What is ultraswarm?

Ultraswarm is a Node.js tool that coordinates multiple AI coding assistants working in parallel on software tasks. Rather than using a single AI tool on a codebase directly, Ultraswarm acts as an orchestrator: it breaks a coding plan into tasks, assigns each task to one of several supported AI tools (called workers), runs them in isolated Git branches so they do not interfere with each other, and then merges the results after a review step. The supported worker tools include Codex, Claude Code, Grok, and a provider-agnostic CLI called pi that can also run local models through Ollama. Each worker operates in its own Git worktree, a separate checkout of the repository, so changes from one task are isolated until they pass review. The main brain for decomposing plans and reviewing results defaults to Claude, though an API key can substitute for the Claude CLI. A plan is defined as a JSON file listing tasks with descriptions, target files, complexity scores, and optional constraints on which paths the worker is allowed to touch. Ultraswarm ranks available workers based on capability fit and historical cost and latency data, then routes each task accordingly. After workers finish, the run waits in a review state. You inspect logs and status before separately approving the merge, which is a fast-forward-only Git operation. If the base branch moved while workers were running, a recovery command brings the run back to a mergeable state. Configuration controls which workers are enabled, how many can run in parallel, which file paths are off limits, spending caps in dollars, and whether each run requires explicit approval before starting or before merging. Workers can also be run in containers for stronger isolation. Custom aliases let you register your own named worker configurations that extend a built-in worker type with different models or invocation settings, useful for running several locally hosted models through the same CLI binary. State is stored in a local SQLite database so runs survive interruptions. Commands cover the full lifecycle: run, status, logs, cancel, resume, merge, and an export command that produces a provenance record of what each worker did. Node 22 or newer is required.