Comma

Where do Claude Code routines post their reports?

Anthropic shipped Routines for Claude Code in April 2026: schedule a recurring task, and the agent runs it without you opening a session. A nightly eval sweep, a Monday-morning metrics digest, a weekly dependency audit — the scheduling problem is solved.

The destination problem isn't. A routine defines when the agent runs and what it does. It does not define where a report should land. Out of the box you get three options, and each one drops part of the loop.

The three defaults, and what each loses

Session logs. The work happened; the transcript proves it. But a transcript is not a document. Nobody on your team opens last Tuesday's session log to check whether the conversion funnel moved, and there is no way to comment on it.

A Slack notification. Fine as a doorbell — "the report is ready" — but wrong as the artifact. The HTML becomes a snippet or an attachment, feedback scatters across thread replies, and next week's run buries this week's. We've written up the failure modes in Where should my agent post its work? — Slack is option one of four, and it falls apart the same way for scheduled runs, just faster, because the runs keep coming.

A pull request. Exactly right when the routine's output is code. Wrong when it's a report: a rendered analytics digest reviewed as an HTML diff is unreadable, PR comments anchor to source lines rather than rendered content, and merging a report into main every night is version-control noise, not review.

What's missing from all three is the second half of the loop: the team reads the report as a document, leaves feedback on the content, and the agent reads that feedback before its next run.

The shape that closes the loop

  1. The routine runs on its schedule and produces HTML.
  2. The agent publishes it to Comma over MCP — create_report on the first run, update_report on every run after, so the share link your team bookmarked refreshes in place instead of multiplying.
  3. The team reviews it like a Google Doc: comments anchored to the exact paragraph or table cell, threads, resolution. Reviewers don't need a seat — commenting is free on every plan.
  4. The next run reads the feedback. The agent calls list_comments, sees each comment with the text it's anchored to ("this regression table is missing the p95 column"), addresses it, replies with reply_to_comment, and resolves the thread with set_comment_status.

Schedule → publish → review → revise, with the routine driving the schedule and Comma carrying everything after.

Wiring it up

One line installs the Comma plugin for Claude Code:

curl -fsSL https://commareports.com/install.sh | sh

The installer attaches Comma's hosted MCP server, registers the /comma-publish, /comma-search, /comma-routine, and /comma-link slash commands, and walks you through creating a comma_sk_… token. The token is scoped and revocable from Settings → API tokens; it gates the MCP server, the REST API, and anything the routine publishes — one revoke pulls the agent's access everywhere.

Then make the destination part of the routine's task. The instruction is ordinary prose in the routine definition:

Run the weekly retention analysis. Before finishing:
1. Call list_comments on report <id> and address any open threads.
2. Publish the updated HTML to the same report with update_report.
3. Reply to each comment you addressed and mark it resolved.

On the first run, swap step 2 for create_report and post the returned share link to your team channel — Slack goes back to being the doorbell, and the document lives at a stable URL.

One distinction worth knowing

Comma also has its own scheduling primitive, called routines — a hosted cron that re-runs a skill and republishes a report, daily on Pro and hourly on Team. The two compose either way:

  • Claude Code Routines drive, Comma is the destination. The schedule and the execution live in your Claude Code environment; Comma receives the output over MCP. Pick this when the task needs your local code, credentials, or tooling.
  • Comma routines drive end to end. The schedule, the run, the cost cap, and the publishing all live in Comma. Pick this when the report is the whole job and you'd rather not keep a machine awake for it.

Either way, the review surface — anchored comments, share roles, revision history on a stable link — is the same.

Try the loop

The Free plan covers the whole loop on demand: publish over MCP, share the link, collect anchored comments, let the agent respond. Scheduling with Comma's own routines starts at $9/month on Pro.

Install the plugin or create your first report →

Related