#8080th of 143 in Coding agents and AI editors
atlas
Local coding agent that plans, tests and repairs code using small GGUF models
- Stars
- 2.1k
- License
- AGPL-3.0
- Last commit
- Oct 2026
- Last release
- Oct 2026
- Language
- Python
1,345 on 31 Mar 20262,107 on 20 Oct 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
ATLAS is a terminal coding agent that runs a GGUF model through llama-server on your own GPU. A Go proxy drives a multi-phase pipeline of planning, diverse candidate generation, scoring, sandboxed testing and repair, with a Bubbletea TUI as the client. The README marks it experimental and lists known gaps in 3.1.x verification, with fixes planned for 3.2.0.
Who it is for: Self-hosters with a 16 GB+ GPU who want a local terminal coding agent
Strengths
- Runs fully local on a GGUF model; no hosted API or per-token fees
- Sandbox compiles and tests candidates in Python, Rust, Go, C and Shell
- Supports NVIDIA CUDA, AMD ROCm, Apple Metal and Vulkan (preview)
- One-command bootstrap script, with a pinned-release install option
Weaknesses
- A failing V3 candidate can be written and reported as verified (3.1.0 to 3.1.6)
- No current benchmark result; the V3.0 LiveCodeBench figure was withdrawn
- Geometric Lens gives no usable signal with default settings in 3.1.3 to 3.1.6
- Needs 16 GB+ VRAM; prebuilt CUDA image targets Blackwell, older GPUs need a rebuild
What it needs
- GPU required
- Compose
- Needs Docker or Podman, nvidia-container-toolkit (NVIDIA), llama-server, Python 3.9+
- Models: GGUF (Q6_K, Q4_K_M), Qwen3-14B
Also in Coding agents and AI editors
See all 143| Rank | Project | Score |
|---|---|---|
| 1 | ArchonYAML workflow engine that runs AI coding agents through fixed dev steps | 84 out of 100 |
| 2 | claw-codeRust CLI coding agent harness using Anthropic and OpenAI-compatible APIs | 83 out of 100 |
| 3 | Open SWEAgent that turns issues into pull requests, reviews PRs and monitors CI | 83 out of 100 |
| #4 | piTerminal coding agent with extensions, skills, and an SDK | 80 out of 100 |
| #5 | oh-my-openagentTerminal coding agent with multi-model orchestration and git-backed memory | 80 out of 100 |