The hardware
A wristband built for voice.
Screenless, lightweight and ready for life away from a keyboard. The hardware work will cover electronics, sound, battery and enclosure alongside the intelligence it connects to.
02 Project / Hivemind
In planning · Come build it with us
Hivemind.
One assistant.
From wrist to world.
We’re building an open source architecture for personal AI, together with a voice wristband. The ambition: turn the context you choose to share into useful memory and coordinated action across your digital world.
A physical wristband. A cloud architecture. A web application. Mobile is under consideration.
No prior AI experience required. You’ll have people to learn with.
“What do I need for our next project meeting?”
Ask in the web application
A planned home for your assistant, connected nodes and task activity.
Connect what you already shared
Find the relevant decisions, notes and updates in your chosen sources.
Work through connected tools
The cloud agent gathers material it has permission to access.
Review your meeting brief
See the result, activity and any approval requests in one dashboard.
Two ways into the same Hivemind. The software can also run without the wristband.
The software architecture
Intelligence that connects the pieces.
Cloud execution, a web application, shared context and connected agents. The architecture will be open source, so people can run it independently or build new nodes around it.
Building on open source
Why Hivemind
if Hermes exists?
Less briefing.
More follow through.
Our ambition is personal AI that carries the context of your day into the work you need done.
Context you allowMemory that stays relevantHelp across your devices
Hermes · the agent foundation
A capable engine to build on.
Hermes already brings tools, skills, delegation and automatic memory. Our cloud proposal builds on that foundation.
Hivemind · our design goal
A connected personal AI system.
Design the wristband, context architecture and connected agents together. Carry a request from your voice to the right tools, with the relevant memory and one place to follow the work.
You can extend Hermes yourself. Hivemind’s ambition is to make these pieces work as one open source system, so you spend less time assembling context and managing handoffs.
Explore the Hermes foundationWhere we want to push further
Context, continuity and control. Our engineering goals.
Context before the next prompt.
Connect conversations, tasks and sources you allow. Background processing would organize useful details, so you have less to explain next time.
Pick up where the task stopped.
Your laptop sleeps. The cloud finishes its part. When the laptop returns, Hivemind would check what actually completed before continuing, with a clear receipt for each result.
Action receipts would record and verify each step, prevent duplicate actions where supported and flag uncertain outcomes.
Useful context. On your terms.
“Use these notes for this project until Friday.” Give context a purpose and expiry, then see which memories depend on it and where its removal is still pending.
Context permissions would follow data into derived memories, device caches and waiting tasks within Hivemind.
These are proposed system capabilities. Hermes already has durable task queues, and its memory providers offer provenance and deletion. Our aim is to connect those foundations across devices, with verified outcomes and enforceable context permissions.
What would we add to Hermes?The technical distinction, with sources
Recover actions, beyond saving a task.
Hermes documents persistent queues, but a running delegated task becomes uncertain after a process restart. We propose a shared action protocol: record intent, check the external result after reconnection, then resume or ask for review. A receipt means evidence of the outcome, rather than an agent’s completion message.
First scope: cloud and laptop, with one supported file action. We cannot guarantee duplicate prevention on every website.
Hermes delegation and recovery ↗Withdraw permission across the system.
Hindsight, a Hermes memory provider, already tracks sources and removes derived observations. We propose an additional permission layer that checks purpose and expiry before retrieval or action, cancels affected waiting work and collects removal acknowledgments from managed nodes.
The dashboard would show pending devices. This cannot erase information already exported or sent to another service.
Hindsight sources and deletion ↗Research reviewed September 11, 2026. We found no documented, integrated Hermes flow covering these complete proposals. These are integration goals, not claims of exclusive inventions or demonstrated advantages.
Go a little deeper
The idea, under the hood.
Open what interests you.
Explore at your own pace.
How the pieces fitNodes, memory, tools and the dashboard
Nodes & coordination
Different environments, one task.
A node is a connected part of the system with its own context and abilities, such as the wristband, cloud, browser or computer. Hivemind is the software architecture connecting these parts. We plan cloud execution and a web application for requests, connected nodes and activity, with a mobile application under consideration. A node contract is a shared format for capabilities, status and results, so teams can build parts that work together.
Memory & context
Useful continuity, with a choice about sharing.
The proposal separates your personal memory from explicitly chosen team spaces. Retrieval means finding the pieces of past context that matter for a request. We’re exploring a background process that organizes useful facts, connects related information and updates outdated details, with consent and deletion rules to define before external use.
Tools & permissions
Access you understand and control.
The cloud proposal builds around Hermes and tool connections such as MCP, a common way for an agent to communicate with other software. We plan credentials limited to the access each task needs, approval before irreversible actions and a record of tool activity. Testing these controls is part of the engineering work ahead.
