fractal

fractal, published by the plasma-ai organization under the Apache-2.0 license, runs autonomous agent loops that arrange themselves into a tree. Each node iterates toward a goal in its own git worktree and can spawn child nodes for separable subtasks. Run metadata, including cost, is stored in a local SQLite database and can be viewed in a terminal UI.

What It Is

fractal is a Python command-line tool and agent skill for orchestrating hierarchical agent loops. The root node is the user's current branch; top-level nodes branch from it, and child nodes branch from their parent. Agents iterate in tmux sessions, or in detached process groups with the headless option, and all state (runs, iterations, steps, costs, signals) is tracked locally.

How Nodes Are Launched and Bounded

The /fractal skill takes plain-language directives, drafts node instructions and a parameter table, asks for anything it could not infer, and launches the node once approved. Hard caps keep each loop bounded: iterations, depth, children, descendants, cost, and time limits per run, iteration and step. An operator can steer or stop a loop at any point, and the dashboard opens with fractal open.

Agent Backends and Setup

Five backends are supported, selected per node with an agent option: Claude Code, Codex, Grok Build, OpenCode and Oh My Pi. Claude and Codex can route through OpenRouter. Install with pip from PyPI as plasma-fractal, and install the skill through the Claude Code or Codex plugin marketplace or with the fractal install command.

Safety Considerations

The README warns that nodes run their agents without permission prompts by default, so a node can run any command its agent chooses with the user's credentials. The per-node git worktree isolates the committed branch, not the filesystem or network. The README recommends a sandboxed or disposable host for tasks that are not trusted to run unsupervised.