ATLAS is a local coding agent that combines planning, verification, and repair around small models to run software tasks on local hardware. It includes a TUI, a proxy orchestrator, a multi-phase V3 pipeline, and a geometric lens for scoring.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (2 so far).
What it is
ATLAS is described as a local coding agent that brings frontier-style reasoning and verification to compact open models. It emphasizes planning, candidate generation, quality scoring, sandboxed testing, and repair to enable smaller models to tackle software work locally without hosted APIs.
How it works
ATLAS comprises several components:
- atlas-tui: a native Bubbletea terminal UI serving as the chat client with a live pipeline view, slash commands, and multiple input modes.
- atlas-proxy: a Go agent loop that orchestrates the system, classifies file operations by complexity, enforces grammar for output shapes, and applies mitigations to steer edits.
- V3 Pipeline: multi-phase code generation including PlanSearch, DivSampling, Budget Forcing, PR-CoT Repair, Refinement Loops, and Derivation Chains.
- Geometric Lens: an energy-based scorer using a C(x) cost field and an XGBoost-based G(x) quality predictor, plus RAG/PageIndex for retrieval and a Confidence Router for routing.
- Sandbox: isolated execution for build verification, supporting multiple languages and pre-screening with tests.
- llama-server: local LLM inference options with GPU acceleration across NVIDIA, AMD, Apple Silicon, and Vulkan, plus grammar-constrained decoding and self-embeddings.
Full documentation is available in the docs/ directory, covering setup, architecture, configuration, and benchmarks.
Getting started
One-shot install is available via a bootstrap script:
curl -fsSL https://raw.githubusercontent.com/itigges22/ATLAS/main/scripts/atlas-bootstrap.sh | bash
There are alternative, non-piped options for reviewing and pinning a release, for example:
# Pinned to a release: script, checkout, and images all at the signed tag
curl -fsSL https://raw.githubusercontent.com/itigges22/ATLAS/v3.1.3/scripts/atlas-bootstrap.sh \
| ATLAS_BOOTSTRAP_REF=v3.1.3 bash
# Review before running
curl -fsSL -o atlas-bootstrap.sh https://raw.githubusercontent.com/itigges22/ATLAS/main/scripts/atlas-bootstrap.sh
less atlas-bootstrap.sh
bash atlas-bootstrap.sh
Running the script from any project directory launches the stack with atlas.
Requirements include GPU with 16 GB+ VRAM and support for NVIDIA CUDA, ROCm, or Apple Metal; Docker or Podman runtime; Python 3.9+; disk around 20–22 GB for model weights and images.
Recent releases
Latest release is v3.1.3 Maia (2026-07-06) focusing on production-platform upgrade/rollback, SQLite state store, signed manifests, structured logs, interactive permissions, and session resume. Previous releases include v3.1.2 Maia (2026-06-17) and v3.1.0 Maia (2026-05-12) with features like native Bubbletea TUI and streaming activation.
Traction
Stars: 2064 Forks: 182 Open issues: 13 Language: Python License: AGPL-3.0 Created: 2026-02-01 Last push: 2026-07-20
Behind the repo
ATLAS is hosted by GitHub at itigges22/ATLAS and references a community around local-first AI tooling, with documented architectural components and release notes.
Caveats
License is AGPL-3.0. Known limitations discuss Linux Docker stack with a native macOS path, current registry models not formally benchmarked, occasional inconsistencies with new feature additions, and slower grammar-constrained decoding (around 51 tok/s on llama-server). The latest roadmap items describe V3.1.3 and V3.1.2 as current and recent milestones; hardware support spans NVIDIA, AMD ROCm, and Apple Silicon with Vulkan as a preview. The setup notes include specific GPU, runtime, Python, and disk requirements, plus macOS setup paths.