Fabro

Fabro is a workflow engine for AI coding agents. You define the process as a version-controlled graph, agents execute it, and you intervene at the points you chose rather than reviewing a large diff after the fact.

Workflows are written in the Graphviz DOT language, with branching, loops, parallelism, and human gates. Each node is a stage: agent nodes run multi-turn LLM sessions with tool access, command nodes run shell scripts for builds, tests, and linters, human gates pause for approval or revision, and parallel nodes fan out and merge. Because the graphs are plain files, they diff and review like any other source file.

Stylesheets written in CSS-like syntax route each node to a model and provider, with automatic fallback chains, so planning can run on a cheap model while coding runs on a frontier one. Verification is part of the graph too: test suites, linters, type checkers, and LLM-as-judge stages sit alongside the agent nodes, and a failure triggers a fix loop.

Runs can execute in isolated cloud virtual machines with snapshot-based setup, network controls, and automatic cleanup. You can open a shell into a running sandbox or expose one of its ports for live debugging.

Every stage commits code changes and execution metadata to Git branches, so a run can be resumed, reverted, or traced. Durable events, checkpoints, conclusions, and stage outputs make a finished run inspectable and exportable, and each run produces a retro covering cost, duration, and files touched.

Fabro runs as a server, either on your own machine or self-hosted from a Docker image, and exposes a REST API with server-sent event streaming alongside a React web interface for queueing and tracking runs. It is distributed as one compiled Rust executable that needs no Python, Node, or Docker runtime.