Agent hardware · In development

The always‑on
computer for AI agents.

Nanisyamkumar builds private, dedicated hardware where AI agents can run, remember and act — on your desk, around the clock, under your control.

Join the waitlist How it works
NVIDIA RTX Spark N1XSuperchip
128GBUnified memory
Qwen3.8‑27BOn-device
Core · Concept render Always on
24/7 runtimeAgents never sleep
Local firstPrivate by default
You approvePhysical consent
01 — The problem

Agents are becoming coworkers. They still don't have a desk.

/01

They sleep when your laptop does.

Today's agents live in browser tabs and terminals. Close the lid and the work stops — no persistent runtime, no long-term memory.

/02

They need your most private context.

Useful agents need email, files, calendars and the physical world around you. Right now that context is shipped to someone else's servers.

/03

They act, but nobody's watching.

Agents are starting to send, book and pay. There is no physical, human checkpoint between a plan and an action.

Our answer

A dedicated, private computer for agents — with the senses to understand the real world and the controls to keep a human in charge.

Who it's for

Designed first for the people already running agents.

Agent developers

A local, always-on runtime with an SDK, sensor APIs and 128GB of memory to build and test real agent workloads.

Privacy-first professionals

Clinics, legal and finance teams who want agent help without sensitive data leaving the premises.

Power users & small teams

People who want their agents reliable 24/7 — without leaving a laptop open or paying per token for every routine task.

02 — Why now

Three curves just crossed.

For the first time, a capable agent model, the silicon to run it locally, and the demand for agents that act can all meet in one quiet box.

Aug 202601

Open models got good enough.

Qwen3.8‑27B is an open-weight model built for coding and long-horizon agentic tasks, compact enough for consumer hardware.

Oct 202602

The silicon arrived.

NVIDIA RTX Spark N1X combines a Grace CPU, a Blackwell GPU and up to 128GB of unified memory in a laptop-class chip.

Now03

Agents moved from chat to action.

Agents that act need what a chat window can't give them: a persistent runtime, long-term memory and a human approval step.

03 — The system

One brain. Three senses.

Core runs your agents. Sense, Dock and Clip give them eyes, ears and a conscience — all connected over an encrypted local mesh.

The brain

Core

A whisper-quiet agent computer that stays on, keeps memory local, and runs a frontier-class open model without the cloud.

  • ProcessorNVIDIA RTX Spark N1X
  • Memory128GB unified, standard
  • On-device modelQwen3.8‑27B
  • Size12.7 × 12.7 × 4.6 cm
Target price $2,499
Eyes & ears

Sense

A 4K camera and six-mic array with a physical shutter. Faces and screens are redacted on Core before any agent sees a frame.

Target price $249
The conscience

Dock

A desk display that shows what every agent is doing, with a physical Approve Key and a hardware kill switch.

Target price $349
On the go

Clip

A light wearable that captures context when you tap it — and glows whenever a sensor is active. No exceptions.

Target price $199
04 — Inside Core

Six layers.
Nothing wasted.

A laptop-class superchip, given a desktop-grade thermal system and a body you'd happily leave on a shelf — forever on, barely heard.

  1. 01
    Glass halo lidStrengthened glass with a printed halo.
  2. 02
    Halo light guide24 LEDs show agent status at a glance.
  3. 03
    Logic boardNVIDIA RTX Spark N1X with 128GB unified memory.
  4. 04
    Thermal systemCopper vapor chamber and a slow, quiet impeller.
  5. 05
    Unibody enclosureMachined from recycled aluminum.
  6. 06
    Base and powerBuilt-in power supply. No brick.
01 02 03 04 05 06

Processor

NVIDIA RTX Spark N1X

CPU · GPU

20-core Grace · 6,144-core Blackwell

Unified memory

128GB LPDDR5X

On-device model

Qwen3.8‑27B at peak token speed

05 — How it works

Hardware, OS and models, designed as one.

Nanisyamkumar is a full stack: devices that sense and approve, an operating system that runs agents safely, and a model layer that is local by default.

Layer 1 · Devices

Core, Sense, Dock, Clip

Compute, perception and physical consent — paired out of the box over an encrypted local mesh.

  • Always-on agent compute
  • Camera, mics, wearable capture
  • Approve Key and kill switch
Layer 2 · Agent OS

A runtime built for agents

Every agent runs in its own sandbox, with scoped permissions and a long-term memory that stays on Core.

  • Sandboxed agent containers
  • Encrypted Memory Vault
  • Permissions for every tool and sensor
Layer 3 · Models

Local first, frontier when needed

Qwen3.8‑27B runs on-device for everyday work. Harder reasoning can route to frontier models via API.

