A searchable record of nearly everything
Screenpipe turns a computer into a personal activity archive. It watches operating-system events, captures changed screens, reads accessibility trees, falls back to OCR, records microphone and system audio, transcribes speech, and stores the result locally. A timeline lets the user browse the day, while full-text and natural-language search can retrieve a window, conversation, or meeting from the archive.
The event-driven design is smarter than saving video continuously. An app switch, click, scroll, or typing pause can trigger a frame, while idle periods need fewer captures. Accessibility data gives structured text without OCR when an application exposes it. The README estimates typical CPU use of 5 to 10 percent and 0.5 to 3 GB of memory. That is still a permanent background workload, but it is plausible on a modern work machine.
Agents make it more than a recall app
A local REST API exposes search results, frames, audio, elements, health information, and pipe management. The JavaScript SDK wraps that interface. Screenpipe also runs an MCP server, so Claude Desktop, Claude Code, Cursor, and other compatible clients can query recent screen history and meeting transcripts. Registration with Claude is one command after recording starts.
Pipes are scheduled agents defined in Markdown. Built-in examples create meeting summaries, daily recaps, stand-up updates, time breakdowns, prompt journals, and video exports. A pipe can query the local archive, call APIs, write files, and take actions. This closes the gap between remembering work and acting on it.
The permission design deserves credit. Pipe frontmatter can allow or deny applications and windows, restrict content types, limit time ranges and days, and block raw SQL or frames. The README says restrictions are enforced through skill exposure, agent interception, and server middleware using per-pipe tokens. That is a better boundary than asking a model to obey a sentence in its prompt.
One related feature is not ready for casual use. High-priority issue 5496 describes continuous memory synchronization into global Claude Code and Codex instruction files. A long-running profile produced more than 1,600 eligible memories; the selected entries would occupy roughly 60,000 to 75,000 tokens. Existing Codex and Claude files exceeded 160 KB and remained after the connection was removed. MCP and skills are the safer initial integration until memory writes are bounded, previewable, and reversible.
Local-first does not mean risk-free
The archive is an exceptionally sensitive dataset. Screens can contain passwords, customer records, health information, private messages, source code, and financial details. Audio can capture other people. Keyboard input adds another category of exposure. Local SQLite storage limits routine cloud disclosure, but malware, an unlocked session, an over-permitted pipe, or an unencrypted backup could reveal more than a normal document folder.
Screenpipe offers app, window, browser, password, and PII filters plus optional encryption at rest. macOS users can exclude selected applications from system-audio capture after enabling the experimental CoreAudio path. These controls need deliberate setup and periodic testing. Recording conversations also requires compliance with workplace policy and applicable consent rules.
The default is not completely offline. The FAQ says PostHog analytics is enabled and uses a stable installation identifier; signed-in use may associate an email with device and feature metadata. Sentry receives diagnostics while telemetry remains on. Cloud transcription, hosted AI, and encrypted sync process the information required for those services remotely. Users seeking the smallest network surface should disable analytics, leave sync off, choose local Whisper and a local model, then monitor outbound traffic themselves.
Storage claims need clarification. The short specification says about 20 GB per month, while the detailed reference and FAQ say roughly 5 to 10 GB. Actual usage depends on screen changes, monitors, audio, and retention. Either way, continuous history needs disk monitoring, deletion rules, and a backup decision.
The license is a purchasing constraint
Screenpipe changed to a commercial source-available license. Personal non-commercial, nonprofit, educational, and research use can be free. An organization receives only a seven-day evaluation window for development and testing. Business or production use, revenue-supporting use, product integration, hosted service operation, and commercial distribution require a paid license.
The license also says the licensor retains rights in modifications and patches. Official signed builds are governed separately by subscription terms, starting at $25 per month according to the README. Five or more users in one company require Team or Enterprise plans, and shared or unattended machines need a commercial agreement. Developers should not mistake visible source and an npm command for permission to adopt it freely at work. Earlier MIT-licensed versions retain their old license, but current code does not.
Setup, health, and the decision
The desktop application is the easiest route and includes automatic updates. The CLI starts with npx screenpipe record; setup can then connect agents. macOS and Windows receive native installers. Linux is supported through a source build. Every platform still needs sensitive operating-system permissions, and local transcription adds model and compute demands.
The repository was pushed on August 10, 2026. GitHub lists app-v2.5.176 from August 4 as the latest release, and 158 open issues and pull requests combined. That release fixed macOS capture stalls and enterprise sync data loss, among other AI and interface changes. Current issue traffic includes audio transcription gaps and a scheduled-task view that wrongly reports the backend unavailable while capture continues. Fast work is obvious, but the README itself warns that the main branch moves quickly and can break.
Screenpipe is a compelling personal memory system for someone who understands the privacy bargain. It is harder to recommend as a casual utility or an assumed open-source building block. Start with a short retention window and a few allowed applications, verify local-only settings, and expand capture only after the archive proves useful enough to justify its cost and risk.