The framework for running simple, secure, and durable autonomous AI agents.
OpenRoutines helps you run autonomous multi-player assistants that know how to work on a team. Scaffold one in minutes, test it locally, deploy it on your own infrastructure – then maintain it like any other software you ship.
curl -fsSL https://get.openroutines.dev/install.sh | bash
Why OpenRoutines?
Plenty of tools will run a prompt on a schedule. OpenRoutines helps you build autonomous agents that work with your team to do a job.
-
Agent-as-code
No chat interface mystery meat or "cross fingers" updates. Every agent is a single git repository that you ship like the rest of your software.
-
Teamwork ready
OpenRoutines agents knows how to work on teams out of the box. You focus on the routines that do the work; the teamwork takes care of itself.
-
Secure by design
A sandboxed, deny-by-default design and encrypted credentials keeps OpenRoutines agents secure, trustworthy, and enterprise-ready.
Example agents
See what kind of agents you can build with OpenRoutines
-
Product manager
Gathers research from support and sales calls, grooms the roadmap, and updates docs as work ships.
-
IT operations
Triages the helpdesk queue, tracks software licenses and unused seats, and flags renewals before they auto-bill.
-
Release notes
Turns merged pull requests into release notes a customer can read, then posts them the moment a tag ships.
-
Incident scribe
Builds the timeline while an incident is still open, then drafts the postmortem from what it recorded.
-
Market watch
Follows competitor changelogs, pricing pages, and job postings, and reports the changes that matter each week.
-
Recruiting coordinator
Screens inbound applications against the role, schedules the first round, and keeps the pipeline current.
Frequently asked questions
What exactly is an OpenRoutines agent?
A single git repository that builds into one Docker container. Routines are markdown files: frontmatter declares the scope, the body is the prompt.
Where can I run one?
Anywhere a container runs -- a VPS, Fly, Render, your own homelab. There's no database, queue, or secrets platform to stand up behind it.
Which model providers are supported?
Any provider opencode supports. The provider is configuration, so you can move an agent between models without rewriting its routines.
How are credentials handled?
They're encrypted in the repository and scrubbed from logs. A routine only gets the credentials its frontmatter explicitly grants.
What can an agent actually reach?
Nothing until you say so. Runs are sandboxed and the shipped container listens on no ports, so an agent's entire reach is greppable in its own files.
How do I try one before deploying it?
Scaffold an agent and run it locally against real routines, then deploy the same container to your own infrastructure when you're happy with it.