  • Qwen3.8‑27B on-device
  • Claude and other models via API
  • Bring your own open weights
A task, end to end
01
PerceiveSense or Clip captures context, redacted on Core.
02
Plan locallyQwen3.8‑27B plans the task on-device.
03
Escalate, if neededHard steps can go to Claude via API, with permission.
04
Ask youDock shows the proposed action in plain language.
05
Act & rememberApproved actions run; memory stays on Core.
Hybrid intelligence

Local by default. Frontier when it matters.

Most agent work — triage, summaries, routine tool calls — runs privately on Core. For the hardest reasoning steps, a policy can route a task to a frontier model such as Claude through its API. Sensitive details are redacted on-device first, and nothing leaves without your permission.

Agent task Core POLICY Qwen3.8‑27B On Core · default · private Claude Via API · opt-in · redacted
06 — Built for developers

Ship agents to real hardware.

An SDK, sensor APIs and a local runtime, so you can give an agent memory, senses and an approval step in a few lines of code.

  • Python and TypeScript SDKs
  • Every sensor exposed as a permissioned tool
  • Local model endpoint, plus frontier models like Claude via API
  • MCP-compatible tool servers and sandboxed containers
Claude APIQwen3.8‑27B · localMCPPythonTypeScriptContainersWebhooks

We build our SDK and internal tooling with Claude Code.

inbox_agent.pypolicy.yaml
1from nanisyamkumar import Core, Policy
2
3core = Core.connect()  # finds Core on your local network
4
5agent = core.agents.create(
6    name="inbox-triage",
7    model="local/qwen3.8-27b",       # private, on-device
8    escalate_to="anthropic/claude",  # hard reasoning, opt-in
9    tools=["mail", "calendar", "sense.camera"],
10    memory=core.vault("inbox"),
11    policy=Policy(
12        redact=["faces", "account_numbers"],
13        approve_on_dock=["send", "pay", "delete"],
14    ),
15)
16
17agent.run(schedule="always")  # keeps working after you log off
Dock · Concept render
07 — Control by design

Agents propose.
You decide.

Safety isn't a settings page. It's built into the hardware, so the most important controls are ones you can touch.

Approve Key

Nothing that sends, pays or deletes happens until you press it.

Physical off switches

Sense's mechanical shutter and Dock's kill switch cut sensors in hardware.

Local by default

Memory and models stay on Core. Cloud calls are opt-in, per task.

Yours to keep

Open formats and full export. Your agents' memory belongs to you.

08 — Planned configurations

Same brain. Pick your setup.

Every Core is planned with the same superchip and the full 128GB of unified memory. Configurations differ in storage, networking and included devices.

Core

Target
$2,499target price
  • ProcessorNVIDIA RTX Spark N1X
  • Memory128GB unified
  • Storage2TB SSD
  • Networking10GbE · Wi‑Fi 7
  • DevicesCore only
Join waitlist

Core Plus

For developers
$2,999target price
  • ProcessorNVIDIA RTX Spark N1X
  • Memory128GB unified
  • Storage8TB SSD
  • NetworkingDual 10GbE · Wi‑Fi 7
  • DevicesCore only
Join waitlist

Core Kit

Target
$3,599target price
  • ProcessorNVIDIA RTX Spark N1X
  • Memory128GB unified
  • Storage4TB SSD
  • Networking10GbE · Wi‑Fi 7
  • DevicesCore, Sense, Dock, Clip
Join waitlist

Target pricing and specifications for planned products; subject to change before launch. Every configuration ships with Qwen3.8‑27B on-device.

09 — Roadmap

Where we're headed.

Software first, so developers can build before the hardware ships. Dates are current targets.

Now · Q4 2026

Concept & design

Industrial design, concept renders and Agent OS architecture.

Q1 2027

Developer preview

SDK and local runtime for early developers on N1X-class hardware.

Q2 2027

Prototype hardware

First engineering prototypes of Core and Dock.

H2 2027

Founding beta

Small-batch units for waitlist members.

2028

General availability

Core, Sense, Dock and Clip.

10 — Company

Building the home for AI agents.

Our mission is to give every AI agent a safe, private place to work — and to keep people firmly in charge of what agents do.

ND
Nanisyamkumar Dara

Founder

Stage

Pre-launch. Product in development.

Focus

Agent hardware and the Agent OS that runs on it.

How we build

Software first, with Claude Code across our SDK and tooling.

Contact

hello@nanisyamkumar.com

Join the waitlist

Give your agents
a home.

Get updates on the developer preview and founding beta. Developers get invites first.

Thanks — you're on the list. We'll be in touch.

We store your email to send developer-preview and beta updates. You can unsubscribe by email. Prefer email? Write to hello@nanisyamkumar.com