Understanding Taxiway
Taxiway creates isolated labs for running and comparing AI agent orchestrators on Lima or Docker, each with a LiteLLM gateway, Caddy proxy, and optional Langfuse observability. This documentation covers how to install, use, understand, and contribute to the tool.
Reference
| Page | Purpose |
|---|---|
| Concepts | Core product model |
| CLI Usage | CLI command reference |
| Configuration | Runtime paths, drivers, and observability |
| Architecture | System design |
Drivers
Drivers create and operate labs.
| Driver | Best fit | Requirement | Details |
|---|---|---|---|
lima | Local labs on macOS and Linux | limactl on PATH | Lima |
docker | Container-backed labs, Linux development, Docker-backed end-to-end coverage | docker on PATH and daemon | Docker |
Taxiway auto-selects Lima when limactl is available, then Docker when Docker is
available.
Use --driver <name> to force a driver for one command, or set LAB_DRIVER to
choose a default driver for the current shell.
Orchestrators
Orchestrator adapters define what gets installed, verified, authenticated, and started inside a lab.
| Type | Description | Supported Agents | Details |
|---|---|---|---|
claude-code | Anthropic Claude Code CLI | Claude Code | Claude Code |
codex | OpenAI Codex CLI | Codex | Codex |
gastown | Gas Town HQ and workspace shell | Claude Code | Gas Town |
Inspect what an adapter provides for a given type:
taxiway describe <type>
Typical flow:
- Choose an adapter type.
- Run
taxiway up <lab> --type <type>, optionally with--repo <url>. - Complete adapter authentication when required.
- Attach to the lab with
taxiway shell <lab>. - Run the workflow in the attached shell or orchestrator session.