# Fabro
Fabro is an open-source platform for **orchestrating [[AI Agents|AI coding agents]] with version-controlled workflow graphs**. Instead of the interactive prompt-wait-review loop, you define a process once as a graph that wires together AI agents, shell commands, and human decision gates into a repeatable, long-horizon coding pipeline that can run unattended.
> The dark software factory for expert engineers.
## The premise
Interactive agent sessions don't scale: every run is a manual, one-off conversation. Fabro turns the *process* into a durable artifact ; a deterministic [[Cook (Orchestration CLI)|workflow graph]] (authored in Graphviz DOT, with branching, loops, and parallelism) so the same long-running coding workflow can execute reproducibly, 24/7, via API. It targets small teams of expert engineers who want auditable, structured automation rather than unpredictable autonomous agents.
## Key capabilities
- **Workflow-as-code** ; processes defined as version-controlled DOT graphs with branches, loops, and parallel stages
- **Multi-model routing** ; assign providers/models per node via **CSS-like stylesheets** (route cheap tasks to small models, hard ones to frontier models)
- **Human-in-the-loop** ; approval gates and mid-turn steering of a running agent
- **Quality gates** ; layer deterministic verifications (tests, linters, LLM judges) into the graph
- **Isolated cloud sandboxes** ; agents run in throwaway sandboxes with SSH access and automatic cleanup
- **[[Git]] checkpointing** ; a checkpoint at each stage for traceability and reversibility
- **Observability** ; queryable event streams, checkpoints, and automatic retrospectives (cost, duration, execution metrics)
- **Interfaces** ; REST API + a React web UI
## Stack
A single compiled [[Rust]] binary with zero dependencies (plus an API server and web UI). Open source.
## Where it fits
- **vs interactive harnesses** ([[Claude Code]] et al.): those are great for one session; Fabro encodes the repeatable multi-step *process* around them, like [[Cook (Orchestration CLI)|Cook]] but with explicit DOT graphs, sandboxes, and Git checkpoints
- **vs [[Paseo]]**: Paseo multiplexes interactive agents across devices (human-in-the-loop, mobile); Fabro is about *unattended, version-controlled pipelines* with verification gates ; different ends of the orchestration spectrum
- **vs [[Ralph TUI|Ralph-style]] task-list loops**: same long-horizon ambition, but graph-structured and deterministic rather than a single progression gate
## References
- Site: https://fabro.sh/
- GitHub: https://github.com/fabro-sh/fabro
- Documentation: https://docs.fabro.sh/getting-started/introduction
- Roadmap: https://fabro.sh/roadmap
- Blog: https://fabro.sh/blog
- Changelog: https://docs.fabro.sh/changelog/2026-05-26
## Related
- [[Cook (Orchestration CLI)]]
- [[Paseo]]
- [[Ralph TUI]]
- [[AI Agents]]
- [[AI Agent Harness]]
- [[How Coding Agents Work]]
- [[Claude Code]]
- [[Git]]
- [[Rust]]