Concept app · v0.2.0 beta · Notarized

Heimdall

Keeps watch over your Mac. Checks memory, CPU, heat, battery, disk, network and device management, explains what it finds in plain words, and answers questions — with an AI model that runs entirely on the Mac. Deploy your IT guides and it answers from them, naming the document.

Concept app. Heimdall explores whether a small on-device model can make Mac diagnostics understandable for end users and IT support — with a rule engine deciding what's wrong and the model only explaining it. It's a beta, not a supported product: the Jamf Pro and Intune checks were built against documented log formats and tested with sample logs; 0.2.0 fixes issues found in the first tests on a managed Mac, and has had no wider testing. All testing so far was in a demo environment, not in production.

Download Heimdall (2.4 GB) Read the Docs

Heimdall overview: 3 things to look at (low disk space, CPU overloaded, not managed by an MDM) and a list of checks with no problems found

Features

Diagnostics people can actually read

🩺

Checks Without AI

A deterministic rule engine evaluates the Mac and lists findings with severity (critical, warning, info), explanation, recommendation and the measured values in human units.

💬

Ask in Plain Language

“Why is my Mac slow?”, “Is my battery OK?”, “Is this Mac managed?” The on-device model explains the rule engine's findings — it never invents its own. Every question carries the current results, so it doesn't ask users to run commands or paste logs. It introduces itself as Heimdall and declines requests unrelated to the Mac or IT.

🧠

On-Device Model

Gemma 4 E2B (text-only) is bundled in the app and runs via MLX. Loaded on the first question, released after 5 minutes idle. Load Anyway when memory is short.

🖥️

System Checks

Memory pressure, swap, memory and CPU per app, CPU load, thermal throttling via kernel_task, temperature, battery health and cycles, disk space, uptime, network, Low Power Mode.

🏢

Device Management

MDM enrollment and user approval, Automated Device Enrollment, MDM profile signature, MDM errors in the system log, Jamf Pro and Microsoft Intune agent health. The managing vendor is read from the MDM profile, so a Jamf-managed Mac with Company Portal no longer gets an Intune warning. Kandji, Workspace ONE, Addigy and Mosyle agents are detected.

📚

Knows Your Organization

Deploy IT guides (Markdown, text, PDF, Word, RTF, HTML) and Heimdall searches them locally, answers from the matching sections and names the document. A support contact you set is added when an answer recommends contacting IT.

🗂️

Conversations

Saved per session on the Mac (newest 50 kept), with New Conversation and a Conversations menu. The model sees the last 10 messages.

⚙️

MDM-Configurable

Managed Preferences can force the on-device model, point Heimdall at an OpenAI-compatible server, set the documentation folder and the IT support contact. Forced settings show as locked. A Jamf Pro schema and example profile are included.

📦

Easy to Deploy

A signed, notarized package that installs only the app — no LaunchAgents, no daemons, no root helper, no system extensions.

🔒

Private by Design

No telemetry, no account, no cloud. Data leaves the Mac only if an administrator configures an external server.

Status

Where things stand

System Requirements

macOS 26 or later, Apple Silicon.

8 GB RAM or more — the model needs about 2.7 GB of free memory while answering.

About 2.5 GB of disk space.

Build Status

Version 0.2.0 — beta. Signed with a Developer ID certificate and notarized by Apple.

The rule engine, MDM probe, documentation search and chat sessions are covered by unit tests (64 passing).

What's Verified

System checks and the bundled model were verified live on an 8 GB M1. 0.2.0 follows up on the first tests on a managed Mac in a demo environment; nothing has been tested in production.

The Jamf Pro and Intune checks were built against documented log formats and unit-tested with sample logs; vendor detection by MDM profile, the documentation feature and the support contact have not had wider testing across tenants.

Coming Next

More checks (login items, crash reports, Spotlight, Time Machine, software updates), trends over days, and MDM-configurable thresholds.