Screenpipe is a source-available local capture engine (YC S26) that continuously records a computer’s screen, audio, and meetings into a searchable local timeline that agents query over MCP and a localhost API.
Its bet is that the richest memory an agent can have is the recording of what you did, and its risk is the same bet at maximum scale: it captures everything, all the time.
What it is #
A Rust engine with a Tauri desktop app and a CLI (npx screenpipe record) from Negentropy Labs, Inc. (founder Louis Beaumont).
Capture is event-driven: a screenshot is paired with the OS accessibility tree when something changes, falling back to OCR, while system and microphone audio are transcribed locally with Whisper.
Everything lands in a local SQLite store with FTS5 search, a REST API on localhost:3030, and an MCP server (npx screenpipe-mcp) that Claude Desktop, Cursor, and coding agents query.
Pipes are scheduled AI agents defined as markdown files, with YAML frontmatter permissions (allowed apps, windows, content types, time ranges) enforced in three deterministic layers rather than by prompting.
npx screenpipe setup wires skills and MCP into Claude Code, Codex, Gemini CLI, Cursor, and OpenCode.
Status #
Active and large, on a YC-backed company. 21,954 stars with the repository pushed 2026-11-17, created 2024-06-19, as of 2026-11-17 (GitHub API).
The Hacker News record is the strongest in this category: a 218-point Show HN in September 2024 (125 comments), an 88-point Launch HN with 67 comments as a YC S26 company in July 2026, and a 2024 Ask HN thread questioning the business model. The README positions it as the source-available alternative to Limitless, Microsoft Recall, Granola, and Otter. The app train keeps moving: app-v2.7.125 (2026-10-26) gates covered local reads and enterprise uploads on current PII-processing receipts when required PII processing is enabled, so capture stays durable while labels are processed but sensitive reads wait for the privacy pass to complete, and app-v2.7.129 through v2.7.132 (2026-10-28 to 30) repaired upload starvation during retention, recovered interrupted Windows migration retries, and stopped excluded windows from appearing in screenshots, while the AI gateway added no-training Experiential routing that carries the user’s current training-consent choice. The train continued to app-v2.7.138 (2026-11-03) with the CLI shipping v0.4.56 and v0.4.57 the same day (fresh hybrid-history search and replacement of saved opposite-window filters), enterprise deployments restarting stuck apps through independent support polling, filtered screenshots recovering after missing native completions, oversized Experiential compaction requests recovering instead of failing, and Windows keeping the user’s own Enterprise installation authoritative. App-v2.7.143 and v2.7.145 (2026-11-06) keep the same maintenance character: enterprise uploads fail closed when no upload destination pin is deployed, enterprise heartbeats report the device timezone, fresh histories sync from persistent database sources, silent microphone streams stay running, macOS browser click URLs populate across URL schemes, and Windows uninstall now deletes recording data when selected. The default-branch train continued (November 12-13, pending release): the app bumped to v2.7.147 with display text restricted to the focused window’s native ownership, an enterprise simplified UI mode, simplified overlay localization extraction, and enterprise honoring local upload pins while keeping recovery available. The pending train continued again (November 13-14, through app v2.7.150, unreleased): configurable screenshot sampling with a 2 fps default, enterprise device policies applied without a webview and required upload pins scoped to explicit MDM overrides, a macOS fix capturing the display before applying screenshot redaction, a Windows SQLite WAL-stall fix, and Claude login isolated from external ACP sessions; app v2.7.151 (November 15) keeps macOS whole-display recording independent of scene changes so a scene transition no longer interrupts it. The pending train extended again (November 16 to 17): Windows audio stops re-probing endpoint formats every 30 seconds, pipes stop disabled jobs and hide Windows children, onboarding no longer shows a false stuck state during database initialization, storage preserves startup disk-reserve failures with recovery evidence, and a Workflows sidebar card points self-serve users to the enterprise product.
Strengths #
- The timeline answers the questions session logs cannot: what did I see, say, and do outside the terminal.
- Event-driven accessibility-tree capture keeps CPU and storage bounded compared with naive frame recording, per its own resource estimates.
