Your team is already using AI - dozens of tools, every day, mostly without telling IT. Northbeams discovers them, scores the risk, and gives you one place to govern it. No new behaviour to enforce. No data to ship to a third party. One afternoon to install.
Northbeams is a browser extension and a Mac/Windows app, paired to one dashboard, that catches every AI tool your team actually uses - across the browser, desktop apps like Claude Desktop and Cursor, CLI tools like Claude Code and Aider, and the MCP servers your coding agents call - then helps you govern it from one screen.
Your team isn't waiting for IT to approve their AI. They're using ChatGPT, Claude, Gemini, Copilot, and 20-odd other tools you've never been told about. Their coding agents are calling MCP servers wired up to GitHub, Postgres, Stripe, and the corporate filesystem. Today, you have no way to see this. Tomorrow, when sensitive data leaks - a client contract pasted into a chatbot, a customer table read by an agent over MCP - you'll find out from a regulator or a customer, not from your stack.
Northbeams sits where the work actually happens. The browser extension sees AI tools opened in the browser. The Northbeams app for Mac and Windows watches outbound connections and process names on the laptop - so desktop apps (Claude Desktop, Cursor) and CLI tools (Claude Code, Aider) show up too. The new MCP Gateway sits in the path between coding agents and the MCP servers they call, classifies the arguments on-device, and lets you allow, warn, or block per tool. One platform, four surfaces, one dashboard. The original prompt text never leaves the user's device. Neither do MCP argument values.
The MCP Gateway is a local stdio proxy that sits between your coding agents (Claude Desktop, Cursor, Claude Code) and the MCP servers they call. It classifies every tool argument on-device, ships only categorical labels and a hash to your dashboard, and lets you set per-tool allow / warn / block rules per MCP server.
Coding agents have grown teeth. The same Cursor window that writes a unit test can also call delete_file on the GitHub MCP, create_refund on the Stripe MCP, or query on a Postgres MCP wired to a production replica. None of that was visible to your DLP yesterday. Now it is.
Install once via the Northbeams desktop app. The Gateway scans your laptop's Claude Desktop, Cursor, and Claude Code configs, wraps each MCP server with an atomic, idempotent rewrite, and keeps a timestamped backup. Disable in one command and the wrap unwinds cleanly. No new install, no new dashboard, no new procurement cycle. Sentinel and Fleet only.
The minute someone in marketing tries a new "AI writer", or someone in engineering pastes a stack trace into a free chatbot, it's on your dashboard - by tool, by user, by category. No more reading about a vendor in a board meeting.
The on-device classifier flags credentials, PII, source code, customer data, and contracts. Each flag carries a severity and an estimated exposure value. You walk into the next risk review with a number, not a feeling.
Mark a tool Sanctioned, Unknown, or High risk and Northbeams enforces it in the browser. Allow ChatGPT Team. Block consumer ChatGPT. Sandbox the new tool finance just discovered. The MCP Gateway extends the same three states down to the MCP tool level: allow read_file, warn on create_pull_request, block delete_file. Roll back in one click if a team yells.
EU AI Act readiness checks, SOC 2 control mappings, tool-by-tool incident history, CSV exports. When the auditor or your biggest customer's security team asks "how do you govern AI?", you have a 30-second answer and a 30-page export.
A · Classifier
The classifier runs locally. Original prompt text never leaves the device. The dashboard sees category labels and a redacted snippet only.
B · Policy editor
Mark each tool Sanctioned, Sandboxed, or Blocked. Northbeams enforces the policy in the browser. No procurement cycle, no admin escalation.
C · Audit log
Immutable signed CSV exports map directly to SOC 2, EU AI Act, and HIPAA controls. Hand it to the auditor and answer the AI-governance question in 30 seconds.
One-click extension across Chrome, Edge, Brave, Arc - plus the Northbeams app for Mac and Windows. The desktop app brings the MCP Gateway with it on Sentinel and Fleet. Push both via MDM (Intune, Jamf, Kandji, Google Workspace). No certificates, no proxies, no DNS rewrites.
Northbeams catalogs every AI tool, model, plugin, and integration touching your environment. Each one is named, categorized, and risk-scored automatically.
The classifier runs inside the user's browser. It detects credentials, PII, source code, customer data, contracts. Only category labels and a redacted, secret-masked snippet hit your dashboard.
Mark each tool Sanctioned, Unknown, or High risk. Block, allow, or sandbox. Roll back in one click. Export an audit log for the next compliance review.
Nothing on this page requires a procurement cycle, a network change, or a re-architecture. The whole flow above runs without ever touching the data plane your security team is already trying not to break.
The point of governance isn't to slow your company down - it's to let it actually use AI without paying for the next $670K incident. Here's what changes the day after you turn Northbeams on.
The average shadow-AI-driven data breach costs $670,000. One year of monitoring 200 users runs roughly $12K. The math is not subtle.
You don't have to ban AI to be safe - you have to see it. Sanction the tools your team needs, sandbox the rest, kill only the genuinely dangerous ones. The work keeps moving.
Tool inventory, risk classification, incident logs, retention controls - exported as CSV or shown live to an auditor. Article 4 readiness mapped, SOC 2 evidence pre-built.
"How do you govern employee use of AI?" used to be a deal-stopper. With Northbeams it's a one-paragraph answer with attached evidence - the same paragraph that lands you the enterprise contract.
The classifier runs entirely inside the user's browser. The dashboard only sees category labels and redacted snippets - never the raw prompt. Privacy and visibility, both.
This isn't a six-month CASB rollout. Push the extension, sign in, watch the dashboard fill. Most companies have their first AI Discovery Report inside an afternoon.
Shadow-AI breaches add $670K of premium to an incident, plus $160 per leaked record. Northbeams is the line item that takes that off your risk register.
You inherited "AI governance" without a budget or a team. You need a defensible answer to "what AI are we using?" by Monday.
You need evidence of inventory, classification, and enforcement, in a format the auditor will sign off on without a follow-up call.
You want AI to make the company faster without putting the next funding round, the next enterprise customer, or the next regulator at risk.
Composite, based on customer interviews.
We don't watch keystrokes. We don't read non-AI browsing. We don't log Slack DMs. We classify AI prompts on the device, before they leave - and report categories, never raw text, to your dashboard.
The fastest way to wreck a security rollout is to make it feel like spyware. The classifier only fires on input to known AI surfaces. Outside of that, it does nothing. No DOM hoovering, no clipboard listening, no idle telemetry.
When a policy fires, the user sees it: "this tool isn't approved for source code, use Copilot instead." Not a silent disciplinary log. They get their own AI usage dashboard too, so the system stays in sunlight.
The point is to stop your team from accidentally pasting the customer list into a free summarizer. Not to grade them on output.
If your works council, ER team, or general counsel needs the formal version, we have a one-page memo and a sample employee-comms template. Email hello@northbeams.com.
CASBs, proxies, and DLP sit on the network and inspect packets. They miss anything off the corporate network, anything inside a native AI app, and anything on a personal laptop. Northbeams runs on the device itself - browser extension plus Mac and PC desktop apps - so it catches the prompt where it's typed, not where it lands.
Every workspace starts on Sentinel for 14 days, free. After that, drop to Beam (aggregate-only visibility for up to 10 users, runs 60 more days), or subscribe to Lighthouse / Sentinel to keep per-user attribution, blocking, and audit-ready logs. No card required at signup.
Most workspaces have their first AI Discovery Report inside 24 hours.