Skip to main content
Glama
abhitsian

myday-mcp

by abhitsian

My Day

A private record of what you worked on, kept on your Mac.

It writes a short note every ten minutes. Weeks later you can ask it what you were doing — and so can your AI assistant.


The problem

You finish a week and cannot account for it. Your manager asks what you shipped. You open a half-finished document and spend ten minutes reconstructing why you opened it. You read something useful in March and cannot find it in August.

Meanwhile your AI assistant starts every conversation knowing nothing. You re-explain the project, the file, the thing you already tried, every single time.

Your computer knew all of it and wrote none of it down.

Related MCP server: BuildAutomata Memory MCP Server

What My Day does

Every ten minutes it writes one small Markdown file describing what you were doing:

---
start: 09:10
end: 09:20
title: Tracing the webhook retry loop
summary: You read Stripe idempotency docs, then edited retry.ts.
apps: Visual Studio Code, Google Chrome
sites: stripe.com, github.com
project: payments-api
---
- Reading Stripe's idempotency-key docs on retried deliveries
- Editing retry.ts, focused on the key-reuse path

That is the whole storage format. No database, no account, no upload. You can grep it, edit it, or delete a file you would rather not keep.

Install

npm install -g @abhitsian/myday
myday init      # explains exactly what gets read, then asks
myday start     # begins recording

macOS, Node 18 or later. Prefer an app? Download My Day.app, drag it to Applications, and it walks you through the same six screens with a menu bar icon.

Removing it is one command and takes everything with it:

myday uninstall

How it works

flowchart LR
    A["Every 15s<br/>which app is in front"] --> D
    B["Your browser's own<br/>history database"] --> D
    C["Claude Code<br/>session transcripts"] --> D
    D["Every 10 minutes<br/>one Markdown note"] --> E["~/.myday/memories/"]
    E --> F["The app<br/>timeline, threads, friction"]
    E --> G["MCP server<br/>your AI assistant reads it"]

Three sources feed it, and you switch each one on or off separately. Reading your browsing is a different decision from reading your terminal work, so they are different switches.

It works before you grant it anything

Tier

Needs

You get

1

nothing at all

Which app was in front, and for how long

2

nothing at all

Page titles and addresses, from your browser's own history

3

Accessibility, one small binary

Window titles — the document, the chat, the folder

Most tools in this category want an invasive permission before they show you anything, so you have to trust them on a promise. Run tiers 1 and 2 for a week and read the output first.

For tier 3, myday build-helper compiles a 60-line Swift program that reads two attributes and prints them. You grant Accessibility to that one binary rather than to osascript, which would hand the same power to every script on your machine.


The timeline

Your day in order, with real application icons and clickable page links. Claude Code sessions sit on the same rail, each with a resume button that copies the command to reopen it.

Threads

A tracker tells you that you spent 1h55m in Slack, which is a fact about software. Threads tell you the checkout bug ran across four days and has been quiet since Tuesday, which is a fact about your work.

Nothing is declared. There is no project field to fill in — threads are derived from the files you touched and the pages you opened. Each carries a state (today, yesterday, this week, quiet), and the ones you picked up and put down surface first.

Friction

The costs your day hides from you, each with the evidence attached:

  • Signing in to the same host 29 times a week, which is a session-timeout setting rather than a habit

  • The same search typed on 13 separate days, because the answer never got saved anywhere

  • A page you navigate to most days and never bookmarked


Your AI assistant can read it

This is the part that changes how the tool feels. My Day ships an MCP server, so any assistant that speaks MCP can look up what you did without you re-explaining it.

Claude Code — add to ~/.claude.json:

{ "mcpServers": { "myday": { "command": "myday-mcp" } } }

Cursor — add to .cursor/mcp.json. Windsurf, Zed, Continue and anything else with MCP support take the same two lines.

Then, mid-task:

You: pick up where I left off

Claude: You were tracing a double-fire in the webhook retry loop — you'd read Stripe's idempotency docs and edited retry.ts, and retry.test.ts was still failing on the duplicate-delivery case (2026-08-14 15:30).

Four tools are exposed: history_search, history_window, history_resume, and history_time_by.

Notes are built from web page titles and window titles, which are text other people control. A page titled "ignore previous instructions" would otherwise reach an assistant that can run commands. Every response comes back inside a fenced envelope with the brackets escaped, so the payload cannot break out of it, and the rule is restated on every single response.


Where this came from

OpenAI shipped Computer History in the ChatGPT Mac app in August 2026: record interaction events rather than screenshots, roll them up every ten minutes into Markdown memory files, read those back when the user asks about past work. Windows Recall and Rewind had taken the screenshot route before it and spent their public lives defending that choice.

The text-only decision is the good one, and My Day keeps it. Where the two part ways:

ChatGPT Computer History

My Day

Before you grant anything

nothing

app names, then full browsing detail

Page detail

scraped from the window

read from the browser's own history

Works with

ChatGPT

any MCP client — Claude Code, Cursor, Zed

Storage

local Markdown

local Markdown, and you pick the folder

Recurring work

derived threads with a state per thread

Recurring costs

friction, with the evidence attached

Your own history

starts empty

reconstructed from months of browsing on day one

That last row matters more than it sounds. Threads need two or three weeks before they mean anything, so a fresh install has nothing to say at exactly the moment you decide whether to keep it. My Day reads the browsing history already on your disk and writes notes for the past sixty days, so the first screen you see is populated.

Privacy

Rules apply before anything is written, so an app you exclude never reaches disk. There is no later filtering step to get wrong.

myday permissions apps +Signal          # never record Signal
myday permissions sites include         # switch to an allow-list
myday clear hour                        # delete the last hour, notes and events both

Password managers are excluded out of the box. Private browsing is never included, because your browser does not record it. Summaries are written locally with no model by default, and enabling one logs every send to a file you can read.

Clearing removes the notes and the raw events behind them. Deleting a note alone would leave the events on disk and the next rollup would write it straight back.

Commands

myday init | start | stop | status | uninstall
myday show | timeline | apps | browse | sessions
myday threads | friction | search <q> | ask "<question>"
myday sources | permissions | clear | backfill
myday view                       # the app, in a browser

What it does not do yet

  • Window titles need a permission, so without tier 3 a note names the app and not the document.

  • Threads bind 41% of consecutive notes. Continuous work still fragments more than it should.

  • Nothing pushes. Every view waits for you to open it.

  • macOS only. The capture layer is Cocoa and LaunchServices.

MIT licensed. Built by Abhishek Sivaraman.

A
license - permissive license
-
quality - not tested
B
maintenance

Maintenance

Maintainers
Response time
Release cycle
Releases (12mo)
Commit activity

Resources

Unclaimed servers have limited discoverability.

Looking for Admin?

If you are the server author, to access and configure the admin panel.

Related MCP Servers

View all related MCP servers

Related MCP Connectors

  • Long-term memory for AI assistants. Hybrid retrieval, query expansion, auto-topics.

  • Persistent memory and knowledge graph for AI assistants — keyword + vector + graph search.

  • Persistent memory for AI agents. Search, store, and recall across sessions.

View all MCP Connectors

Latest Blog Posts

MCP directory API

We provide all the information about MCP servers via our MCP API.

curl -X GET 'https://glama.ai/api/mcp/v1/servers/abhitsian/myday'

If you have feedback or need assistance with the MCP directory API, please join our Discord server