Your screen time, compiled into agent-ready workflows

What it does

activity-frames is a local screen-capture compiler that turns raw screen snapshots and input events into structured records of what you actually did. It bundles your day into bounded “activity frames” — which app, which site, for how long, what you typed — and surfaces them as context blocks or executable step sequences for LLM agents via MCP. Everything stays on your machine: capture, storage, and compilation never touch a network.

The interesting bit

The unusual bit is its two-tier honesty contract. Tier one is strictly measured — session durations, page entities, click counts — derived deterministically from capture data with no LLM in the compile path. Tier two quarantines any inferred intent in a namespaced, confidence-tagged annex, so facts and guesses never mix. It is episodic memory with an audit trail and a bouncer.

Key highlights

• Runs fully offline; capture, SQLite storage, and compilation are local-only
• Recurring tasks compile into deterministic replay scripts that agents can execute at zero model calls on the happy path
• Ships as an MCP server with six tools for context retrieval, pattern detection, and step extraction
• No LLM in the compile path: frames are produced by deterministic code, not interpretation
• Typed text content is opt-in (--include-text); audio recording is off by default

Caveats

• macOS-first: the bundled capture engine is prebuilt for Apple Silicon, and Linux users must supply their own recorder
• Entity parsing covers 25+ sites, with generic fallbacks for anything unrecognized
• Still v0.2, and plugging in a third-party recorder requires matching database schemas

Verdict

Worth exploring if you want local agents to stop re-deriving your daily routines from scratch. Skip it if you are not comfortable with local screen recording, or if you need a mature Linux capture engine out of the box.