A.T.L.A.S. CLI
Local coding agent with a terminal UI: it plans, drafts several candidates, tests them in a sandbox and repairs, so compact open models can handle code
- 2,108 ★ on the repository
- Last commit
- AGPL-3.0
Broad access
Why this level
- The install script sets up a container runtime and driver layer, downloads about 20 GB and starts several services on the machine
- The agent writes files and runs commands, and sandbox commands have outbound network access by default
- In 3.1.x an unverified edit can be written with a "verified" mark
Install
In your terminal, with SkillFoxx CLI
npx skillfoxx add cli/a-t-l-a-sDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The install script sets up a container runtime and driver layer, downloads about 20 GB and starts several services on the machine
- The agent writes files and runs commands, and sandbox commands have outbound network access by default
- In 3.1.x an unverified edit can be written with a "verified" mark
This is third-party code. Review the repository files before installing.
What it does
A.T.L.A.S. runs its own coding agent on your computer on a compact open model in GGUF format. The `atlas` command in a project folder opens a terminal UI: a chat, a `!bash` mode and the slash commands `/add`, `/diff`, `/commit` and `/run`. A Go proxy drives the agent loop, and the multi-phase V3 pipeline turns a request into a verified piece of code: constraint-driven planning, several candidates, quality scoring, checks in an isolated sandbox and repair. Simple edits take a short path, while harder ones get more candidates, reasoning and validation. The sandbox compiles and tests Python, Rust, Go, C and shell, running both generated and existing tests. The Geometric Lens component scores candidates over the model's own embeddings and keeps lessons from earlier sessions. Everything runs on your hardware: NVIDIA (CUDA), AMD (ROCm), Apple Silicon (Metal) or Vulkan. According to the author, the repository and prompts are not sent to external models or services, and no paid API key is needed. The project is marked experimental.
Who it is for. For developers with a 16 GB or larger GPU or Apple Silicon who want to try a coding agent on a local open model without a paid API.
Good fit when
- You want a coding agent that runs fully locally, with no paid API and no repository leaving the machine
- You have an NVIDIA, AMD or Apple Silicon GPU with memory to spare and want more from a small GGUF model
- Code has to be compiled and run through tests in a sandbox before it is written to a file, rather than taken as a raw generation
- You study verification and repair approaches around small models and want to watch the pipeline stages in a side pane
- You need your own model: `atlas lens build` trains the lens for another GGUF
Not a fit when
- You need a dependable helper for production code: the author advises not to rely on a "verified" mark and to review everything the agent writes
- Your GPU has less than 16 GB of memory: smaller setups are not listed in the requirements
- You need a Windows build or an in-editor plugin for Claude Code or Cursor: the README describes a Linux Docker install and a native macOS path
- You do not want to install Docker and download about 20 GB of images and model weights
Example request
Run atlas in this folder and ask it to add a date parsing function with tests to utils.py, then show the diff with /diffLimitations
- The project is experimental. In 3.1.x, if no candidate passes its tests even after repair, ATLAS still writes the best-scoring one and tells the agent the edit was verified. The fix is planned for 3.2.0.
- C, C++, Swift, JSX, Vue and Svelte files are checked as Python, so their candidates fail and the same limitation applies.
- The Lens component in 3.1.3 to 3.1.6 gives no usable signal: it cannot veto a candidate or raise an alert.
- There is no current benchmark result: the 74.6 percent LiveCodeBench figure for version 3.0 was withdrawn by the author.
- Requirements: a GPU with 16 GB or more of memory, Docker or Podman, Python 3.9 or newer, about 20 GB of disk for images and weights. The prebuilt CUDA image targets Blackwell (RTX 50xx); older NVIDIA cards need a local rebuild.
- Sandbox commands have outbound network access by default; set `ATLAS_SANDBOX_NET_INTERNAL=true` to disable it.
- Complex feature additions can be inconsistent, and grammar-constrained decoding is slower than unconstrained decoding.
How to disable. Stop the ATLAS Docker stack containers and delete them along with the images, model weights and the repository folder. The main README gives no dedicated uninstall command.
FAQ
How do I install A.T.L.A.S.?
Run `curl -fsSL https://raw.githubusercontent.com/inferstep/ATLAS/main/scripts/atlas-bootstrap.sh | bash`. The script detects your distro and GPU, installs the container runtime, downloads model weights and starts the stack, which takes 10 to 30 minutes. Then run `atlas` in any project folder. Be careful with piping a script: you can download and read it first or use the version pinned to a release.
Which GPU does A.T.L.A.S. need?
One with 16 GB of memory or more. NVIDIA (CUDA), AMD (ROCm, community-tested), Apple Silicon (Metal) and Vulkan in preview are supported. Intel Arc (SYCL) is on the roadmap. About 20 GB of disk is needed.
Does A.T.L.A.S. send my code to outside services?
According to the author, the repository and prompts are intentionally not uploaded to hosted models or ATLAS-operated services. Sandbox commands do have outbound network access by default; set `ATLAS_SANDBOX_NET_INTERNAL=true` to turn it off.
Can I trust the "verified" mark?
Not until 3.2.0. In 3.1.x, when every candidate fails, ATLAS can write the best one and report the edit as verified. Review what it writes and run your own tests.
How do I train A.T.L.A.S. for another model?
For your own GGUFs there are `atlas lens build` and `retrain`, and ASA calibration runs through `atlas asa check`, `build` and `publish`. Ratings inside the terminal UI are entered with `/good`, `/bad`, `/review` and `/deny`.
Discussion
No questions or reviews yet.
Related
CLI toolsA Go coding agent for open models: it turns a chat into tasks on separate branches, checks and merges the result and runs all projects in one window
A coding agent tuned for small local models, with a minimal system prompt, a planning mode and background sub-agents for research
A local-first AI agent running on your own machine via llama.cpp that drives a browser, edits files, runs approved commands and remembers context across session
Late is a coding AI agent in one Go binary: an orchestrator plans while disposable subagents code and research, on llama-server or cloud models
A local model server for Apple Silicon Macs in OpenAI, Anthropic and Ollama formats: Claude Code, Codex and OpenCode connect with one command
A Go CLI that configures the AI coding agents you already use: it gives them shared memory, a workflow and deterministic steps without replacing the agents them
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Repository stars | Main difference |
|---|---|---|---|
| A.T.L.A.S. (this entry) | no | 2,108 | |
| Aider | no | 49,706 | connects to hosted and local models and edits code in git, while ATLAS installs its own stack with a sandbox and candidate verification |
| OpenCode | no | 220,388 | a general terminal agent for many providers, while ATLAS is built around compact local GGUF models |
README in short
The README starts with an experimental-status warning and briefly says why the project exists. It continues with release and vulnerability news, a description of six parts of the system (terminal UI, proxy, V3 pipeline, Geometric Lens, sandbox, llama-server), a one-command install and a requirements table. A known limitations section lists problems in 3.1.x, followed by the roadmap, project support and license.