K3rnel.ai
Local-first intelligence / In development

Intelligence,
without the cloud.

Your browser is the computer.
Your device is the datacenter.
A private AI workspace, powered by hardware you already own.

Explore the workspace See the architecture

On-device inference. No remote prompt processing.

WEBGPU / LOCAL COMPUTE
K3 / INFERENCE CORE
QUANTIZED MODEL · CLIENT-SIDE
DESIGNED TO STAY ON YOUR DEVICE
Local by designOffline-capableOpen-model foundation
SCROLL TO EXPLORE   ↓
BUILT ON AN OPEN
TECHNOLOGY FOUNDATION◈WebGPU⌁WebLLM▧IndexedDB◇OPFS⟐Web Workers
01 / A different foundation

Less dependency.
More possibility.

Cloud-first AI comes with trade-offs. K3rnel explores a simpler direction: move the intelligence closer to you.

01 / COST

Compute you already own.

Local inference avoids per-prompt cloud API charges. Put your device's GPU to work instead of renting every response.

02 / PRIVACY

Your context stays yours.

Designed to process prompts and documents on-device, without sending them to a remote inference provider.

03 / FLOW

One connected workspace.

Bring conversations, documents, and useful tools together. Spend less time switching tabs and moving context around.

02 / The architecture

A small kernel.
A whole new workspace.

An operating-system-inspired architecture. A local model coordinates context and tool requests while browser APIs handle execution and storage.

MODEL + TOOLS + PERSISTENT CONTEXT
01

Intelligence at the edge.

The proposed WebLLM runtime uses WebGPU to run supported quantized models locally. Model choice depends on your device's memory and browser support.

WebGPU · @mlc-ai/web-llm
02

Tools that do the work.

A structured interpreter dispatches allowed tool calls to background workers. Keep the interface responsive, with explicit permissions for sensitive actions.

Function calling · Web Workers
03

A memory of your own.

Threads and preferences live in IndexedDB. Larger files use the Origin Private File System. Export and backup matter: browser storage can be cleared.

Dexie.js · IndexedDB · OPFS
03 / The experience

Big ideas. One little tab.

Interactive interface preview ↙
K3RNEL / WORKSPACE PREVIEW● CONCEPT DEMO
A WORKFLOW, REIMAGINED
YOU
Summarize my project notes and turn the next steps into a checklist.
K3
A local workflow could read your selected document, identify the key ideas, and build an actionable checklist.
◇ read_document → project-notes.mdEXAMPLE
User-approved file · On-device processing
Run the walkthrough to preview the experience.
Preview a local document workflow
Scripted interaction only. No model is loaded and no files are accessed.

⌁   Local model runtime

Designed for supported, compact open models.

◇   Permission-aware tools

Explicit actions, clear boundaries, useful widgets.

▧   Persistent local context

Your threads and files, organized on your device.

04 / From concept to kernel

Small steps.
A deliberate build.

A proposed four-week implementation sprint, moving from the inference foundation to a cohesive browser workspace.

Week 01 / Foundation

Light up the core.

Scaffold the Vite project, integrate WebLLM, and validate local prompt execution on supported hardware.

Week 02 / Tooling

Give it useful tools.

Build the tool dispatcher, background worker execution, and the first dynamic workspace widgets.

Week 03 / Persistence

Make context last.

Add local threads with Dexie.js and document storage through OPFS, with clear storage controls.

Week 04 / Experience

Bring it all together.

Refine the interface, add command shortcuts, and prepare a static deployment for early testing.

Proposed milestones, not completed releases. Timing may change during development.

05 / Planned pricing

Your hardware. Your advantage.

A free local foundation, with optional services for a more connected workspace.

Local

Your personal AI foundation
$0 / month

The planned core experience, running on your own device.

  • Supported local open models
  • On-device threads and file storage
  • Offline use after setup and caching
  • No per-prompt cloud inference fee
Explore the roadmap
PLANNED UPGRADE

Pro

A workspace that travels
$10-15 / month

Optional cloud services and advanced capabilities.

  • Everything in the local foundation
  • Planned encrypted backup and sync
  • Advanced multimodal plugins
  • Developer SDK for custom tools
Understand the details

Marketplace

Build on the foundation
Create & share

A proposed home for third-party productivity widgets.

  • Publish useful workspace micro-apps
  • Extend the tool-calling schema
  • Reach the K3rnel community
  • Revenue-share terms to be defined
Meet the builders

Concept pricing. Plans are not available to purchase. Final features and pricing may change.

Does this page run an actual language model?

No. This is a landing website and interactive interface preview. The walkthrough is scripted. The WebGPU model runtime, file tools, and persistent workspace are implementation work described in the roadmap.

What does "local-first" mean here?

The core design keeps inference on your device. Initial downloads need a network connection. Optional backup, sync, or connected plugins would involve external services and should have explicit controls and clear disclosures.

Will the workspace work on every device?

The planned runtime needs a compatible WebGPU browser and enough available memory for the chosen model. Performance and model availability depend on hardware. Offline use also depends on retaining cached application files and model weights.

Are Web Workers a security sandbox?

Workers move tasks off the main UI thread, but are not a complete security boundary for untrusted code. A production tool runner needs restricted capabilities, resource limits, permissions, and an isolation strategy appropriate to its threat model.

Built with a different belief

The datacenter
should be optional.

K3rnel.ai is in active development. We're exploring a future where a useful, private AI workspace starts with the device in front of you.

RS
Ritesh ShahLead Architect
NR
Nitin RajCo-Author
Explore the foundation