Fractal

Fractal runs autonomous coding-agent loops as a tree. Each node iterates toward a goal inside its own git worktree and can spawn child nodes for separable subtasks, so the tree grows to fit the problem rather than following a fixed plan. The root node is your current branch, top-level nodes branch from it, and child nodes branch from their parent.

Every loop is bounded by hard caps on iterations, depth, direct children, total descendants, wall-clock time per run, per iteration and per step, and cost in US dollars per run, per iteration, and per step, with a reserved cleanup budget. An operator can steer or stop a node at any point. Agents iterate in tmux sessions, and run metadata — runs, iterations, steps, costs, and signals — lands in one local SQLite database that a terminal dashboard reads live.

Five agent backends are selectable per node, with children inheriting the choice, and two of them can additionally route through OpenRouter using a key from the launching shell. Nodes are driven either through a /fractal skill installed into a coding agent, which turns a plain-language directive into node instructions, a parameter table, and a launch, or through the fractal CLI directly.

Nodes run their agent with the approval gate disabled by default, since an unattended loop cannot stop to ask; the per-node worktree isolates the branch a node commits to, not the filesystem or the network.

Features

• Worktree-per-node: each loop iterates in its own git worktree on its own branch, branching from its parent
• Recursive spawning: a node creates child nodes for separable subtasks, bounded by depth, children, and descendant caps
• Cost and time ceilings: per-run, per-iteration, and per-step limits in dollars and time, plus a reserved cleanup budget
• Local run ledger: runs, iterations, steps, costs, and signals in one SQLite database with a live terminal dashboard
• Five agent backends: selected per node with --agent and inherited by children
• OpenRouter routing: two backends can route through OpenRouter with a key from the launching shell; the others reach it through their own model ids
• Skill and CLI entry points: a /fractal skill interprets plain-language directives, and the CLI exposes the same operations
• Scoped commits: a node can be restricted to committing only within named subdirectories of its worktree
• Scheduling controls: fixed iteration intervals, sleep between iterations, detached per-step sessions, and local-only mode that skips pushing