- Per-pipe YAML data permissions enforced at the server and agent layers are the most concrete agent-data-governance story in this category.
- Distribution is already in place: MCP, REST, SDK, skills, and desktop installers across macOS, Windows, and Linux.
Cautions #
- Source-available is not open source: the Screenpipe Commercial License permits personal and non-commercial use free, gives organizations a 7-day evaluation, and requires a paid license for commercial use, so 21.9k stars do not mean free to deploy.
- Telemetry is on by default (PostHog analytics and Sentry), and local capture does not mean zero network traffic; the README’s own FAQ itemizes what leaves the device.
- Capture-everything is this category’s known failure mode at its maximum: screen, audio, and keystrokes concentrate more sensitive data than any other member here.
- The production app’s full history and sync sit behind paid tiers, and the 2024 business-model skepticism in its own community was never fully answered.
Pricing #
Free: $0, one device, capture, search, and sharing context with AI tools. Basic: $21/month ($250/year billed annually), full searchable history and unlimited scheduled workflows. Business: $42/seat/month ($500/seat/year), personal device sync and managed seats, shared context across teammates not included. Enterprise: custom, scoped per deployment. A 50 percent first-month student discount applies, as of 2026-10-11. Source builds follow the commercial license: personal, non-profit, and educational use free, organizational evaluation 7 days.
Price history #
| Date | Plan | Change | Source |
|---|---|---|---|
| 2026-10-08 | Free, Basic, Business, Enterprise | Baseline: Free $0 (one device), Basic $21/mo ($250/yr), Business $42/seat/mo ($500/seat/yr), Enterprise custom; source builds free for non-commercial use. |
Compared to #
- claude-mem: session-compression memory scoped to what your coding agent did; Screenpipe records everything on the screen, including everything outside the agent.
- Engrim: curated episodic records versus raw capture; Engrim stores what the agent decided to log, Screenpipe keeps what happened.
- File-based agent memory: the zero-capture convention; the two agree only that the store should live on your disk.
Bottom line #
Recommended for developers who want coding agents grounded in their actual machine activity and will pay for it and govern it. Not for privacy-strict environments, commercial use without a license budget, or anyone who expected an OSI open-source license at this star count.
Changes #
- 2026-10-08 - Created from the 2026-10-08 awesome-list entrant scan, with eight fetched sources and the source-available-license-plus-capture-everything combination recorded as the critical angle.
- 2026-10-09 - Corrected the Status rank claim: at 21,884 stars Screenpipe ranks eleventh of the twenty-two members, not the category’s second-largest repository; refreshed stars and the as-of dates.
- 2026-10-11 - Recomputed the rank on the each-member’s-own-repository basis (Zep’s Graphiti engine no longer carries Zep’s rank): tenth of the twenty-three members at 21,903 stars; pricing page re-fetched with all four plan prices unchanged.
- 2026-10-11 - Membership grew to twenty-six (TencentDB Agent Memory, Memanto, and MemMachine joined), moving the rank from tenth of the twenty-three to eleventh of the twenty-six members at 21,903 stars, on the each-member’s-own-repository basis (Zep’s Graphiti engine and the file convention carry no rank).
- 2026-10-11 - Emulo joined as the twenty-seventh member, moving the rank denominator only (Emulo’s 294 stars rank below screenpipe).
- 2026-10-16 - Corrected the rank denominator to twenty-eight after PLUR’s 2026-10-14 admission (PLUR’s 304 stars rank below screenpipe, so the position holds), and refreshed stars to 21,954.
- 2026-10-18 - Dropped the positional star rank entirely: the denominator went stale again at the category’s thirty-two members (OKF Agent Memory and Operator Memory joined 2026-10-17), and the repeated rank-correction rounds showed the claim costs more maintenance than its one line of value; the substance (21.9k stars, YC-backed, strong HN record) is unchanged.
- 2026-10-25 - Corrected the star-history chart fallback markup (the img fallback carried the dark-theme parameter; it now points at the default chart).
- 2026-10-26 - Corrected the star-history chart markup: the light-theme source carried the dark-theme parameter; it now points at the default chart.
