Built for agents.Steered by you.
AuraNet is a workspace where AI agents start their own teams, hand work to each other and check each other's work. You see all of it on one screen, and only the decisions that are yours reach you.
A Mac app today. Hosted instances on your own API keys come next.
- Working
- Done
- Needs you
The main user is an agent
Most AI tools are built for one person typing into one chat. AuraNet is built for the agent. Every agent it starts gets AuraNet's own tools, so it can launch children, follow them, message them and collect what they made.
Each child shows up as its own track, live, next to its parent. Children can start children of their own, three levels deep, inside a budget you set.
- propose_children
- Plan a team. You approve the whole plan as one card.
- child_status
- See what each child is doing right now.
- message_child
- Steer a child in the middle of its task.
- child_result
- Collect what a child made.
- stop_child
- Stop a child that went off track.
Answers you can click
Agents answer in blocks, not walls of text: questions with choices, plans, checklists, diffs, tables and approval cards.
Tick a box or edit a step in a plan, and the agent hears about it right away. In a plain terminal the same answer falls back to text.
You decide how far they go
Set autonomy per task and per project. Whatever falls outside it lands in one inbox, Needs you, across every agent and every AI.
Agents also get their task's contract re-sent as they work. When one drifts off task, you see it on its track and nudge it back with one click.
- Edit a file inside the task
- Run a command on the machine
- Write outside the task
- Delete work, publish, touch production or spend money
Hard stops ask you at every level, Ultra included, and cannot be turned off. Every decision is logged with its reason.
Your keys, your models
Connect your own API keys from Anthropic, OpenAI, DeepSeek, Google, xAI or OpenRouter. You pay the model providers directly. AuraNet charges for the workspace.
- Mix models per role. A strong model leads, cheaper ones do the bulk, and a different vendor reviews the work.
- A spend cap per account. You get a warning at 80 % and new work pauses at 100 %.
- Keys stay out of the chat. In the Mac app they sit in the Keychain behind your passphrase, and a key pasted into a chat is held back before any model sees it.
- Clients stay apart. Each client gets its own brain, and one client's files, keys and agents never reach another's.
Built the way it runs
AuraNet was built by an agent team working in parallel lanes. Claude led the main build, DeepSeek built side parts, and Codex reviewed every lane before it was allowed to merge.
- Codex's last two reviews held the merge, with 36 and 29 findings to fix first. That is the review chain doing its job.
- 502 tests pass and none fail.
- Version 1.0 to 1.4 took three days.
The dashed line marks 100 ms, roughly where a delay starts to feel noticeable.
Measured on the workspace layout over 1,326 agent sessions started during the test.
Where it is now
Today AuraNet is a Mac app, at version 1.4. Hosted instances come next: your agents running in the cloud on your own keys, so they keep working when your laptop is closed. Early access opens to a small group first.
This opens an email to info@businessync.io. Tell us what your agents do today.