Open Source · MIT Licensed
The dark software factory
for expert engineers
AI coding agents are powerful but unpredictable. Define your process as a graph, let agents execute it, and intervene only where it matters.
Stop babysitting agents.
Start orchestrating workflows.
You either babysit every step or review a 50-file diff you don't trust. Fabro gives you a middle path: version-controlled workflow graphs that orchestrate AI agents, shell commands, and human decisions into repeatable, long-horizon coding processes.
Key features
Workflow Engine
Deterministic workflow graphs
Define pipelines in Graphviz DOT with branching, loops, parallelism, and human gates. Diffable, reviewable, version-controlled.
Control
Human-in-the-loop
Approval gates pause for human decisions. Steer running agents mid-turn. Interview steps collect structured input.
Intelligence
Multi-model routing
CSS-like stylesheets route each node to the right model and provider, with automatic fallback chains.
Isolation
Cloud sandboxes
Run agents in isolated cloud VMs with snapshot-based setup, network controls, and automatic cleanup. SSH in with fabro ssh.
Versioning
Git checkpointing
Every stage commits code changes and execution metadata to Git branches. Resume, revert, or trace any change.
Analytics
Automatic retrospectives
Each run generates a retro with cost, duration, files touched, and an LLM-written narrative. Your workflows improve over time.
Interface
API server + web UI
REST API with SSE event streaming and a React web UI. Run workflows programmatically or as a service. Queue runs, track progress, manage at scale.
Distribution
Single binary, no runtime
One compiled Rust executable with zero dependencies. No Python, no Node, no Docker required.
Workflow-as-Code
Define how work gets done in version-controlled graph files. Each node is a stage — agent, command, human gate, conditional — executed by the workflow engine.
- Agent nodes run multi-turn LLM sessions with tool access
- Human gates (
hexagon) pause for approval or revision - Parallel nodes fan out for concurrent execution, then merge
- Command nodes run shell scripts for builds, tests, and linters
digraph PlanImplement {
graph [
goal="Plan, approve, implement, and simplify a change"
model_stylesheet="
* { model: claude-haiku-4-5; }
.coding { model: claude-sonnet-4-5; }
"
]
start [shape=Mdiamond, label="Start"]
exit [shape=Msquare, label="Exit"]
plan [label="Plan", prompt="Analyze the goal and codebase. Write a step-by-step plan."]
approve [shape=hexagon, label="Approve Plan"]
implement [label="Implement", class="coding"]
simplify [label="Simplify", class="coding"]
start -> plan -> approve
approve -> implement [label="Approve"]
approve -> plan [label="Revise"]
implement -> simplify -> exit
}
Verification is a
first-class concept
Fabro doesn't just produce output — it validates it. Builds and tests are gates, not suggestions. Failures trigger automatic fix loops. Human gates let you approve, revise, or steer.
Mergeable results that survive CI and code review.
Multi-model, multi-provider
CSS-like stylesheets assign models to workflow stages by ID, class, or shape. Use fast models for breadth, frontier models for correctness — with automatic fallback chains.
- Anthropic, OpenAI, and Gemini models out of the box
- Selector specificity:
#id>.class>shape>* - Ensemble strategies to maximize quality and minimize cost
/* All nodes default to fast + cheap */
* {
model: claude-haiku-4-5;
reasoning_effort: low;
}
/* Coding stages use a frontier model */
.coding {
model: claude-sonnet-4-5;
reasoning_effort: high;
}
/* Cross-critique with a different provider */
#review {
model: gemini-3.1-pro-preview;
}Ground-up observability
Full traceability across runs — every model call, every tool invocation, every decision point. Query run data with SQL. Automatic retrospectives rate each run and surface friction points.