OverviewCanvasSandbox

VaultEd

Learn AI.
Not the cloud bill.

The private, on-device AI sandbox built for Canvas, MacBook Neo, and modern Computer Science classrooms.

Launch Canvas SandboxTry Standalone Demo
VaultEd — Lab 04 · Agentic Refactor · running on-device

Specification

Tools granted

read_filewrite_filerun_testsnetwork
Run agent against spec

Agent trace

0 tokens billed

◆ plan · 4 acceptance criteria identified

› read_file("intervals.py") → 0 bytes

› write_file("intervals.py", 18 lines)

› run_tests() → 3 passed, 1 failed

◆ plan · sorted in place — preserve input, retry

› write_file("intervals.py", 18 lines)

› run_tests() → 4 passed, 0 failed

✓ all criteria satisfied · 2 iterations · 0 tokens

Backend

WebGPU

Token cost

$0.00

Egress

0 B

$0

per prompt, forever

100%

on-device inference

0 B

student data egress

1

click to launch in Canvas

Two ways to deploy

One platform. Inside your LMS, or on its own.

LMS Embedded Mode

Lives inside Canvas.

Plugs into Canvas, Moodle, or Blackboard as an interactive module via LTI 1.3 (with 1.1 fallback), keeping students inside their existing course workflow — with automated grade passback.

See the launch flow

Standalone Web Platform

Or runs on its own.

Operates independently in the browser without requiring a backend LMS — easy to share as a demo, run standalone workshops, or sell directly to institutions.

Open the sandbox

Hybrid deployment

Inside Canvas via LTI for enrolled courses, or standalone for workshops and demos.

Air-gapped privacy

Code, prompts, and evaluation data stay on the machine they were typed on.

Predictable budgeting

A fixed license instead of volatile per-prompt API bills.

Agentic pedagogy

Students define specs and evaluate behaviour instead of typing boilerplate.

LMS Integration

Plugs directly into your existing workflow.

Instructors install VaultEd once as an LTI tool. Students click the module inside their course and land in a working AI lab — already authenticated, already rostered, grades flowing back automatically.

  • Single-click LTI 1.3 deployment, 1.1 fallback for legacy tenants
  • Canvas, Moodle, and Blackboard supported out of the box
  • Grade passback via Assignment & Grade Services
  • Deep Linking places a specific lab into a module
  • Names & Roles rosters sync — zero onboarding friction
See the Canvas launch flow

LTI 1.3 launch sequence

  1. 1

    Student clicks the module in Canvas

    OIDC third-party init

  2. 2

    Platform posts a signed id_token

    JWT verified against JWKS

  3. 3

    VaultEd renders inside the course

    roster + context claims applied

  4. 4

    Model loads to the student's GPU

    WebGPU · cached after first run

  5. 5

    Score posts back via AGS

    only the grade leaves the device

The Perfect School Recipe

Purpose-Built for MacBook Neo & Modern Classrooms

VaultEd combines browser-native WebGPU AI with entry-level school hardware to deliver 100% private, zero-token AI learning.

On-Device Neural Engine Efficiency

Leverages the local Neural Engine and unified memory on devices like the MacBook Neo (and M-series Macs). Small language models (SLMs) run 100% locally in WebGPU without thermal throttling or needing high-end GPUs.

Zero Cloud Overhead

Plugs Right Into Canvas

Embed VaultEd directly inside your Canvas modules. Students launch interactive local AI sandboxes in one click — while grades automatically sync back to your gradebook.

Instant LTI Deployment

Built for Institutional Compliance

Student code, prompts, and work never leave the local laptop. Zero student data leakage, full FERPA/FIPPA compliance, and zero recurring API token invoices for school IT budgets.

Zero Data Leakage

Privacy & Security

100% air-gapped.
Zero data leakage.

On-device execution bypasses strict institutional data governance hurdles. There is no inference vendor to review, no data processing agreement to negotiate, and no cross-border transfer to disclose — because the data never leaves the laptop it was typed on.

FERPAFIPPAPIPEDAGDPRNo training on student data
Model weightsCached in browser storage after first load
InferenceWebGPU on the student's own GPU — never a server
Prompts & codeNever serialized off-device
Evaluation tracesStored locally; only the score is passed back
Compliance postureFERPA / FIPPA / PIPEDA — no third-party processor
Network at runtimeZero outbound inference calls

Cost Savings

Unlimited iterations. Zero cloud tokens.

High-performance 4B–7B parameter models run locally via browser WebGPU. A student can re-run an agent two hundred times debugging one spec, and the line item is still the same fixed license.

Term cost estimator

Cloud API · 13 weeks

$2,471

≈ 411.8M tokens

VaultEd · 13 weeks

$0

fixed license

Illustrative: ~2,200 tokens per agentic turn at a blended $6 / 1M tokens.

Bundled local models

ModelQuantVRAMSpeed
Qwen2.5-Coder 3Bq4f16~2.4 GB~42 tok/s
Llama 3.2 4B Instructq4f32~3.1 GB~35 tok/s
DeepSeek-Coder 6.7Bq4f16~4.6 GB~24 tok/s
Mistral 7B Instructq4f16~5.0 GB~21 tok/s

Measured on Apple M2 Pro, Chrome WebGPU. WASM fallback for unsupported devices.

2030 Agentic Pedagogy

From typing boilerplate to directing agents.

01

Specify

Students write the spec and the acceptance criteria before a line of code exists. The rubric grades the specification, not the keystrokes.

02

Direct

A local agent plans, edits, and runs tests inside the browser sandbox. Students steer it, interrupt it, and constrain its tools.

03

Evaluate

Students grade model behaviour against their own criteria — hallucination, regression, edge cases — and defend the verdict.

Deploying MacBook Neo or Chromebook fleets this Fall? VaultEd delivers agentic AI education with zero server costs and total student data privacy.

Talk to our education teamSee it inside Canvas

Pick your deployment mode.

Same platform, same runtime, same privacy guarantee — inside your LMS or on its own.

Launch Canvas SandboxTry Standalone Demo