Loom: the first principle AI agent.
An intuitive, productive coding agent. Tools, memory, long-running jobs and browser use all work the way you would expect, with no new concepts to learn.
It writes the code and finishes the job. It drives multiple tasks in parallel. It talks to you not only in text and images, but builds and interacts with task dedicated UIs. It learns from its own work, so the next run is more efficient than this one.
Free to use · bring your own model provider
What it is like to use
You reach a running session from the browser, from another machine, or from your phone over the network - a long job does not tie you to one desk.
An agent is not limited to text and pictures: it can mount a real page into the conversation - buttons, forms, a task board, an file editor, or a page you are building.
You don't feel it installing new tools, new skills, nor conversation compaction. They takes effect with no restart nor waiting.
A very high cache hit rate (typically >95%). The long part of a conversation comes back from the provider's prompt cache and saves your cost.
It maintains its own library: the trap a run hit, the endpoint that had moved, the recipe that can be reused, written down and read back by the next run.
A agent hands jobs to a fleet of sub-agents, and several tasks move at once instead of one at a time in a single queue.
Programming work, end to end: it writes and runs real code, drives a real browser, and finishes the task instead of describing it.
Install
macOS, with Homebrew
brew install loombot-ai/tap/loom
Linux
curl -fsSL https://downloads.loombot.dev/install.sh | sh
Get Started
Start it with:
loomd
Then open http://127.0.0.1:7777 - that is the web UI, and it is where a session runs.