How does Privent protect my n8n workflows without disrupting them?

Privent runs as a native node inside your n8n workflow. It checks the data your agents are about to send, masks anything sensitive, and passes a safe version to the model. Your workflow keeps running. No proxy, no rebuilding.

Does Privent block my agents or slow them down?

No. Privent transforms the data instead of blocking, so the model still receives useful input, returns a useful output, and the workflow never stops.

Can Privent handle PHI under HIPAA?

Yes. PHI is masked before it reaches any external model and raw patient data is never stored. Every event leaves an audit trail that supports HIPAA technical safeguards. For full isolation, the entire stack deploys on-prem.

Which agent frameworks does Privent support?

n8n is available today via our official node, n8n-nodes-privent. The @priventai/mcp-server package gives any MCP client our tokenize, risk, and audit tools, and the @priventai/mcp-gateway package tokenizes MCP tool calls in front of any stdio server you run. LangGraph and CrewAI are next on the roadmap.

Can we deploy Privent on-prem?

Yes. The full stack, AI models included, deploys inside your network or private cloud. Nothing leaves. Most teams start on Privent Cloud and move on-prem as they scale.

How is Privent different from DLP tools or AI gateways?

Traditional DLP sees only the final prompt string from the outside. Gateways pass through or block, but they don't transform data in place inside the workflow. Privent runs inside the workflow, sees everything the agent has gathered, and rewrites risky data so the pipeline keeps working.

Documentation

Two solutions. Two guides.

Everything you need to deploy Privent inside your n8n workflows or across your organization's browsers.

Agent Security · n8n

Protect your n8n workflows

01Install n8n-nodes-privent from Settings → Community Nodes
02Add your PriventApi credential (API key, base URL)
03Drop Session, Tokenize, and Detokenize into the canvas

LangGraph · CrewAI · MCP · SDK (coming soon)

Open the n8n guide

AI Monitoring

Roll out the monitoring extension

01Push the extension via MDM: Google Workspace or Intune
02Paste your managed config (apiKey + deploymentGroupKey)
03Detection events appear in the dashboard in real time

Open the extension guide

Both solutions deploy fully on-prem: we install the full stack, ML inference included, inside your network.

See pricing

Data & Privacy

Designed to analyse without retaining. Your prompts never leave the detection pipeline as plaintext.

What we store

ACARS composite score and per-signal breakdown
Decision outcome (allow / warn / block) and APE stages applied
Risk category (PII, Financial, Source Code, Legal, Strategic)
Timestamp, deployment group, and org policy snapshot
Execution graph node identifier and framework type

What we never store

Raw prompt or agent payload text, processed in-memory only
Tool call outputs or retrieved document contents
End-user or employee PII or identity information
The text behind any stored embedding: learned-block vectors keep category and risk metadata, never the content they derive from
Any data that crossed the agent boundary, only pre-boundary state is read

Prompt text is processed ephemerally in-memory for the duration of the scoring request, then discarded. It is never written to disk, never indexed, and never used for model training. For maximum control, the entire pipeline (ML inference included) is also available as a fully on-prem deployment, so even this ephemeral processing never leaves your network.

Read the full privacy policy →

Document jobs

The same tokenization, applied to whole files. Async jobs through one API, backed by the same vault and session as text.

The job API

–POST /v1/documents/jobs creates a job; progress comes from polling the job status
–Pipeline stages: Queued → Extracting → Scoring → Tokenizing
–Terminal statuses: Completed, Completed with warnings, Failed
–"Completed with warnings" is its own terminal state, never collapsed into success or failure
–Every download of a tokenized result is recorded on the access trail
–Jobs appear in the session audit view beside text and tool-call events

Files, tokens, and storage

–docx, xlsx, and PDF, capped at 25 MB per file
–Scanned PDFs are read with OCR as an opt-in engine build; models are mounted locally, never fetched at runtime
–The same reversible [KIND_NNN] tokens, the same vault, the same session as text
–Tokenized output lands in object storage you choose, local disk by default, never in the database
–Stored artifacts expire after 24 hours by default

The pipeline boots fail-closed: if the document engine is missing or stubbed, a production deployment refuses to start instead of silently skipping scans. The whole path can run fully air-gapped, and the engine makes zero outbound calls. In MCP clients, document tools ship behind preview flags.

Read the security overview →

FAQ

Frequently asked questions

Deployment, privacy, and how Privent protects GenAI prompts.

Never. Raw prompt and agent payload text is processed in memory only; detection events keep metadata: risk score, category, decision, timestamp. When the LLM judge confirms a block and a vector store is configured, an embedding vector derived from the content (never the text itself) is stored with category and risk level. Self-hosted deployments that don't configure a vector store persist nothing beyond event metadata.