About
Built to disappear
into your work.
Umbra is an AI agent designed around one principle: it should feel like an extension of you, not a tool you have to manage.
The problem
Every model has a token budget, and on real projects you hit it fast. The bigger the codebase and the longer the session, the sooner the model loses context — responses degrade, and you either start over or pay for a bigger plan.
Most AI coding tools make this worse: they dump entire files into the prompt and keep the full session history forever. We wanted the opposite — an agent that stays efficient by design, so it stays sharp across long sessions and large codebases.
How it works
Umbra runs as a local daemon — a background process that keeps your project loaded so the CLI/TUI starts instantly. On first run it builds a compact AST map of your codebase: function and class signatures instead of full files, so the model sees structure without burning tokens on raw text.
Past sessions are remembered too, via a local vector database that recalls relevant solutions from earlier work. Give it a task and it plans, edits files, and runs commands — in --exec mode it reads failing checks and retries automatically until they pass.
Open source
Umbra is open source under the MIT license — the agent, the tool layer, the provider integrations, all of it. Nothing hidden, nothing phoning home beyond the model provider you configure yourself.
Bring your own API key for OpenAI, Anthropic, and others, start free with OpenCode Zen, or run everything locally through Ollama or LM Studio.
Who builds it
Umbra is developed in the open, with a public roadmap guiding what comes next.
Issues, ideas, and pull requests are welcome — see the repo for how to get involved.