JuggrOS Get in touch

Pre-launch · used daily · starter kit v0.8.1

Juggr

The operating system for guarded AI agents at work

Memory, guardrails and an audit log, with approval gates on anything risky.

Become a design partner Get in touch
Juggr's Today screen for a demo user, Sam: the first task on today's plan with an Open Atlas button, a four-item plan for the day, and a Needs you panel with two Atlas approvals waiting for review.
Juggr's Today view on a phone: the first task on the plan, an Open Atlas button and the Needs you list with Review buttons.
The Today screen on desktop and phone, with demo data.

01 · The problem

Agent safety today is per-tool and per-person.

Each developer sets their own rules. Nothing is shared across the team, and nothing is tracked across tools.

  1. No shared team memory

    What one person's agent learns stays with that person.

  2. No team policy

    Rules live in individual setups, one developer at a time.

  3. No cross-tool audit log

    There is no single record of what agents did, across tools.

02 · How it works

Every agent write passes guard checks.

Juggr sits between your agents and your work, with three layers that apply to everyone on the team.

Memory

Shared knowledge bases, called Brains, for a person, a team or the whole company, so context isn't trapped in one setup.

Guardrails

  • Every agent write passes guard checks
  • Guard hooks block writes to protected paths
  • Secrets are scanned

Audit log

One record of what agents did, plus approval gates on anything risky.

03 · Approvals and Atlas

Atlas asks before anything high-impact.

Atlas is Juggr's assistant. It runs locally, with daily and weekly loops, and puts a human in charge of the actions that matter.

The Needs you screen in Juggr. Atlas asks to save the launch announcement draft, with Approve, Deny and Reply to Atlas buttons. The open details show the file it will write, the three lines it writes, the risk (medium, adds or replaces a file in your vault), the knowledge and setting it runs under, and that unanswered requests are denied after 10 minutes.
Every approval shows what will change, how risky it is, and which setting it ran under.

Approval before high-impact actions

Risky actions wait for a yes: approve, deny, or reply to Atlas. Unanswered requests are denied, not done.

Right where you work

The same approval cards show up in the Atlas chat and in one Needs you list.

Runs locally

Atlas runs on your machine, on daily and weekly loops.

Financial actions stay suggest-only

Atlas can suggest. It does not act on money.

Two approval cards in the Atlas chat. One asks to save the launch announcement draft, marked medium risk, showing the file and the lines it writes. The other asks to email the launch checklist to the design team, marked high risk. Each has Approve, Deny and Reply buttons and is denied after 10 minutes without an answer.

04 · Guardrails

You choose how careful Atlas is.

Four levels, and each session keeps its own. Whatever the level, every agent write still passes guard checks.

  1. Ask first Recommended

    Reads freely. Asks before it runs anything, sends anything or changes a file outside your notes.

  2. Edit freely

    Writes and changes files in your knowledge on its own. Asks before anything else.

  3. Plan first

    Looks around and writes a plan. Changes nothing until you OK the plan.

  4. Auto

    Gets on with it. A safety check stops risky steps and asks you about those.

The How careful is Atlas? picker in Juggr with four levels: Ask first (recommended and selected), Edit freely, Plan first and Auto. Each level says what Atlas does on its own and what it asks about first. Auto warns that it can still make mistakes.

05 · Memory

The What Atlas remembers card in Juggr: four short notes about how the user works, such as keep answers short and always ask before sending anything outside the team, saved with the user's notes, with an Edit button and other notes Atlas keeps on working hours and writing style.

Atlas remembers, in notes you can read.

What Atlas learns about how you work is kept as plain notes in your own vault, next to the rest of your notes.

  • Open and edit what Atlas remembers at any time
  • Only you can see it, unless you share it
  • Sharing always asks first and shows who will see it

06 · Team brains

Knowledge moves up only with approval.

A Brain is a knowledge base for a person, a team or a company. Knowledge moves from yours, to team review, to company, step by step and never silently.

  1. Only you

    Your knowledge

    Your notes and the apps Atlas can read stay in your own space. Nothing is shared unless you share it. The team spaces you belong to sit right next to it.

    My knowledge in Juggr: what Atlas knows (your notes, plus Gmail, Google Calendar, Google Drive and Slack to connect), a banner saying nothing here is shared unless you share it, and a Your Team knowledge list with brains such as Sample Co, Accounting, Business Operations and Finance, each showing how many people are in it.
  2. Needs approval

    Team review

    A shared note waits in review. The reviewer sees exactly what the team knowledge will gain, then approves or declines.

    The Team brain review tab in Juggr. A note called Planning calendar for next year, shared by Morgan Reyes to the Sample Co brain, waits for review. It shows what the Sample Co brain will know if approved, with Approve and Decline buttons.
  3. Shared with the company

    Company knowledge

    Once approved, the note lands in the shared space with who added it and who OK'd it. Your own knowledge stays private.

    The Sample Co brain, shared with 13 people. Under What's in it, the Planning calendar for next year note, added by Morgan Reyes and OK'd by Skyler Grant. A What stays private card lists your notes, email, calendar, files, chats and drafts.

Live Multi-user team review, with approval before anything is shared.

07 · Projects

Your work, on one board.

Projects sit next to your plan and your approvals: active, blocked, paused and ideas, at a glance.

The Projects board in Juggr with four columns: Active (Spring product launch, Customer onboarding guide, Website refresh, Customer survey), Blocked (Hire a support lead), Paused (Podcast pilot) and Ideas (Team offsite, Monthly newsletter).

08 · Local-first on NVIDIA

Uses your local NVIDIA GPU when you have one.

If the machine has an NVIDIA GPU, Juggr puts it to work.

Today

~0.4 s

Voice input, per clip

Voice input runs through whisper.cpp on the local GPU, at about 0.4 seconds per clip.

Planned

Open models

On rented NVIDIA GPUs

Planned: running open models on rented NVIDIA GPUs.

09 · Plans

Free to start, paid to scale.

Pre-launch, but already a working product used daily. Pricing for paid tiers is not set yet.

Starter kit

Free

The way in. Current version: v0.8.1.

Pro

Paid · pricing to come

Paid tier. Details to come.

Team

Paid · pricing to come

Paid tier for teams. Details to come.

Done-for-you

Implementation

Done-for-you implementation for businesses.

Become a design partner.

Juggr is pre-launch. If your team runs AI agents at work and wants shared memory, team policy and one audit log, let's talk.

Become a design partner Get in touch

Sylvain Nehme, founder · Habibi Capital