Workflows and schedulers
Your coding agents, on a schedule. On your machines, under your control.
Chain Claude Code, Codex, Antigravity and GitHub Copilot into multi-step workflows. Start them on a schedule, from a webhook or with a click, and keep control of what matters.
Steps that hand over to each other
Each step is a prompt given to a CLI, with its own model and machine. Steps run one after another, in parallel, or wait for a condition.
One CLI and one model per step
Claude for the analysis, Codex for the tests, Antigravity for the docs: each step picks, and can run on another machine.
Parallel steps, conditions and loops
Steps that run at the same time, others that only start when a condition holds, bounded loops.
Parameters at launch
A branch, a ticket, a keyword: filled in at launch and used in the prompts.
Chosen secrets and MCP servers
Only the secrets the workflow declares reach the run’s terminal, masked in the output. MCP servers are allowed one by one.
Four ways to start them
The same workflow can start from a schedule, an outside call, an event or you.
On a schedule
A cron scheduler: every night at 2:00, Mondays at 7:00. As often as every 15 minutes on Pro, 5 on Business.
From a webhook or the API
A secret URL per workflow, to call from your CI, GitHub Actions or any tool.
On an event
A document added to the Drive, a task that changes status, another workflow that completes.
By hand
From the cockpit, on desktop or phone, or by an agent when the workflow allows it.
They move on their own. You decide what matters.
Unattended does not mean unguarded.
Approval steps
A step that waits for your decision before going on, with no time limit or with a deadline.
Tool policy
Allow, deny or ask, tool by tool. On an ask, Claude Code stops, then picks up where it left off once you answer.
Interactive steps
The CLI waits for you in a cockpit terminal until you approve the step.
Decide from anywhere
From the timeline or the cockpit, and with Telegram buttons on Business. Alerts by email, Telegram or webhook.
Every run leaves a trace
Real status, duration, end reason, exit code and the terminal’s last lines on failure, tokens per model. Costs are in the cockpit, by workspace, model or agent.
Machine off at the scheduled time
The launch is retried twice, after 1 then 5 minutes, then you get an alert.
A step fails
The execution stops, and you restart it from the failed step.
Start from a template
Ready-made workflows to import in a click, then adapt. All that is left is to fill in the secrets they ask for.
- Project overview in Memory: An AI agent reads your repository and writes a project overview in your SideHub Memory. Nothing to set up, no secret, no extra tool.
- Project kickoff, from idea to backlog: Start a project on an empty repository. Analyst, product manager, architect and product owner agents shape it with you, then turn the backlog into tasks.
- Daily pull request review: Every weekday morning, an AI agent reviews your open pull requests against the repository conventions and posts its findings before the stand-up.
- Nightly issue triage: Every night, an AI agent labels new GitHub issues, spots duplicates, asks reporters for missing details and leaves a prioritized digest for the morning.
- Monday engineering report: Every Monday morning, an AI agent collects last week's merged pull requests, releases, incidents and CI failures, and writes the team report for you.
- Weekly dependency updates: Every week, an AI agent audits outdated and vulnerable dependencies, applies the safe updates on a branch, runs the tests and opens a pull request.
- SEO content pipeline: keyword to pull request: From one keyword, an AI agent researches the search intent, writes the article, waits for your editorial approval, then opens a pull request on your site.
Frequently asked questions
Which plan do I need?
Workflows started by hand are included in Starter, for free. Schedulers, webhooks, triggers, secrets and MCP servers need Pro (€10 a month, 14 days free) or Business.
What if my machine is off?
A scheduled launch is retried twice, after 1 then 5 minutes. If it still fails, you get an alert and the failure shows in the history.
What if a step fails?
The execution stops and running steps are stopped. The error shows the terminal’s last lines, and you restart from the failed step.
How long can a step run?
60 minutes by default, adjustable per workflow and per step. An interactive step waits up to 72 hours.
Can I mix CLIs in one workflow?
Yes. Each step picks Claude Code, Codex, Antigravity or GitHub Copilot, with its model, and can run on another agent.
Does my code go through SideHub?
No. The workflow’s prompts are stored in SideHub, which sends them to the agent. The code stays on your machine, and model traffic goes straight from your side.
Your first scheduled workflow, tonight
Connect a machine, import a template, pick the time.