Repolode
Submit
inferstep

inferstep/ATLAS

A self-hosted local coding agent that adds planning, candidate scoring, sandboxed testing and repair around compact open GGUF models.

Open-source alternative to:GitHub CopilotCursor

View on GitHubWebsite
Stars
2,111
Forks
201
Open issues
117
Last commit
2 days ago

What is it?

ATLAS is a local coding agent designed to get more out of compact open models. Instead of relying on a single generation, it wraps the model with a multi-phase pipeline: it plans the change, generates several diverse candidates, scores them with a learned quality model that uses the model's own embeddings, runs them in an isolated sandbox for compilation and tests, and repairs failures before accepting an edit. It runs on your own hardware and does not require a hosted model API or per-token fees.

The system is made of a terminal UI (atlas-tui), a Go agent loop (atlas-proxy) that routes tool calls and enforces grammar-constrained output, the V3 generation pipeline, a Geometric Lens scoring component, a sandbox for multi-language execution, and llama-server for local inference. It targets NVIDIA CUDA, AMD ROCm, Apple Silicon Metal and Vulkan GPUs, and the bootstrap script sets up Docker containers, model weights and the stack.

The project is labeled experimental by its authors. Its README lists known gaps in verification, including cases where a failing candidate can be written as verified, so generated code should be reviewed and tested before use.

Key features

  • 01Local inference on consumer GPUs (NVIDIA, AMD, Apple Silicon, Vulkan) with no hosted API required
  • 02Planning and multi-candidate generation followed by scoring to select the best candidate
  • 03Sandboxed compilation and test execution for multiple languages before accepting edits
  • 04Repair loops that correct code after failed verification
  • 05Terminal UI with slash commands for adding files, viewing diffs, committing and running commands
  • 06Grammar-constrained decoding to keep model output in expected JSON shapes

Use cases

  • Running a coding assistant on your own machine where repository contents must stay local
  • Making multi-file or structural code edits with a compact open model and automated testing
  • Experimenting with agent orchestration around a self-hosted GGUF model
  • Studying how planning, scoring and repair layers can improve small-model code generation

Quick start

  1. 1

    Run the one-shot bootstrap installer from the README (review it first if preferred)

  2. 2

    Wait for the install (10-30 minutes, mostly model download) and ensure Docker and a 16 GB+ VRAM GPU are available

  3. 3

    In any project directory, launch the terminal UI

    atlas

Things to know

  • Status is experimental, and the README says verification has known gaps; review its output yourself.
  • Requires a 16 GB+ VRAM GPU and Docker, and older NVIDIA cards need a local rebuild.
  • Full Docker stack is Linux-focused; Apple Silicon uses a separate native macOS path.
  • Licensed under AGPL-3.0, a copyleft license with network-use source obligations.

In collections

AI coding agentsCoding with local LLMsOpen-source AI agent frameworks & appsAI coding assistants & agentsSelf-hosted appsRun LLMs locallyPrivacy & local-first tools

Topics

#ai-agent#coding-agent#local-llm#llm#self-hosted#developer-tools#local-first#ai-harness

FAQ

What is ATLAS?
A self-hosted local coding agent that adds planning, candidate scoring, sandboxed testing and repair around compact open GGUF models. ATLAS is a local coding agent designed to get more out of compact open models. Instead of relying on a single generation, it wraps the model with a multi-phase pipeline: it plans the change, generates several diverse candidates, scores them with a learned quality model that uses the model's own embeddings, runs them in an isolated sandbox for compilation and tests, and repairs failures before accepting an edit. It runs on your own hardware and does not require a hosted model API or per-token fees.
Is ATLAS free to use? What license does it have?
ATLAS is open source under the AGPL-3.0 license. Check the license terms before using it in commercial products.
Is ATLAS still maintained?
Actively maintained. The last commit to ATLAS was on Nov 5, 2026.
What can I use ATLAS for?
Running a coding assistant on your own machine where repository contents must stay local Making multi-file or structural code edits with a compact open model and automated testing Experimenting with agent orchestration around a self-hosted GGUF model Studying how planning, scoring and repair layers can improve small-model code generation
How do I install ATLAS?
Run the one-shot bootstrap installer from the README (review it first if preferred) Wait for the install (10-30 minutes, mostly model download) and ensure Docker and a 16 GB+ VRAM GPU are available In any project directory, launch the terminal UI: atlas
Where can I get ATLAS?
The source code and installation instructions are on GitHub: https://github.com/inferstep/ATLAS

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects