Building reliable AI agents often feels like wrestling with cloud APIs and opaque prompt chains. openfabriq, created by Matthieu Guirlinger, offers a refreshing alternative by treating agent workflows as code stored directly in your repository. Released as version 0.1.0 on November 17, 2026, this open-source tool allows developers to define teams of Claude agents that triage issues, implement features, and review pull requests. The key differentiator is that these agents run on your machine or your team's server, not a black-box cloud service.

Factory as Code

The core concept is the 'factory,' represented by an openfabriq/ directory within your repo. This folder contains a factory.yaml file that sets the boundaries, including who can start runs and which checks must pass. Agents are defined in markdown files under agents/, specifying their job descriptions, tools, and procedures. Automations in the automations/ folder dictate triggers, such as a GitHub label being applied. This structure ensures that the logic governing your AI team is version-controlled and reviewable, just like your application code.

Security and Local Execution

A major selling point is the strict security model. Agents never hold GitHub credentials; they propose outputs, but only the engine applies them to GitHub. The engine runs in a sandbox that cannot access ~/.ssh or GitHub CLI tokens, preventing agents from accidentally pushing broken code or leaking secrets. Furthermore, the tool enforces spending limits, such as usd_per_run and usd_per_day, defined in the configuration. This local-first approach means you maintain full control over the execution environment, avoiding the data privacy concerns often associated with third-party AI coding services.

Integration and Workflow

openfabriq integrates seamlessly with existing developer tools. It supports triggers from GitHub issues, labels, comments, and pull requests. For broader automation, it can connect to Slack, Linear, Sentry, and custom webhooks via a small relay server. The CLI provides commands like openfabriq init to scaffold the factory, openfabriq plan to preview changes, and openfabriq start to run the engine. A built-in studio serves a local web interface for monitoring runs, accessible only via 127.0.0.1 by default, ensuring that the management layer remains private.

Key Takeaways

  • Agents run locally on your infrastructure, keeping credentials and code private.
  • Workflows are defined as code in the openfabriq/ directory, enabling version control and review.
  • Strict sandboxing prevents agents from directly writing to GitHub or accessing sensitive keys.
  • Supports triggers from GitHub, Slack, Linear, Sentry, and webhooks.
  • Currently supports Claude Code, with Codex support planned for the future.

The Bottom Line

openfabriq is a pragmatic answer to the chaos of AI-assisted coding. By anchoring agents in your repo and running them locally, it restores control and predictability to automated development workflows. This tool is for builders who want AI agents that behave like disciplined team members, not wildcards. With version 0.1.0, it is early but the architecture is sound for serious adoption.