Work on Repeat documentation

AI routines as a service — bounded responsibilities that run on a schedule, use approved tools, and leave a trace.

Work on Repeat turns recurring operational work into a routine: a named responsibility with stable instructions, a model, a schedule, an explicit tool policy, and a durable record of every run.

These docs are organised in three parts.

The operating model

Name the responsibility

Choose recurring work that deserves a schedule — a weekly security review, a monthly analytics report, a quarterly revenue summary. Give it instructions that stay stable from run to run, and put the values that change into variables.

Bound it

Pick a model, a five-field cron schedule with an IANA timezone, and execution limits (up to 25 steps and up to 20 minutes per run).

Approve its tools

Connect GitHub, PostHog, Zernio, or any Streamable HTTP MCP server. Every discovered tool starts disabled, and letting one run automatically takes a second, separate confirmation.

Prove one run

A routine cannot be scheduled until one manual run of its current configuration has succeeded. Inspect the model steps, tool calls, token usage, and result before you commit to a cadence.

Activate the schedule

The scheduler claims due routines every minute, runs them in their timezone, and records everything. Signed webhooks tell your systems when a run finished.

Start here

Routines people run

RoutineCadenceTools it usually needs
Weekly security auditMonday morningGitHub (read)
SEO and AI-SEO auditWeeklyGitHub (read, and write a PR)
Blog research and draftingWeeklyResearch sources, draft output
Social creation and publishingSeveral days a weekZernio MCP
Quarterly revenue reportQuarterlyStripe MCP (read-only)
Analytics reportWeekly or monthlyPostHog MCP (read)

Each of these has a template in the product and a guide here. Start with Getting started.

On this page

Edit this page on GitHub