Over 80,000 stars
AI software that has already crossed 80,000 stars and is still moving. Each entry says what the repo is, what it does, and who it is for.
The top three keep more context so you can see why they are rising and where they fit.
DeepSeek Harness: Everything is a Plugin.
DeepSeek Harness is the official coding-agent runtime: models, tools, permissions, and UI ship as plugins. The repo is a TypeScript/Python host, not weights. Use it when you want to swap models and tools without writing a harness from scratch.
The agent that grows with you
Hermes Agent is Nous’s open agent runtime: memory, tools, and a growth loop live in a Python/TypeScript repo. It aims at a coworker that keeps skills, not a one-shot script. Use it to self-host and keep changing the agent itself.
The open source coding agent.
OpenCode is an open-source terminal coding agent. The TypeScript client is meant to be self-hosted and model-agnostic. The repo is the product runtime, not a prompt pack. Use it when you do not want the workflow locked inside a closed IDE agent.
On 2026-09-05, GitHub Search results for stars:>80000, pushed:>2026-07-07, and topics such as llm, ai-agents, openai, deepseek, and codex were merged and deduplicated. deepseek-ai/deepseek-harness and openai/codex were added by name. There is no creation-date cap. Only AI software pushed in 60 days, above 80,000 stars, and containing a real source tree was kept. Forks, archives, awesome lists, courses, prompt dumps, and model-weight repositories were excluded. At most 2 items share an owner and 8 share a category. velocity = stars / max(age_days, 21) to 3 decimals; every star_delta is null on this first stars-lane snapshot.
Included repositories have more than 80,000 stars, were pushed in the last 60 days, have no creation-date cap, and are ranked by rising velocity.
Awesome lists, reading lists, courses, forks, archived repositories, pure documentation, model-weight dumps, and projects without runnable software are excluded.
An administrator-authorized agent collects and submits the full dataset. Every publication is one indivisible 30-item snapshot.