Observability for the agentic workforce.
Signal, not surveillance. See how your engineers use Cursor, Claude Code, and Copilot — catch leaked secrets, PII, and confidential data, without blocking them or reading their prompts.
- Manager sees
- Signal, not content
- Prompts leave device
- Never by default
- Rollout
- MDM / Jamf
The problem
The fastest enterprise tool adoption in history shipped without a control plane.
Cursor, Claude Code, Copilot, and Windsurf became default stack in under 24 months. Nobody sees what goes into the prompt — not the engineer in the moment, and not the leader accountable for the whole team. HeimWall gives you that visibility, on-device, without reading a single prompt.
Samsung 2023 quietly repeats every day
Engineers paste API keys, customer records, and proprietary code into Cursor and Claude Code by accident. Most never know. Managers definitely don't.
Classic DLP breaks the engineer workflow
Nightfall, Symantec, and Purview were built for email and file uploads. Agentic chat is an entirely different surface. Block rules kill velocity. They get turned off.
Native controls only see their own tool
Cursor can tell you about Cursor. Copilot about Copilot. Your team uses four tools this quarter, and nobody owns the cross-tool signal. HeimWall watches all of them from one on-device agent.
How it works
Capture, detect, redact — all on your machine.
The whole loop runs on-device. Raw prompts never leave the engineer’s machine; the manager dashboard only ever sees redacted signal.
- 01
Capture on-device, before the network
A 15MB menu-bar agent reads the composer itself — via macOS Accessibility, a clipboard poll, and a local TLS proxy — so it sees the prompt even when the tool talks straight to its own cloud. Nothing is written to disk.
- 02
Tiered detection in under 50ms
25+ hand-written regex rules catch the hard-shaped things — AWS keys, tokens, private keys, SSNs — deterministic and explainable. A pretrained on-device classifier is the fallback for fuzzy leaks. We never train on your prompts.
- 03
Redacted at the source
The moment a match is found, the value is masked before any record is written. What could ever leave is a category, a severity, and a masked snippet — plus a hash so the same secret seen twice is correlated without storing it.
- 04
Manager sees signal, not content
For teams, redacted metadata becomes a Safety Score, a category breakdown, and a trend. Raw prompt text never leaves the machine. Opening a sealed payload needs 2FA, written justification, and employee notification.
Principles
What we will not compromise on.
- We observe. We don't surveil.
- Observability is the product. Coaching is the outcome. Investigation is a last resort, gated by 2FA, justification, and employee notification.
- Friction budget is sacred.
- Soft notifications beat hard blocks in 19 out of 20 cases. A three-second toast at the source teaches. A modal interrupts and gets bypassed.
- Manager sees signal, not content.
- Dashboards show categories, trends, and safety scores. Raw prompt text never leaves the engineer's machine by default. Contractually, not just culturally.
- Model- and tool-agnostic.
- Cursor today, Claude Code tomorrow, whatever ships next quarter. One agent. One dashboard. One policy surface.
For teams · early access
Bring HeimWall to your team.
Signal, not surveillance — a Safety Score and cross-tool trend for every engineer, without ever reading a prompt. Rolls out over MDM/Jamf. Tell us your role and we’ll line up early access to the manager dashboard.