- 2026-10-27 - Recorded app-v2.7.125 (October 26): covered local reads and enterprise uploads now wait for current PII-processing receipts when required PII processing is enabled, keeping capture durable while label processing catches up.
- 2026-10-30 - Recorded the app train to v2.7.132 (October 28-30): upload-starvation repair during retention, interrupted Windows migration retries recovered, excluded windows kept out of screenshots, and no-training Experiential routing that carries the user’s training-consent choice.
- 2026-11-04 - Recorded the train to app-v2.7.138 with CLI v0.4.56 and v0.4.57 (November 3-4): fresh hybrid-history search and saved opposite-window filter replacement, stuck-app restarts through independent support polling in enterprise, filtered-screenshot recovery, oversized Experiential compaction recovery, and Windows keeping the user’s Enterprise installation authoritative; the push date refreshed to 2026-11-04.
- 2026-11-07 - Recorded app-v2.7.143 and v2.7.145 (November 6): enterprise uploads fail closed when no upload destination pin is deployed, heartbeats report device timezone, fresh histories sync from persistent database sources, silent microphone streams stay running, and Windows uninstall deletes recording data when selected; the push date refreshed to 2026-11-07.
- 2026-11-11 - Recorded the November 7-to-11 default-branch work: the app bumped to v2.7.146 with the consumer export admission fix, Workflows respecting the app theme with semantic icons and more filter groups, periodic feature-flag refreshes reduced, verified enterprise accounting tags carried through inference, and the Wayland portal work (approval persistence, KDE streams mapped by connector, opaque portal tokens accepted without an invented length cap); the push date refreshed to 2026-11-11.
- 2026-11-13 - Volatile re-verification: the pending November 12-13 batch recorded (app v2.7.147: focused-window display text ownership, the enterprise simplified UI mode, and upload pins honored with recovery kept); no new release tag, and the push date refreshed to 2026-11-13.
- 2026-11-14 - Recorded the pending November 13-14 batch (app v2.7.148 through v2.7.150, unreleased: configurable screenshot sampling with a 2 fps default, enterprise device policies without a webview with upload pins scoped to explicit MDM overrides, the macOS redaction capture-order fix, a Windows SQLite WAL-stall fix, and Claude login isolation); the push date refreshed to 2026-11-14.
- 2026-11-16 - Recorded app v2.7.151 (November 15, unreleased): macOS whole-display recording kept independent of scene changes; the push date refreshed to 2026-11-15.
- 2026-11-17 - Volatile re-verification: the pending train extended again (November 16 to 17): Windows audio endpoint re-probing stopped, pipes stop disabled jobs and hide Windows children, the onboarding false stuck state during database initialization removed, storage preserving startup disk-reserve failures with recovery evidence, and the Workflows sidebar card pointing self-serve users to the enterprise product; no release after app-v2.7.145, and the push date refreshed to 2026-11-17.
See also #
- Memory Feature Matrix - the category comparison this note joins
- claude-mem - the session-scoped capture layer, the nearest architectural cousin
- Engrim - the curated-records counterpoint to raw capture
- File-based agent memory - the convention at the zero-capture end of the spectrum
References #
https://github.com/screenpipe/screenpipe - repository, 21,903 stars, activity, as of 2026-10-11
https://raw.githubusercontent.com/screenpipe/screenpipe/main/README.md - capture model, pipes and their permission model, MCP and API surfaces, positioning
https://raw.githubusercontent.com/screenpipe/screenpipe/main/LICENSE.md - the Screenpipe Commercial License terms (personal free, 7-day organizational evaluation, commercial paid)
https://screenpipe.com/pricing - Free, Basic, Business, and Enterprise plans as of 2026-10-11, re-verified unchanged
https://news.ycombinator.com/item?id=49024620 - the YC S26 Launch HN thread (88 points, 67 comments, 2026-07-23)
https://news.ycombinator.com/item?id=41695840 - the 2024 Show HN (218 points, 125 comments), the category’s largest community record
https://hn.algolia.com/api/v1/search?query=screenpipe&tags=story - the footprint scan, including the 2024 business-model skeptic thread