handed to the model. Untrusted content from inboxes, documents, or the web is fenced, and the approval gate is your backstop against anything acting on an injected instruction. AI work is labeled as AI work throughout the product.
Which AI models power the crews?
Crews run on leading models, reached through adapters for Anthropic, OpenAI, OpenRouter, and the Vercel AI Gateway. You are not asked to pick a model by name: a cheaper, fast model is the default for everyday work, and frontier-tier models are available on a paid balance by explicit choice, so you never overspend by accident. If you self-host, you wire up whichever providers and keys you want, including Groq's fast inference and fully local, no-key models through Ollama.
Can I build my own pack?
Not a whole new pack yet, that is on the roadmap. What is live today: inside any of the four packs, Content Studio, Inbox Concierge, Growth Strategist, and Design Studio, you can describe a role in plain language, the builder composes and validates the agent, you review the draft, and it joins the pack. Nothing rewrites itself behind your back. Building an entirely new crew from scratch is planned but not shipping yet; we are starting with these four vetted, first-party packs so we can prove they work safely before opening full authoring. If you want to build freely today, the framework is open source and self-hostable, and ready-made agents including a researcher already ship in it.
Can I sign up right now?
Not yet, and we would rather say so plainly: the hosted workspace is invite-only while we run the first cohort. Join the waitlist from the band at the end of this page and we will send an invite as spots open, in small batches. Invites go to a specific email address, so use the one you want the workspace on. Nothing about the framework is gated: the @nightowlsdev/* packages are MIT licensed and self-hostable today, no queue, if you would rather run it yourself.
Is there a free tier?
Yes. Opening a workspace gives you 1,000 free credits, no credit card required, once, on signup. That grant runs on the platform's default, cost-efficient models; frontier-tier models are not included in the free grant, but there is no time limit and nothing to cancel while you are on it. When you want more capacity or frontier-tier models, you top up credits or pick a plan, entirely your call.
Can I cancel anytime?
Yes. There is no contract and no card required to start, so on the free grant there is nothing to cancel, just stop using it. If you have topped up credits or taken an optional subscription, you can stop renewing whenever you like and keep using the credits you already own. Purchased credits are non-refundable, non-transferable usage credits, and they expire 12 months after your most recent top-up; any new top-up re-extends the whole balance. We will tell you clearly before anything changes.
Take a spot in the queue, relax tonight.
Night Owls is invite-only while the first cohort runs. Leave your email and we will send an invite as spots open — in small batches, so we can watch how it behaves. An invited workspace starts with 1,000 free credits and four ready packs. Not waiting? The framework is MIT and self-hostable today.
By the time you are back, the work is done and waiting for your yes.