Screenpipe launches: record your screen, pipe it to your agents
It turns recorded work into MCP agent context and scheduled Pipes, local-first. Plus OneCLI keeps raw API keys out of your agents entirely.
Copy markdown
Your agent finally sees what you actually did
Screenpipe records screen and mic 24/7 on-device, snapping a screenshot plus the accessibility tree on every app-switch, click or typing pause (OCR as fallback) and indexing it in local SQLite. That's ~300MB per 8 hours versus ~2GB for continuous video, at 5-10% CPU. The YC S26 launch lands with ~20k GitHub stars.
Pipes turn that timeline into scheduled agents
The real unlock is Pipes: markdown + YAML files that run an agent against your activity timeline on a schedule - meeting summaries, standups, day recaps grounded in what you did, not what you typed into a prompt. Record yourself doing a task once and a Pipe can write the SOP your agent replays.
One line to wire it into Claude, Cursor or Codex
It runs as an MCP server, so `claude mcp add screenpipe -- npx -y screenpipe-mcp@latest` gives Claude Desktop, Cursor, Cline or Continue direct query access to your screen history. There's a REST API on :3030 and a JS SDK for custom automation, with local Ollama or any OpenAI-compatible endpoint and Whisper handling transcription on-device.
The catch: source-available, not open source
Despite the 'open' framing, screenpipe ships under a commercial license - free for personal use, but commercial or SaaS use needs a paid tier ($25/$50/$150 per seat). Analytics (PostHog, Sentry) are on by default, and the maintainers warn main breaks often, so pin a release for anything real.
Meanwhile OneCLI keeps your API keys out of the agent
Timely after this week's token thefts: OneCLI (Apache-2.0) is a Rust proxy on :10255 that swaps placeholder creds like `FAKE_KEY` for the real ones at request time, with AES-256-GCM at rest. Point your agent's HTTP through localhost and it never handles an actual key. 2.4k stars, now at v1.36.0.