screenpipe (YC S26): Local Work Memory, SOPs, and AI Agents via MCP

On July 14, 2026, screenpipe (@screenpipe) posted a launch thread that reframed the product in one line: record how you work → searchable memory, SOPs, and AI agents. The pitch — open source, local-first, npx screenpipe record, ~20K GitHub stars — cleared 260K+ views in hours.

The idea is not a new chat window. It is a continuous work ledger on your machine: screen + audio → local SQLite → MCP/API → agents. Developers already report piping that ledger into Codex and Claude for continual memory; skeptics ask whether always-on capture is compatible with real privacy.

explainx.ai verifies GitHub stats, walks install + MCP wiring, and maps exclude-app controls, Pipes, and how this differs from Limitless/Rewind — without treating a marketing thread as a security audit.

TL;DR

What screenpipe actually captures

The README architecture is straightforward:

screen + audio → local storage → AI

Event-driven capture is the technical bet. Instead of saving every second of video, screenpipe listens for app switches, clicks, typing pauses, scrolls, and grabs a screenshot + accessibility tree when something changes. Text comes from the OS accessibility layer first; OCR (Apple Vision, Windows native, Tesseract on Linux) fills gaps for games, RDP, or broken a11y.

Audio uses local Whisper Large-V3-Turbo by default, with optional Deepgram cloud transcription. System and mic audio can be transcribed with speaker diarization.

Storage lands in local SQLite with FTS5 plus JPEG frames on disk. The project claims ~300 MB per 8 hours vs ~2 GB for naive continuous recording — a meaningful difference if you leave capture on for weeks.

explainx.ai read: This is closer to structured activity logging than DVR screen recording. That design choice is why MCP integration matters — agents query events and text, not raw multi-hour MP4s.

Install: CLI vs desktop app

Two entry points:

CLI (fastest trial):

npx screenpipe record

Desktop app: screenpi.pe/onboarding — signed builds, auto-updates, subscription features.

MCP hook for Claude-family hosts:

claude mcp add screenpipe -- npx -y screenpipe-mcp@latest

Then ask in-session: "what did I see in the last 5 mins?", "summarize today conversations", or "create a pipe that updates Linear every time I work on task X".

For protocol background, see explainx.ai's MCP guide. For permission boundaries when agents read personal context, pair with context-mode MCP sandboxing thinking — screenpipe is powerful precisely because it is broad.

Pipes: SOPs and scheduled agents

Pipes are markdown files (pipe.md) with prompts and schedules. screenpipe runs an agent (pi, Claude Code, etc.) against your local timeline and can call APIs, write files, or patch records.

Built-ins listed in the README include:

Custom pipes live in ~/.screenpipe/pipes/. That is the SOP automation layer the Jul 14 tweet highlights — not just search, but recurring agent outputs grounded in what you actually did.

YAML frontmatter supports deterministic data permissions: allow-apps, deny-apps, deny-windows, content-type gates, time ranges, and endpoint blocks (allow-raw-sql: false). Enforcement is three-layer (skill gating, agent interception, server middleware with per-pipe tokens) — closer to policy-as-code than "please don't look at my bank tab."

This overlaps conceptually with agent skills (how agents behave) but adds OS-level capture as the data plane. Loop engineering is the natural orchestration layer for nightly standup pipes or verification gates on generated SOPs.

Privacy, exclude apps, and the Jul 14 X debate

The launch thread split reactions:

Filters documented in README: window, app, browser extension, password fields, optional Screenleak PII model (May 2026 alpha). Audio exclusions on macOS 14.4+:

{ "excluded_apps": [{ "bundle_id": "com.spotify.client", "name": "Spotify" }] }

File: ~/.screenpipe/audio-exclusions.json (hot-reloads; override path via SCREENPIPE_AUDIO_EXCLUSIONS_PATH).

Telemetry caveat: README FAQ states PostHog analytics and Sentry are on by default; disable under Settings → Privacy → Analytics. Local capture does not mean zero network — only that frames and transcripts stay on disk by default.

OpenClaw angle: README lists OpenClaw among integrations. explainx.ai covered OpenClaw harness risks separately — if you route screenpipe MCP + OpenClaw messaging, treat OAuth and third-party harness bans as live policy, not solved problems. See OpenClaw safety analysis.

screenpipe vs Limitless / Rewind / Recall

Use this table as a buyer/engineer checklist, not a scorecard — features shift quarterly.

Limitless question from X: "How's this different besides local storage?" — screenpipe's answer in docs is API + Pipes + permissions YAML + Linux + multi-monitor, not storage alone.

Claude Code /rewind note: Anthropic's /rewind restores session context after /clear — different problem from weeks of desktop memory. See Claude Code 2.1.191 changelog for harness-level rewind vs capture tools.

GitHub stats — verifying the launch claims

Repo URL: github.com/screenpipe/screenpipe (org migration from mediar-ai/screenpipe). Stars prove developer interest, not enterprise readiness — read LICENSE.md before shipping capture inside a commercial product.

Practical setup for builders

1. Start CLI — npx screenpipe record on a non-production machine; measure CPU/disk for your monitor layout.
2. Deny list first — banking, password managers, personal messaging before enabling MCP.
3. Add MCP — claude mcp add screenpipe -- npx -y screenpipe-mcp@latest; test narrow queries ("last 5 minutes in VS Code only").
4. One Pipe — deploy standup-update or meeting-summary before exotic automation.
5. Telemetry off — if policy requires, disable analytics in Settings before capture.
6. Team mode — if evaluating screenpi.pe/team, map admin filters to your MDM story separately from individual trials.

REST smoke test (default port):

GET http://localhost:3030/search?q=meeting+notes&content_type=all&limit=10

JS SDK entry point: @screenpipe/js with pipe.queryScreenpipe({ q, contentType, limit, startTime }).

Risks and limitations

• Capture breadth — even with excludes, misconfiguration leaks sensitive windows into FTS5 index.
• License — commercial redistribution or SaaS embedding requires a commercial license, not "it's on GitHub."
• Cloud toggles — Deepgram, cloud AI, encrypted sync send data off-device when enabled.
• Harness coupling — feeding everything to Codex/Claude increases context quality and exfiltration blast radius if MCP scopes are loose.
• Main branch velocity — README warns main breaks often; pin releases for production desktop rollouts.

Summary

screenpipe's July 14, 2026 thread repositioned a mature repo (~20K stars, ~1.9K forks) as local work memory for agents: capture screen+audio locally, search with FTS5, expose context through MCP, automate with Pipes. npx screenpipe record plus screenpipe-mcp is the fastest Claude/Cursor integration path. Exclude-app controls and YAML pipe permissions address part of the privacy backlash, but always-on capture still demands explicit deny lists and telemetry choices. vs Limitless/Rewind, screenpipe sells auditability, API surface, and scheduled agents — not just local disk.

Related on explainx.ai

• Claude Memory Heist — web_fetch PII exfiltration (Jul 2026)
• What is MCP? Model Context Protocol guide
• What are agent skills? Complete guide
• Loop engineering for coding agents
• Context Mode — MCP sandboxing for agents
• Claude Code 2.1.191 — /rewind vs MCP fixes
• Is OpenClaw safe? Harness and OAuth context
• What is MEMORY.md? Agent persistence patterns

Sources: screenpipe README · docs.screenpi.pe · GitHub API — screenpipe/screenpipe, Jul 15, 2026 · screenpipe X launch thread, Jul 14, 2026

Star counts, pricing, and telemetry defaults are accurate as of July 15, 2026. Verify license terms and capture settings before deploying on managed or regulated devices.