One dashboard
A place to ask, follow and review.
The web application will bring together four core views: a registry of connected nodes, a detail page for each node, an activity feed and a command bar for requests. Shared status labels and panels for each device should make the whole system understandable as each team’s integration arrives.
What each subteam will buildSeven entry points into the same project
Wristband
Build a small, screenless voice wristband. Integrate microphones and speakers, develop Bluetooth connectivity and design for efficient power use, a custom circuit board and a comfortable enclosure. Physical hardware is a core part of the project, alongside the software architecture.
USB device
Explore a USB path to mouse, keyboard and screen interaction on Mac and Windows. Research connectivity, reliable actions and coordination with the PC agent.
Cloud
Build the cloud side of Hivemind: agent execution, memory, user separation and task routing. Work with Design/Product on the web application and explore a possible mobile application. Research reliable tool connections and websites that restrict traffic from data centers.
PC / OS
Connect local files and applications with permission. Account for OS restrictions and the fact that this node is available only while the computer is running.
Browser
Explore Chrome and Safari extensions using page and tab context. Work on finding relevant content, choosing between direct software connections (APIs) and actions on the page, and completing tasks efficiently.
Architecture / Research
Design memory, agent tools and context handling. Compare models for task quality, speed and cost, and test the system against real outcomes.
Design / Product
Create the dashboard’s four views and reusable interface components. Define the data each view needs, design device panels and help teams resolve integration gaps.
The hard parts we want to get rightUseful memory, understandable actions and real value
Will it remember what matters?
Explore selective retrieval: bring in relevant facts, update outdated context and keep personal and shared memory properly scoped.
Can I understand what it’s doing?
Design a clear activity trail and approval steps for consequential actions. A confident answer still needs to be checked against the actual result.
Will it save more than it costs?
Measure task completion time, correctness and model usage together. Test model choices and context size instead of assuming a cheaper model is enough.
We’re at the planning stage. Implementation hasn’t started and there’s no public release yet. We’re building toward a physical wristband and an open source software architecture, with cloud execution and a web application. A mobile application is under consideration.
Your first AI project is welcome here
Learn.Build.Belong.
No prior AI experience required. We’ll pair you with a subteam lead. Bring your questions, your perspective and a willingness to try.
Find your place Interest form · Google sign in requiredPeople to learn with
Educational workshops, speaker events and a 450+ member Discord community where questions are welcome.
More than coding
Design, product, outreach, marketing, operations and finance all have a place in the project.
Build thoughtfully with AI
Vibe coding is welcome. Learn what the code does, test it and review it with your team.
What we’re aiming for
Semester targets for real time saved. We’ll validate these with users as we build.
End of Fall
30
people using Hivemind
Target: 2.5 hours saved per person, per week on average
End of Spring
200
people using Hivemind
Target: 5 hours saved per person, per week on average
Northstar
We’ll measure task speed, accuracy and cost against real outcomes. These are research goals, not proven gains.
Speed
How fast can we have agents complete tasks?
Accuracy
How accurately can agents complete tasks?
Cost
How cheaply can agents complete tasks?
Subteams
Find your part of the Hivemind.
Seven technical subteams, building one connected system. Explore a team below; you can contribute across areas as you learn.
- EL
Agentic Wristband
Evan Liu
Build a lightweight, screenless wristband for talking to agents while your hands and eyes are busy.
- EL
Agentic USB Stick
Evan Liu
Explore a USB device that lets agents operate a Mac or PC through mouse, keyboard and screen.
- AI
Cloud Agent
Anand Indukuri
Build the assistant’s cloud service and web app, with a mobile app also being considered.
- IS
PC / OS Agent
Ife Solarin
Lives on your computer with access to your files, doing what the cloud cannot reach.
- BS
Browser Extension Agent
Bala Shukla
Safari and Chrome extensions that act across your tabs, in sessions you are already signed into.
- RK
AI Architecture / Research
Ryan Khalili
Design memory, choose models and build the tools that help the connected agents work together.
- JD
Design / Product
Joslyn DeGuzman
Shape the dashboard, views for connected devices and how the whole Hivemind feels to use.
Meet your team leads
Eight people to build with, learn from and ask questions. Find your starting point; we’ll help you take the next step.
Evan Liu
President / Hardware Lead
Ryan Khalili
AI System Architecture / Research Lead
Anand Indukuri
Cloud Engineering Lead
Ife Solarin
PC Engineering Lead
Bala Shukla
Browser Engineering Lead
Joslyn DeGuzman
Design / Product Lead
Jorge Roji Pezzoli
External Outreach / Finance Lead
Zahra Moosani
Internal Outreach / Marketing / Operations Lead