activity-frames MCP
Records your screen locally and compiles the workday into structured activity frames that an agent reads through MCP.
- 530 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- Continuous screen capture covers passwords, messages and third-party data, and the local database stores typed text.
- The agent receives a summary of user actions, and window text can carry injected instructions.
Install
- Works in
Manual install
pip install activity-framesThe first README command. Then `aframes record` starts capture and MCP is added with `claude mcp add activity-frames -- aframes mcp`.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Continuous screen capture covers passwords, messages and third-party data, and the local database stores typed text.
- The agent receives a summary of user actions, and window text can carry injected instructions.
This is third-party code. Review the repository files before installing.
What it does
activity-frames compresses screen recordings into short documents: which app and site were open, for how many minutes, which pages and how many keystrokes. Computation runs as plain code without a language model, while capture and storage stay on your machine. The `aframes mcp` MCP server gives the agent six tools: `get_context`, `get_activity`, `get_steps`, `get_day_summary` and others. Recurring tasks become a step-by-step script that the agent executes instead of deriving again.
Who it is for. Users of computer-use agents who want memory of what they did during the day and repeatable workflows.
Good fit when
- An agent starts a conversation blind and needs the context of the last two hours.
- You ask which article you had open around nine in the evening.
- The same browser routine is done a hundred times and the agent needs a ready script.
- A daily summary of apps and time is wanted without manual tracking.
- A researcher measures how many tokens an agent spends re-deriving a routine.
Not a fit when
- Screen recording is not allowed on a work computer under company policy.
- You need Linux with a ready recorder, and no prebuilt capture engine exists there.
- The screen shows other people's personal data that must not land in a local database.
Example request
Show what I did in the last two hours and find the workflow for the query "message john doe".MCP
- Transport
- stdio
- Authentication
- not required
Limitations
The version 0.2 project was tested on Apple Silicon macOS, and the Intel Mac build is less exercised. There is no prebuilt capture engine for Linux: bring your own recorder and set `$AFRAMES_DB`. The capture database stores what is visible on screen, including typed text. Text reaches compiled documents only with `--include-text`. Audio is off by default. The last commit dates to August 2026.
How to disable. Stop capture with `aframes record --stop`, remove the MCP entry and the `activity-frames` package. Delete the capture database separately.
FAQ
Do recordings leave the machine?
The README states that capture, storage and compilation run locally and nothing is uploaded.
Which MCP tools exist?
Six, including `get_context`, `get_activity`, `get_steps` and `get_day_summary`.
How do I get context for the last hours?
With `aframes context --hours 3` or the `get_context` MCP tool.
Is a language model needed for compilation?
No, compilation runs as plain code. The capture engine recognises on-screen text locally.
How do I include typed text?
With the `--include-text` flag. By default documents carry only keystroke counts.
Discussion
No questions or reviews yet.
Related
MCP serversLocal persistent memory for coding agents: hooks record every session, and Claude Code, Codex or Cursor recall past decisions through 54 MCP tools
An MCP server for long-term agent memory: compresses dialogues into atomic facts, resolves pronouns and anchors relative time to absolute dates
Local agent memory built on a corrections ledger: tell Claude Code once and a new session reads the rule before it starts work
An MCP server that keeps large tool output out of the agent's context window and stores session memory in SQLite
An MCP server and CLI coordinate specialized agents, retain working patterns and connect tasks with project hooks.
Local memory for agents: decisions from browser chats reach Claude Code, Cursor and Windsurf through MCP
README in short
The README describes local screen capture, compilation into activity frames without a language model, six MCP tools, a CLI and a privacy model.