Security Whitepaper · v1.3 · April 2026

Screenpipe Security Architecture

Local-first data architecture. Source-available Rust capture engine. On-premise AI options. Enterprise policy controls. Source-linked review paths for privacy-sensitive deployments.

1. Security Model

Screenpipe is local-first by design. All screen captures, audio transcriptions, and metadata can be processed and stored on the user's device for core local-mode use. Optional sync, cloud AI, connector, export, and team workflows should be scoped per deployment. The capture engine is written in Rust (13 crates), providing compile-time memory safety — eliminating buffer overflows, use-after-free, and data races that affect C/C++ recording tools.

2. Vision Pipeline Architecture

The vision pipeline uses an event-driven capture model — not continuous polling. Captures are triggered by user activity (app switch, click, typing pause, scroll stop, clipboard change) or visual change detection (histogram diff > 5%). This reduces resource usage while maintaining complete coverage. event_driven_capture.rs

3. Audio Pipeline Architecture

The audio pipeline captures from multiple devices simultaneously, processes 30-second segments with 2-second overlap, runs voice activity detection, speaker diarization, and transcription — all on-device. audio_manager

4. Cryptography

All cryptographic primitives use audited libraries. Parameters are from source: crypto.rs and team-crypto.ts.

5. Data Controls & PII Protection

Capture Controls

PII Detection Engine

Regex-based PII detection engine with 27 pattern categories. Uses RegexSet for single-pass detection. No ML — deterministic, auditable pattern matching. pii_removal.rs

Scheduled Task Permission System

Each scheduled task runs with configurable access controls. Evaluation order: Deny → Allow → Default → Reject. Deny rules always take precedence. permissions.rs

6. Speaker Identification

Screenpipe includes on-device speaker diarization — all processing runs locally with no cloud dependency:

7. Enterprise Deployment

MDM Configuration

// Pushed by MDM to: <app_dir>/enterprise.json
// Or manually at: ~/.screenpipe/enterprise.json
// macOS also checks: ../Resources/enterprise.json
{
  "license_key": "your-enterprise-license-key"
}

Write-Only Archive & Query Gateway

In a write-only archive deployment, devices write capture batches to your own S3 bucket using put-only credentials. Nothing reads that bucket except a query gateway you run on your own infrastructure. Screenpipe servers never hold your capture data. The gateway does, however, answer one question on every request — is this bearer token still authorized? — and it answers from an Ed25519-signed access policy issued by our control plane. That policy has a lifetime, and it is the one place where your archive availability depends on us.

Telling clock skew apart from an outage

The policy timestamps are stamped by our clock and evaluated against your gateway clock, so disagreement between them has two distinct failure modes and both are yours to fix. A gateway clock more than 5 minutes ahead of ours treats a freshly delivered policy as already expired and returns 503 on every query while we are perfectly healthy — a phantom outage. A gateway clock more than 5 minutes behind ours is the security-relevant direction: policies look not-yet-valid, and a revoked token survives that much longer than the stated 5-minute revocation latency. Either way the fix is NTP on the gateway host, so the gateway names the cause in the 503 body rather than letting a local clock problem present as a Screenpipe outage:

Bearer verification is opt-in

Everything above applies only when the operator pins the policy signer public key via SCREENPIPE_GATEWAY_POLICY_PUBKEY_B64. With no pinned key the gateway serves the entire v1 query surface unauthenticated and warns loudly at startup — a posture that is only acceptable on an isolated network. Read the guarantees above as describing a configured deployment, not a default.

What a Screenpipe shut-off can and cannot do

8. Network Requests

Core functionality requires zero network requests. Capture, OCR, local transcription, search, and scheduled tasks all run offline. Network requests occur only for explicitly enabled optional features:

Enterprise deployments can run fully air-gapped with local transcription (Parakeet/Whisper) and local AI (Ollama). In this configuration the only network request is license validation, which supports 14-day offline operation.

9. AI, Transcription & Confidential Compute

Screenpipe supports fully on-premise AI for both transcription and scheduled task execution. Enterprise users are free to use their own AI providers. Cloud AI through Screenpipe is optional and operates under zero data retention (ZDR) policies.

On-Premise Transcription

The following speech-to-text engines run entirely on-device with no network dependency. engine.rs

AI for Scheduled Tasks

Scheduled task execution supports multiple AI providers. Enterprise users can bring their own API keys or use fully on-premise models. pipes/mod.rs

Confidential Cloud AI — Hardware Enclaves

Zero-data-retention is a policy (the provider promises not to log). Confidential compute is a cryptographic guarantee (the provider cannot read the data, even if compelled to). Select Screenpipe Cloud models run inside hardware enclaves so that Screenpipe, its infrastructure provider, and the underlying cloud vendor are all excluded from the trust boundary — only the user and the attested model binary can see the plaintext.

Models & Roadmap

Confidential compute is live today for chat models. Additional workloads are rolling out:

Infrastructure Partner

Confidential inference is operated in partnership with Tinfoil, a confidential-computing platform for AI workloads. Tinfoil provides the attested container runtime, hardware-root-of-trust attestation chain (Intel DCAP + AMD KDS + Nvidia CC), HPKE transport, and the open-source verifier. Their verifier is published under AGPL-3.0 and embedded in the Screenpipe SDK — every session's attestation can be independently checked against Tinfoil's supply-chain measurements. Screenpipe curates the models; Tinfoil handles the cryptographic machinery. Neither party holds decryption keys for user sessions.

Further reading: tinfoil.sh · Tinfoil Containers overview · open-source verifier · RFC 9180 (HPKE) · Intel TDX whitepaper · AMD SEV-SNP whitepaper

Enterprise recommendation: For maximum data control, deploy with on-premise Ollama or a custom OpenAI-compatible endpoint. No data leaves your network. If cloud AI is required (e.g., for frontier model quality), the confidential tier provides the strongest available guarantees short of self-hosting.

10. Testing & CI/CD Pipeline

Screenpipe maintains a multi-layered testing infrastructure across 14 CI/CD workflows, covering unit tests, integration tests, E2E tests, benchmarks, security audits, and longevity testing.

Security-Specific Testing

11. Compliance

12. Source Code Audit

Screenpipe makes its source available under the Screenpipe Commercial License. The following modules are directly relevant to security review:

For documentation, see the Screenpipe Docs including the Scheduled tasks guide, Scheduled Task Permissions, Teams & Encryption, and Cloud Sync Architecture.

Security Contact

To report a vulnerability, request a security review for your organization, or discuss enterprise deployment: louis@screenpi.pe

Document version 1.3 · April 2026 · Screenpipe v2.4.x · Source-linked architecture notes for security review