Toolang

At a glance

Toolang is a command-line language and runtime for building AI agents in near-natural-language .too files. An "agic" is a reusable agent loop that works like a function: it takes inputs, lets a model call tools, and returns text or structured data. A "flow" puts agics in a fixed order, runs them in parallel or loops them, so a prompt cannot quietly skip a step. As of 2026-11-03 it is at v0.4.0, an early release.

Good fit

Developers who want agent workflows they can version, share and rerun, instead of long prompts the model may follow loosely. It fits multi-step jobs such as research fan-out, where the homepage example proposes eight tasks, researches them in four parallel lanes and merges cited findings. It also suits scheduled jobs ("chores") and checking token use and spend per run with too top.

Consider something else when

Skip it if you work on Windows, which the installation guide says is not yet supported, or if you want a visual or hosted agent builder rather than a terminal tool. Skip it too if one chat assistant already covers your task. The point of Toolang is enforcing explicit multi-step structure, and that only pays off when you need it.

Start here

With Python 3.11 or later, run uv tool install toolang (or install with pip), then too --version. Set an API key for a supported provider and run too models to check models are detected. Try a shared agent with too https://toolang.ai/lex.too chat, then create your own with too new eve and too eve chat. Type /exit to leave a chat.

Cost and requirements

Our catalog lists Toolang as open source, and the site links to GitHub, though the captured pages do not name a licence. Model usage is a separate cost: cloud models listed on models.dev (including Anthropic, OpenAI, Google, DeepSeek and OpenRouter) need your own API key and billing. You can avoid API fees with local models served by Ollama or llama.cpp, which need suitable hardware. It needs Python 3.11+ on macOS or Linux.

Limits to know

Version 0.4.0 (as of 2026-11-03) suggests a young project whose syntax and commands may still change. Windows is unsupported. The language is mostly natural language, so an agic's inner steps are still decided by the model; only flows define execution order. You can run agents straight from a URL or a GitHub reference, which means running someone else's instructions and tools on your machine unless you use a sandbox.

Still unclear

• Which open-source licence applies, since the captured pages link to GitHub but do not state one.
• What setup the sandbox option needs and how much it isolates a shared agent's tool access from your machine.
• Where run history and agent data are stored locally, and whether anything is sent beyond the chosen model provider.