AI Footprint Calculator
A two-minute estimate of the AI tools, MCP servers, skills, extensions, personal AI accounts and OAuth grants running across your company, and how much of it your current controls can see.
Estimated AI footprint
2,034
AI tools, MCP servers, skills, hooks, extensions and OAuth grants across your fleet. 1,666 of them were never reviewed by anyone, and your current controls do not see 8 of 11 AI surfaces at all.
Endpoint Sensor
- AI tools, agents and desktop apps788
- of which unsanctioned420
- MCP servers349
- Skills and plugins233
- Agent hooks39
- Local LLM runtimes25
- Plaintext secrets in AI configs87
Browser Sensor
- AI web services in use53
- of which not sanctioned52
- Employees on personal AI accounts510
- Employees who have entered sensitive data427
- GenAI browser extensions200
Cloud discovery
- OAuth grants held by AI appsGoogle Workspace and GitHub400
What your current controls see
Typical coverage for each category of control, not a claim about any one product.
| AI surface | Your controls | Anomity |
|---|---|---|
| AI CLIs, agents and desktop appsEDR and MDM see the binary or app; neither reads its configuration. | Partial | Seen |
| MCP servers and what they can reachAn MCP server is a legitimate process doing what its config allows, so EDR sees nothing wrong. | Not seen | Seen |
| Skills, plugins and IDE extensionsThey live as files inside each tool's own config directory. | Not seen | Seen |
| Agent permissions and hooksAllow and deny rules sit in per-user settings files. | Not seen | Seen |
| Plaintext secrets in AI configsDLP sees a secret only if it leaves the machine, not where it sits. | Not seen | Seen |
| Local LLM runtimesThe process is visible; which models it serves and to whom is not. | Partial | Seen |
| AI web services in useWeb proxies see the domain. | Not seen | Seen |
| Personal vs corporate AI accountsNeeds TLS inspection or tenant restrictions per vendor. | Not seen | Seen |
| Secrets and PII pasted or typed into AIPaste and upload are often covered; typed text usually is not. | Not seen | Seen |
| AI browser extensionsBrowser management can list extensions, but not rate what they can read. | Partial | Seen |
| OAuth grants held by AI appsGrants live in Google Workspace and GitHub, not on the network or the device. | Not seen | Seen |
Email me this estimate
We'll send the numbers and show you how to replace the estimate with your real inventory. No spam.
How we calculate
The estimate multiplies your headcount and developer share by published adoption rates where they exist, and by our own stated assumptions where they do not. Every input to the model is in the table below, with its source. Where a row says Assumption, no primary source we trust measures it yet.
Developers drive most of the endpoint footprint: AI CLIs, IDE extensions, MCP servers, skills and hooks are configured per developer. Browser numbers scale with everyone who uses AI at work. OAuth grants scale with headcount and drop sharply when only admins can approve third-party apps.
Unsanctioned tools are the AI tools outside the ones you ticked. MCP servers, skills, hooks, extensions and OAuth grants count as unreviewed, because none of these inputs is an allowlist for them. The coverage table reflects what each category of control is built to observe, not any one vendor's product.
It is an estimate. Your real inventory will differ, which is the point of measuring it: the Endpoint Sensor, Browser Sensor and cloud discovery replace every number on this page with what is actually there, usually on the first day of a deployment.
| Input to the model | Value | Source |
|---|---|---|
| Developers who use or plan to use AI tools | 84% | Stack Overflow Developer Survey 2025 |
| Developers already using AI agents | 31% | Stack Overflow Developer Survey 2025 |
| Distinct AI tools per AI-using developer (CLI, IDE extension, desktop app) | 2.5 | Assumption |
| MCP servers configured per agent user (survey: most run 2 to 7) | 4.5 | Zuplo State of MCP report, via Nordic APIs (2026) |
| Skills and plugins installed per agent user | 3 | Assumption |
| Agent hooks configured per agent user | 0.5 | Assumption |
| Developers running a local LLM runtime (Ollama, LM Studio) | 10% | Assumption |
| MCP servers configured with a credential written into the config file | 25% | AssumptionGitGuardian found 24,008 unique secrets in public MCP config files in their first year. GitGuardian State of Secrets Sprawl 2026 |
| Non-developers with an AI desktop app installed | 35% | Assumption |
| Employees who use AI assistants at work | 75% | Assumption |
| AI users who reach AI assistants through personal accounts | 68% | TELUS Digital survey of enterprise employees (2025) |
| AI users who admit entering sensitive information | 57% | TELUS Digital survey of enterprise employees (2025) |
| Employees with a GenAI browser extension installed | 20% | LayerX Enterprise Browser Extension Security Report 2025, via The Hacker News |
| Distinct AI web services in use, per doubling of headcount beyond 10 | 8 | Assumption |
| OAuth grants held by AI apps per employee, where self-consent is allowed | 0.4 | AssumptionMaterial Security counted a median of 1,760 OAuth grants per company, and 91% of the AI and automation apps had appeared in the previous 16 months. Material Security OAuth & Google Workspace Risk Report (2026) |
| Share of developer AI use covered by each sanctioned coding tool (floor 30% unsanctioned) | 20% | Assumption |
Frequently asked questions
How accurate is the AI footprint estimate?
It is an order-of-magnitude estimate built from public survey data and stated assumptions, listed with sources under How we calculate. It is meant to size the problem, not to replace an inventory. Deploying the Endpoint Sensor and Browser Sensor replaces it with the real count.
Why can't EDR or MDM see MCP servers?
An MCP server is a legitimate process doing exactly what its configuration allows. EDR watches processes and binaries, and MDM inventories installed applications; neither reads the per-user JSON and TOML files where MCP servers, permissions, skills and hooks are declared.
What counts as an unreviewed AI artifact?
Any AI tool outside the ones you marked as sanctioned, plus every MCP server, skill, plugin, hook, GenAI browser extension and AI OAuth grant, since none of these is covered by an allowlist in the calculator's inputs.
Is any data I enter stored?
No. The calculation runs in your browser and your inputs stay in the page address so you can share the result. Nothing is sent anywhere unless you fill in the form to email yourself the estimate.
How do I get my real numbers?
Get a demo. Anomity's Endpoint Sensor deploys through the MDM you already run, the Browser Sensor through browser policy, and cloud discovery connects to Google Workspace and GitHub, so the first inventory usually lands within a day.
