Recipes
The agent you already run can be the one that fills Koma and works from it. These are the standing orders to paste, one per runner. Each does two jobs: sweep your sources into Koma, and work whatever is waiting on you there. Your agent keeps using its own connections to mail, calendar and meeting notes; Koma only receives what it sends.
Install Koma in the agent first (Connect a client). The two tools the work half uses, read_chat_messages and post_chat_message, are on every account in the preview.
Claude Cowork, scheduled
Type /schedule in any Cowork task, pick hourly, make sure the Koma connector is on, and paste:
Every run: look through my meeting notes, my calendar for today and tomorrow, and unread mail from real people since the last run. Find commitments I made, requests waiting on me, and dates. Search Koma first so nothing lands twice. Add what is new to my Tasks list with create_items, one line of context and the link each. Skip bulk mail, invitations already accepted, and anything I was only copied on.
Then work the Koma tasks I hand you: the ones I name, or everything open on the list "For you". For each: get_item for the task (its body with a handle per block, and its guidance), read_chat_messages for what was said (pass since, the createdAt of the last message you saw, so only what is new comes back), set it to inProgress when you start, and do the work in its body with update_items edits, continuing what is already there. Then post once with post_chat_message on the same item id, with your name as agentName, saying what changed in the body, only if something changed, plus any action you took outside Koma with its link or the one question you need answered. Never set a task to completed; that is mine. Do not narrate steps.
Runs in the cloud, no machine on. Replies you write in a task are picked up on the next run.
Claude Code Routine, scheduled
At claude.ai/code/routines, new routine, any repository (an empty private one is fine, the session only needs a folder), connector Koma, schedule hourly, and the same prompt:
Every run: look through my meeting notes, my calendar for today and tomorrow, and unread mail from real people since the last run. Find commitments I made, requests waiting on me, and dates. Search Koma first so nothing lands twice. Add what is new to my Tasks list with create_items, one line of context and the link each. Skip bulk mail, invitations already accepted, and anything I was only copied on.
Then work the Koma tasks I hand you: the ones I name, or everything open on the list "For you". For each: get_item for the task (its body with a handle per block, and its guidance), read_chat_messages for what was said (pass since, the createdAt of the last message you saw, so only what is new comes back), set it to inProgress when you start, and do the work in its body with update_items edits, continuing what is already there. Then post once with post_chat_message on the same item id, with your name as agentName, saying what changed in the body, only if something changed, plus any action you took outside Koma with its link or the one question you need answered. Never set a task to completed; that is mine. Do not narrate steps.
A routine can also fire on an API call. That is the events guide.
ChatGPT, scheduled task
In ChatGPT, create a scheduled task, hourly, with the Koma plugin connected:
Every run: look through my meeting notes, my calendar for today and tomorrow, and unread mail from real people since the last run. Find commitments I made, requests waiting on me, and dates. Search Koma first so nothing lands twice. Add what is new to my Tasks list with create_items, one line of context and the link each. Skip bulk mail, invitations already accepted, and anything I was only copied on.
Then work the Koma tasks I hand you: the ones I name, or everything open on the list "For you". For each: get_item for the task (its body with a handle per block, and its guidance), read_chat_messages for what was said (pass since, the createdAt of the last message you saw, so only what is new comes back), set it to inProgress when you start, and do the work in its body with update_items edits, continuing what is already there. Then post once with post_chat_message on the same item id, with your name as agentName, saying what changed in the body, only if something changed, plus any action you took outside Koma with its link or the one question you need answered. Never set a task to completed; that is mine. Do not narrate steps.
On Plus and up, a Work chat on chatgpt.com (or in the desktop app with Cloud selected) can instead subscribe to Koma's "task waiting" event and act within seconds, with Koma added on the web Plugins page. See the events guide.
Codex, automation
In the Codex desktop app, new automation, hourly, with the Koma server in Codex's MCP config:
Then work the Koma tasks I hand you: the ones I name, or everything open on the list "For you". For each: get_item for the task (its body with a handle per block, and its guidance), read_chat_messages for what was said (pass since, the createdAt of the last message you saw, so only what is new comes back), set it to inProgress when you start, and do the work in its body with update_items edits, continuing what is already there. Then post once with post_chat_message on the same item id, with your name as agentName, saying what changed in the body, only if something changed, plus any action you took outside Koma with its link or the one question you need answered. Never set a task to completed; that is mine. Do not narrate steps.
Codex is the right runner for tasks that need a repository. Give the automation the repository the tasks are about.
What to expect in Koma
A post from your agent shows in the task's chat as a note under the agent's name, with the dot on the task and a line in your main conversation. Write back in the same chat; the agent reads the chat on its next run. Koma's own assistant keeps answering in that chat too, and it reads the agent's notes as someone else's words.