Atlas

Every other coding agent's marginal cost is flat and its memory resets. Atlas gets cheaper, warier, and more trusted every week you use it.

Atlas is an open-source, model-agnostic, terminal-first coding agent with a persistent intelligence layer, machine-checked receipts on every change, a learning pipeline that compiles repeated LLM work into deterministic zero-token codemods, and a workspace UI. It works in your git repos — and in any plain folder.

Install

npm i -g atlas-agent      # Node 20+; installs a prebuilt SQLite (a C toolchain only if no prebuilt matches your platform)

Or run without installing:

npx atlas-agent "write a scraper for this site"

Quickstart

cd anything/                        # a git repo, or any folder at all (git optional)
export ANTHROPIC_API_KEY=sk-ant-... # or bind a role to a local OpenAI-compatible model — fully offline works
atlas                               # bare `atlas` opens the workspace UI
atlas "fix the failing test"        # or just say what you want
atlas sync                          # index the repo (free, no model calls)
atlas doctor                        # health check: store, toolchain, configs

Tip: keep projects on a local path (e.g. ~/code), not an iCloud/OneDrive-synced folder like ~/Documents — cloud-evicted files make every read a slow network download.

What you get

• Trust that's earned, not assumed — the agent starts by coding (T1), and earns its way up to auto-merge as it proves itself per path.
• Receipts — machine-checked evidence attached to every change (atlas receipts, atlas why <file>:<line>).
• Learning — repeated LLM edits compile into deterministic, zero-token codemods (atlas compile).
• A real TUI — tasks, live streaming transcript, diffs, receipts, trust heatmap (atlas ui).

Full docs, architecture, and the "why Atlas over any other agent" case: https://github.com/klingaling/atlas-agent

License

Apache-2.0