It observes, acts, verifies, and repeats until the job is done, then runs again on its next schedule or trigger. Drive it from your agent (Claude Code, Codex, Cursor, or any MCP client) or the CLI. You stay in control, and every run lands on the record.
floom mcp install --target genericFirst npm i -g @floomhq/floom · hosted workers run today, building your own loops is coming soon.
Turns raw notes into a polished weekly company update.
Drafts follow-up emails after calls, adds CRM notes.
Every morning at 9am, a digest of unread PRs and issues.
Weekly pipeline summary, emailed to the team.
The harness runs the same loop every cadence, so the work keeps getting done without you.
Reads the trigger and pulls the context it needs from your tools and the brain.
Does the work: drafts, updates, posts, screens, whatever the worker is for.
Checks its own output against the bar, and pauses for your approval where it matters.
Runs again on the next schedule or trigger. You only hear from it when it needs you.
Emails, CRM updates, posts: the worker drafts, then waits for your yes. Approve from your agent or the CLI.
Approval gates are on by default for anything that leaves the building. Relax them per worker once a loop has earned it.
Drafted the Acme follow-up with pricing from the brain. Send?
Trigger, tools, context, output. Each step timestamped, each held action visible, every loop logged.
When a run looks off, you read it like a receipt instead of guessing what the model did. Replay or pause the loop any time.
SOPs, pricing, tone. Drop in files once and every run comes pre-loaded, cadence after cadence.
Each run logs which files it used, so you can audit the why behind every output.
A loop touches your tools with your tokens, and your tokens stay yours. Revoke at the source any time.
floom mcp install --target genericnpm i -g @floomhq/floom · then floom login --cloud