Skip to content
MoorAI
Get MoorAI
MoorAIKnowledge & evidence
On this page
Review before the agent,redact when possible.Classify bydata tier.Governance withoutsurveillance.Prompt content, redaction,and coverage.Keep the secretson the device.
// data-loss prevention for AI

On-device AI DLP

On-device AI DLP is data-loss prevention that runs on the developer’s machine instead of in a cloud proxy. MoorAI checks what an AI coding agent is about to take in or send out — a prompt, a file it reads such as .env, a shell command, an MCP tool argument — and coaches, masks, requires a justification or blocks before secrets, personal data or source code reach Claude Code, Codex, Copilot CLI, Cursor or Gemini CLI. The content is inspected on the device, so it never has to leave the machine to be checked.

MoorAI overview →
// the flow

One local check on the device, before a prompt reaches the AI — then only content-free signals flow to the console.

One check, on the device — before the AI sees anything. // moorai · on-device AI guardrails Employee types a prompt MoorAI 85-threat matrix · content rules · local coach alert block justify prompt content never leaves the machine AI agent Claude · Codex Copilot CLI Model API cloud LLM Central console posture · inventory · compliance — content-free signals only // four pillars On-device DLP Prompts and outputs checkedlocally against policy — nothinghas to leave the endpoint. AI-SPM for endpoints Shadow-AI discovery plusagent, tool & MCP governancewhere the work happens. Human-in-the-loop Coach, alert, block, or requirea signed justification — setper policy, tenant, device. Continuous compliance OWASP LLM · MITRE ATLASNIST AI RMF · ISO 42001EU AI Act-aware.
// how it works

Review before the agent,
redact when possible.

The moment a prompt is sent, a file is read into the agent’s context or a tool call is about to run, MoorAI checks it locally against policy: a 85-threat matrix plus content rules. If it carries an API key, a private key, an email, a national ID, a payment card, or PHI, MoorAI can coach the user, mask the sensitive span and let the call proceed, require a logged justification, or block outright, per policy, tenant, and device. A mask replaces the value with a content-free placeholder such as [MOORAI:secret:…], derived from a keyed hash so no part of the value survives, and re-scans the rewritten call before it runs.

Masking rewrites Claude Code tool calls and tool results through the hooks’ updatedInput and updatedToolOutput; where a call cannot be rewritten, the configured fallback applies. Blocking applies on devices enrolled in a MoorAI console; elsewhere MoorAI coaches.

// classification

Classify by data tier.

MoorAI groups detections into four data classes, so you can set one default action per class instead of tuning every rule — and a per-threat setting still overrides it.

Personal data
PII
Emails, phone numbers, national IDs, passports.
Credentials
Secrets
API keys, OAuth tokens, private keys, cloud credentials.
Intellectual property
Source / IP
Roadmaps, architecture, trade secrets, licensed or copyrighted code.
Regulated
PHI / PCI
Health data (HIPAA) and payment-card data — with tighter defaults.
// privacy

Governance without surveillance.

The core invariant: by default, prompts and conversations never leave the device. Security teams receive only redacted, content-free signals (a category, a risk level, and a keyed one-way hash, plus the login and hostname, which the console pseudonymises on arrival) through a central web console. You get evidence of what class of data was at risk and what action was taken, without ever exposing what anyone typed. The one way content is retained is a capture tier an administrator turns on deliberately, per tenant, and that choice is logged.

// faq

Prompt content, redaction, and coverage.

What is on-device AI DLP?

On-device AI DLP is data-loss prevention for AI that runs on the endpoint instead of a cloud proxy. MoorAI reviews prompts, files read into an agent’s context and tool calls locally, and can coach, mask or block sensitive data — API keys, PII, source code, PHI/PCI — before it reaches the AI agent. Because review is local, the content never has to leave the machine.

How is on-device DLP different from cloud DLP?

Cloud DLP inspects data after it is routed through a service, so the content is seen off-device. MoorAI inspects on the endpoint, before the agent sees the prompt or the file, and reports only redacted, content-free metadata to the console — a category, a risk level, and a keyed one-way hash (plus the login and hostname, which the console pseudonymises on arrival).

What kinds of data does MoorAI detect?

MoorAI classifies data into tiers — PII, secrets/credentials, source/IP, and regulated data (PHI/PCI) — with detectors for API keys, tokens, private keys, emails, national IDs, payment cards, IBANs, ICD-10 codes, and more, in English and Hebrew.

Can it mask a secret instead of blocking the call?

Yes, for secrets, PII and regulated data on Claude Code. With the mask action MoorAI replaces the matched span with a content-free placeholder, re-scans the rewritten call and lets it proceed. Where a call cannot be rewritten, the policy’s fallback (notify, justify or block) applies.

Which coding agents does it cover?

Claude Code, where enforcement is validated end to end, plus Codex CLI, GitHub Copilot CLI, Gemini CLI and Cursor through their own pre-tool hooks; those four adapters are tested against each vendor’s documented hook payloads and not yet against the live agents.

Does the prompt content leave my machine?

No. Detection, masking and blocking happen locally; only content-free metadata is reported, and only from a device enrolled in a MoorAI console.

More: Claude Code security · MCP security · quick start · shadow-AI detection · OWASP LLM Top 10 tooling · MoorAI vs Lakera

Keep the secrets
on the device.

Redact or block before the prompt reaches the agent.

Community agent on GitHub