~*~ open source · realtime · early beta · Apache-2.0 ~*~
Stop scrolling, let your agent do it for you
Clockwork sorts your incoming messages in realtime on your own machine, so your AI agent can watch them for you.
Read-only extractors collect your WhatsApp, Gmail, Slack, X and a dozen other sources. A small local model files each item as urgent, later or ignore. Your agent teaches it over MCP, and tells you about what matters.
- It runs on your machine. No account, no telemetry, no Clockwork server.
- It only reads. Nothing can send, reply, react, post or mark as read.
- Your agent stays in charge. The local model gets perfected by your agent over time.
Why
Most of what is on your screens is not worth your time. Some of it is: the message from the school, the invoice, the friend who needs an answer today. That is why you keep checking. Until now, the only way to catch the few things that matter was to look at everything.
An agent can do the looking. For the first time there is a technology that can read all of it for you and pass on only what you needed. You can spend far less time on screens and still get what they were for.
~*~
Install
Give your agent the repository and say:
Install Clockwork on this machine by following agent/INSTALL.md.
Or by hand:
git clone https://github.com/clockworkmcp/clockwork.git cd clockwork ./install.sh clockwork doctor # every line should say ok clockwork start clockwork connect whatsapp # one account at a time clockwork agent # MCP address, token, skill
+---------------+-----------------------------+ | Machine | a Mac that stays on | | Tested on | macOS, Apple Silicon | | Node | 24+ | | Python | 3.10 - 3.13 | | Disk | about 2 GB | | Memory | 2 GB free | | Agent | anything that speaks MCP | | Account | none | | Price | free, Apache-2.0 | +---------------+-----------------------------+
How it works
your accounts
| read-only extractors
v
clockwork, on your machine
|
v
small local model <---------+
| everything it is | labels
| not sure about | teach it
v |
your agent -----------------+
| what matters
v
you
Every source starts in shadow mode: the model predicts and decides nothing while your agent labels the backlog. After 200 labels it trains every night. Once its confident calls match your agent's 95% of the time, it starts filing what it is sure is not urgent.
Everything else goes to your agent. Anything nobody has judged is held as later, never thrown away.
Sources
Sixteen built in, all off until you connect them.
+---------------+-----------------------------+ | WhatsApp | your chats | | Gmail | your inbox | | Slack | conversations you are in | | X | your home timeline | | Outlook | your inbox | | iMessage, SMS | texts on this Mac | | Calendar | new and changed invitations | | Discord | channels your bot is in | | GitHub | your notifications | | GitLab | your to-do list | | Linear | your inbox | | Jira | issues assigned or watched | | Bluesky | replies, mentions, quotes | | Mastodon | mentions, notifications | | Hacker News | replies, your keywords | | RSS / Atom | any feed address | +---------------+-----------------------------+
Write your own from a 50-line template.
Your data
- Items, labels, the model and your account sessions live in one folder on your disk.
- To label an item, your agent reads it. If your agent runs on a cloud model, that content goes to its provider.
- WhatsApp and X are read in ways their terms do not allow. Read the account risks first.
Clockwork
In A Clockwork Orange a man is strapped to a chair with his eyes clamped open and made to watch a screen. Keeping up with every chat, inbox and timeline can feel like that.
Clockwork straps your agent into the chair instead. It watches every feed, all day, with its eyes held open. It does not get tired, anxious or hooked. You get up and leave the room, and it comes to find you when something matters.
–––