ATLAS
Adaptive Test-time Learning and Autonomous Specialization
What is ATLAS?
Adaptive Test-time Learning and Autonomous Specialization
Overview of ATLAS
What it is
ATLAS (Adaptive Test-time Learning and Autonomous Specialization) is a local coding agent that brings frontier-style reasoning and verification to compact open models. It puts more intelligence in the system around the model (planning, candidate generation, quality scoring, sandboxed testing, and repair) so smaller models can tackle real software work entirely on your own hardware, without a hosted API or per-token fees. The project is experimental and open source under AGPL-3.0. It is developed by InferStep and hosted at https://github.com/inferstep/ATLAS. The status is experimental; the 3.1 line's verification has known gaps, and fixes are planned for 3.2.0. Users should not rely on a 'verified' or 'completed' from ATLAS without reviewing its output and running their own tests.
Key features
- atlas-tui: Native Bubbletea terminal UI; the canonical chat client. Type
atlasin any project directory to launch it. Includes live pipeline view, slash commands (/add,/diff,/commit,/run), and input modes (chat,!bash,/slash). - atlas-proxy: Go agent loop that orchestrates the system. Handles tool-call routing, grammar enforcement via GBNF schemas, BiasBusters mitigations, and safety limits (turn caps, token budgets, timeouts).
- V3 Pipeline: Multi-phase code generation that turns a single prompt into a verified candidate. Includes PlanSearch, DivSampling, Budget Forcing, PR-CoT Repair, and Refinement Loops.
- Geometric Lens: Energy-based scoring over the model's own embeddings, no external oracle. Includes C(x) Cost Field, G(x) Quality Prediction, per-step scoring, and a pattern cache.
- Sandbox: Isolated execution for build verification. Supports multi-language execution (Python, Rust, Go, C, Shell, others), compilation and linting before scoring, and runs both generated and existing test suites.
- llama-server: Local LLM inference on one consumer GPU. Provides GPU-accelerated quantized inference (Q6_K / Q4_K_M) on NVIDIA CUDA, AMD ROCm, Apple Metal, and Vulkan. Includes grammar-constrained decoding and self-embeddings.
Use cases
ATLAS is designed for developers who want to run a local coding agent on their own hardware without relying on hosted APIs or per-token fees. It can be used for generating, verifying, and repairing code in a sandboxed environment. The system is suitable for tasks where smaller open models are preferred, and where control over data and execution is important. Use cases include automated code generation, testing, and iterative repair within a project directory.
Getting started
One-shot install:bashcurl -fsSL https://raw.githubusercontent.com/inferstep/ATLAS/main/scripts/atlas-bootstrap.sh | bashPrefer not to pipe a moving script into bash? Same installer, two more careful ways to run it:bash# Pinned to a release: script, checkout, and images all at the signed tagcurl -fsSL https://raw.githubusercontent.com/inferstep/ATLAS/v3.1.7/scripts/atlas-bootstrap.sh | ATLAS_BOOTSTRAP_REF=v3.1.7 bash# Review before runningcurl -fsSL -o atlas-bootstrap.sh https://raw.githubusercontent.com/inferstep/ATLAS/main/scripts/atlas-bootstrap.shless atlas-bootstrap.shbash atlas-bootstrap.shThe script detects your distro (Ubuntu, Debian, RHEL, Fedora, Rocky, Alma) and your GPU vendor (NVIDIA, AMD, Apple Silicon), installs the appropriate runtime, downloads the model weights, builds the ASA steering vector, and starts the stack. Expect 10-30 minutes. Then in any project directory, run atlas. Requirements: 16 GB+ VRAM GPU (NVIDIA CUDA, AMD ROCm, Apple Silicon Metal, or Vulkan).
License and pricing
The license is AGPL-3.0. The pricing type is open source. No other pricing information is stated.
Based on the project's public README and repository metadata.