# HeimWall

> On-device observability for engineering teams using AI coding tools — catches leaked secrets, PII, and confidential data in Cursor, Claude Code, Copilot, and Windsurf without reading prompts.

HeimWall is a macOS-first security observability tool built for engineering teams that have adopted AI coding assistants like Cursor, Claude Code, GitHub Copilot, and Windsurf. It runs a lightweight on-device agent that inspects prompts locally before they leave the machine, detecting leaked secrets, PII, and proprietary code — then surfaces only redacted metadata to a manager dashboard. The product is currently in early access, with a free solo tier available and team/business tiers on a waitlist.

## What It Is

HeimWall sits in the category of AI-era data loss prevention (DLP) and developer observability. Unlike traditional DLP tools built for email and file uploads, HeimWall is purpose-built for the agentic chat surface — the composer windows inside tools like Cursor and Claude Code. Its core job is to give engineering leaders cross-tool visibility into what categories of sensitive data are being pasted into AI prompts, without ever transmitting raw prompt text to the cloud. The company frames this as "signal, not surveillance."

## How the On-Device Architecture Works

The entire detection loop runs on the engineer's machine:

- A 15 MB menu-bar agent reads the composer via macOS Accessibility APIs, a clipboard poll, and a local TLS proxy — capturing prompts even when the tool communicates directly with its own cloud.
- A tiered detection engine runs in under 50ms (p95 on M1 MacBook Air): 25+ hand-written regex rules handle hard-shaped secrets (AWS keys, tokens, SSNs, private keys), while a pretrained on-device ML classifier handles fuzzy leaks as a fallback.
- Matched values are masked before any record is written. What can ever leave the device is a category label, severity, masked snippet, and a SHA-256 hash for deduplication — never raw prompt text.
- The manager dashboard receives only this redacted metadata, showing a Safety Score, category breakdown, and trend lines per engineer.

## Privacy and Investigation Controls

HeimWall's stated design principle is that raw prompt content never leaves the engineer's machine by default — and this is described as a contractual commitment, not just a cultural one. The only path to raw text is Investigation Mode, which requires a second-factor step-up, a written justification of at least 50 characters, automatic notification to the affected employee within one hour, and a 24-hour time-box. Every action in Investigation Mode is audit-logged. The company also states by contract that safety scores and flag history cannot be used in performance reviews, promotion, compensation, or termination decisions.

## Platform and Deployment

HeimWall is macOS-first, supporting Apple Silicon and Intel Macs running macOS 13 and later. Fleet rollout is designed for MDM systems (Jamf, Kandji, Intune) with a signed, notarized package. Individual installs use a standard DMG. The FAQ notes Windows 11 support is on the roadmap, with Linux planned after that. Metadata is currently stored in AWS us-east-1 (Supabase + ClickHouse Cloud), with EU region support and on-premises deployment planned for a future v1.5 release targeting month 12+.

## Supported AI Tools and Extensibility

At launch, HeimWall instruments Cursor, Claude Code, GitHub Copilot, and Windsurf. The detection engine is shared across tools, and the company states that adding coverage for a new agentic tool is typically a one-week release cycle. The Business tier adds SIEM export (Splunk, Datadog) and Slack + PagerDuty integrations. The Enterprise tier adds on-premises deployment, bring-your-own-keys (BYOK), SOC 2 Type II, and HIPAA eligibility.

## Current Status

HeimWall is in early access as of 2026. The solo free tier is available now; team and business tiers are on a waitlist. SOC 2 Type II audit is described as in progress, with a target completion in "Hafta 23" of the company's internal roadmap. On-premises deployment and HIPAA eligibility are targeted for v1.5 at month 12+. The company describes itself as a small team that answers its own email, and sales inquiries go directly to the founders.

## Features
- On-device prompt inspection — raw text never leaves the engineer's machine
- 25+ regex rules for secrets detection (AWS keys, tokens, SSNs, private keys)
- On-device ML classifier for fuzzy/contextual leak detection
- Detection latency under 50ms (p95 on M1 MacBook Air)
- Manager dashboard with Safety Score, category breakdown, and trends
- Cross-tool coverage: Cursor, Claude Code, Copilot, Windsurf
- Redacted metadata only — category, severity, masked snippet, SHA-256 hash
- Investigation Mode gated by 2FA, written justification, and employee notification
- MDM/Jamf fleet rollout support
- SSO (Google, Microsoft) for Team tier
- SAML SSO + SCIM provisioning for Business tier
- Custom policy rules (Business+)
- SIEM export to Splunk and Datadog (Business+)
- Slack and PagerDuty integrations (Business+)
- On-premises deployment (Enterprise roadmap)
- BYOK — bring your own keys (Enterprise)
- SOC 2 Type II (Enterprise, in progress)
- HIPAA eligibility (Enterprise)
- Contractual prohibition on using safety scores in performance reviews
- Fails open — prompts pass through if agent crashes

## Integrations
Cursor, Claude Code, GitHub Copilot, Windsurf, Jamf, Kandji, Intune, Google SSO, Microsoft SSO, SAML, SCIM, Splunk, Datadog, Slack, PagerDuty, Supabase, ClickHouse

## Platforms
WINDOWS, MACOS, LINUX, WEB, API

## Pricing
Freemium — Free tier available with paid upgrades

## Links
- Website: https://heimwall.ai
- Documentation: https://heimwall.ai/how-it-works
- EveryDev.ai: https://www.everydev.ai/tools/heimwall