Skip to content
Nathan Zhou

Sisyphus

A Teams assistant that answers messages while I'm away.

Sisyphus watches Microsoft Teams chats on a dedicated Mac, drafts replies with an AI model, and can send them if allowed. It controls the Teams desktop app through macOS accessibility instead of a Microsoft API, so nothing has to be registered and no tenant admin has to approve it. By default it only writes drafts for review. Coding agents can also ask it to message someone, and nothing goes out until the ask is approved in its panel.

Status
Reading, drafting and sending all work. Sending is opt-in: chat by chat for replies, and ask by ask for agents. Drafts stay the default.
One message answered, others left alone, and a brief ready in the morning.

Overview§

Teams messages keep coming whether or not I'm at my desk. In a few busy group chats most of them need no answer, and the few that do are easy to miss. Sisyphus runs on a dedicated Mac. It reads the chats, works out which messages need a reply, and writes one.

Most of the time you use it through the control panel. The top line shows how many conversations it is watching, and how many messages it answered and left alone today. Its replies in each thread are signed, and each message it chose not to answer is marked with the reason. You choose which conversations it may answer and give each one its own standing facts. Each morning a brief lists the counts, a short summary, what matters, a table of things people asked you to do, and a line on each chat.

It reads the Teams app's interface, not an API, so a Teams update can change what it sees. When that happens it fails loudly instead of reporting no new messages, and the panel stays up to say what is wrong. It also stays out of any chat you are writing in yourself for 30 minutes. Replying on its own is switched on chat by chat, and drafts stay the default.

It now does more than answer. A coding agent can ask it to send one message, hold a short exchange or carry a longer conversation, and the ask waits in the panel with Approve and Reject. Replies land in a mailbox the agent reads. It can also post to WeChat groups I have allowed, send only. Every ten minutes it writes a digest of what came in for my other apps, and Odysseus, my Windows island, shows it, with urgent messages popping up at once.

At a glance§

Role
Built alone, with AI coding assistants
When
18 Sept 2026 to 7 Nov 2026 · 7 weeks
work on 27 different days
Commits
212
saved steps of work
Automated tests
277
checks the code runs on itself
Code
36,722 lines of TypeScript
38,896 in all

Counted from the repository on 12 November 2026: commits on every branch, without merges; tests as written, so one that runs several cases counts once; lines that are not blank, without other people's code, generated files or lockfiles.

Key facts§

How it works§

  1. It types into Teams the way accessibility tools do. In testing this went through the app's full input handling and worked while Teams was in the background. Setting the text field directly changed the screen, but the app didn't notice.

  2. Every Teams message id ends in a millisecond timestamp. Finding new messages is a number comparison, not a diff.

  3. Replies go out under the user's own account, so its own messages come back looking like a person's. Three independent guards, with their own tests, stop it answering itself.

  4. A small local model first decides whether a message needs a reply at all. Overnight, sleep mode answers only when it is certain and otherwise sends a fixed standby message.

    A group chat in the Sent tab: small talk marked “Left alone — judged not worth answering”, two questions to Nathan answered and signed
    Small talk gets no reply. Questions put to Nathan get one, signed as the bot.
  5. Sending has three modes: draft, allowlist and auto. Each conversation has an hourly cap, and every reply is signed as coming from the assistant. An agent's ask is capped too, at one message, 12 over two hours, or 100 over a day, and the agent's token can ask but never approve.

    The conversations list: five it answers, four muted, two showing their own instructions
    The conversations it answers. Each can have its own instructions.
  6. WeChat shows nothing to accessibility tools, so for WeChat it reads the window with on-device text recognition and clicks. It checks the group's name on screen before pasting and sends nothing if any check fails. A rehearsal mode runs every step except Send.

Screens§

  • The control panel's Sent tab with replies signed by the bot

    Replies it sent, each signed as the bot. One turns down a request to move money.

  • Per-conversation instructions and a drafted relay message ready to send

    Standing facts for each conversation, and a relay: you say what to get across, it writes the message, you send it.

  • The brief since yesterday: counts, a one-line summary, what matters, a to-do table with who asked and when, and a row per conversation

    The morning brief shows what came in overnight, what people are waiting on you for, and where each chat stands.

See it work§

Sisyphus runs on my own Teams account, so this is a recording.

Sisyphus

Sleep mode off

Watching 3 of 12 conversations. Today it answered 0 and left 0 alone.

    A replay with made-up messages and names, not real conversations.

    Built with§