Comma

Share what Codex produced

Codex does the work in a sandbox and tells you about it in your terminal. That is the right interface for the person who ran it and a dead end for everyone else: a scrollback buffer is not a document, it does not survive closing the window, and it renders a comparison table as ASCII if it renders one at all.

So the summary of a three-hour refactor gets summarized again, by you, badly, into a Slack message.

Tell it where finished work goes

Attach the MCP server (see Codex setup — a block in ~/.codex/config.toml and a scoped token), then write the rule down where Codex will read it every time. AGENTS.md is exactly the file for this:

## Publishing results

When you finish work that a human needs to read — a refactor summary, a
security review, a benchmark comparison, a postmortem — write it as a
self-contained HTML document and publish it to Comma, then reply with the
URL and a two-line summary. Do not print the full document to the terminal.

Update report `rep_xxxxxxxx` rather than creating a new one, so the link in
our README keeps working.

Now every session ends the same way: a URL. The document is styled, the tables are tables, and the person who asked for it can open it on a phone.

Why a report beats a scrollback paste

  • It renders. The HTML is stored verbatim and served inside a sandboxed iframe with scripts enabled, so a chart the agent generated still works and a diff stays aligned.
  • Readers need nothing. No OpenAI account, no terminal, no repo checkout — visibility is private, team, domain-gated, or link, set per report.
  • Comments anchor to the line. "This assumes we're still on Postgres 14" pins to the sentence that assumed it, and stays there — see commenting on HTML.
  • Codex can read the replies back through the same MCP server, so the next run can start from "resolve the open comments" — see letting an agent respond to comments.
  • One URL over time. Updating the same report appends a revision; the old version stays in the history, and any two revisions can be diffed.

Worth knowing

  • Scope the token to reports:write. An agent that can publish but not read is a much smaller thing to hand a sandbox — see scoped tokens for AI agents.
  • HTML body: 5 MB. Screenshots and archives go in as assets.
  • Nothing re-executes. The published page is a snapshot; a routine is what refreshes one on a schedule.
  • Same pattern, every runner. The instruction file changes name — AGENTS.md for Codex, .cursor/rules for Cursor, CLAUDE.md for Claude Code — and the behaviour is identical.

Try it

Comma is free — unlimited reports, unlimited commenters, unlimited revision history.

Attach the MCP server →

Related