Stars across the open harness class
1.6M+
1,672,761 combined GitHub stars across 13 open agent harnesses, almost all created 2024–2026. Claude Code/Cursor/Windsurf are closed counterparts.
Biggest single harness
393.5K stars
OpenClaw, created Nov 2025 — “The AI that really does things. Any OS. Any Platform.” Model-agnostic gateway harness.
Models harnesses actually call
52% open weights
share of weekly OpenRouter requests (5.13B, week of Oct 26, 2026) hitting open-weight models — DeepSeek-V4.1-flash and GLM-5.3-flash lead.
First, the taxonomy: three layers, three different things
the ecosystem conversation mixes these up — here is the clean stack
LAYER 1
Inference runtimes
They run models — token-in/token-out engines for local or server inference.
llama.cpp · vLLM · Ollama · SGLang · TGI · MLX · Candle · ONNX Runtime · MLC · ExLlamaV2 · llamafile
LAYER 2
Orchestration frameworks
Libraries you build agents with — chains, graphs, RAG, tools, memory. You write the code; they provide the structure.
LangChain · LangGraph · LlamaIndex · DSPy · Haystack · Semantic Kernel · CrewAI · AutoGen · PydanticAI · LiteLLM · smolagents
LAYER 3 — THIS PAGE
Agent harnesses
End-user products that harness models for real work: an agent loop, tools, sessions, and a surface you actually work in.
OpenClaw · Hermes · opencode · pi · Codex · Gemini CLI · Claude Code · aider · goose · Cline · Roo Code · Kilo Code · Crush · OpenHands
Classification: license, model coupling, surface, autonomy, focus
how the major harnesses line up on the five axes that matter for choosing one
| Harness | Stars | Created | License | Model coupling | Surface | Autonomy | Focus |
|---|
| OpenClaw | 393.5K | Nov 2025 | open | model-agnostic | any-platform gateway | autonomous | general work |
| Hermes Agent | 258.9K | Jul 2025 | open | model-agnostic | CLI + desktop + gateway | autonomous + cron | general work |
| opencode | 223.3K | Apr 2025 | open | model-agnostic | terminal TUI | agentic | code |
| Codex | 132.0K | Apr 2025 | open (Apache-2.0) | vendor-leaning (OpenAI) | terminal + cloud tasks | autonomous | code |
| pi | 119.0K | 2025 | open | model-agnostic (unified LLM API) | terminal + TUI | agentic | code |
| Gemini CLI | 108.6K | Apr 2025 | open | vendor-leaning (Google) | terminal | agentic | code |
| OpenHands | 92.3K | Mar 2024 | open | model-agnostic | sandboxed web/agent | autonomous | code |
| Cline | 71.1K | Jul 2024 | open | model-agnostic | VS Code ext + CLI | agentic | code |
| goose | 56.0K | Aug 2024 | open | model-agnostic | terminal + desktop | agentic | general (MCP originator) |
| aider | 49.8K | May 2023 | open | model-agnostic | terminal | interactive, git-native | code |
| Crush | 28.9K | May 2025 | open | model-agnostic | terminal TUI | agentic | code |
| Kilo Code | 27.8K | Mar 2025 | open | model-agnostic | VS Code extension | agentic | code |
| Roo Code | 24.3K | Oct 2024 | open | model-agnostic | VS Code extension | agentic | code |
| Claude Code | — | 2025 | proprietary | vendor-locked (Anthropic) | terminal | autonomous, ACP-interop | code + general |
| Cursor agent | — | 2024 | proprietary | vendor-managed | IDE product | agentic | code |
| Windsurf | — | 2024 | proprietary | vendor-managed | IDE product | agentic | code |
Model-agnostic = any provider, including open weights via your runtime of choice. Vendor-leaning = open source but designed around one vendor’s models. Vendor-locked = closed, one vendor only. Agentic = multi-step tool-use loops; autonomous = runs unattended (schedules, background tasks).
Top models by harness-routed requests · 5.13B total · open weights in blue
deepseek/deepseek-v4.1-flash-20261017
10.9%
z-ai/glm-5.3-flash-20260826
10.3%
openai/gpt-6-luna-20260922
7.8%
deepseek/deepseek-v4.1-flash-20260910
6.6%
deepseek/deepseek-v4-flash-20260423
5.1%
deepseek/deepseek-v4-flash-20260731
4.6%
anthropic/claude-haiku-5.5-20261007
2.8%
openai/gpt-5.6-luna-20260709
2.6%
google/gemini-3.1-flash-lite-20260507
2.3%
Six of the top nine model variants are open-weight DeepSeek or Z.ai GLM flash models — harnesses route most traffic to the cheap, fast, downloadable tier.
The newest category is orchestrating harnesses, not models
| Meta-harness | Stars | Created | What it does |
|---|
| stablyai/orca | 111.1K | Mar 2026 | work with a fleet of agents (Grok Bot, Claude Code…) in one place |
| paperclipai/paperclip | 109.0K | Mar 2026 | the open-source app for managing agents at work |
| wshobson/agents | 42.0K | Jul 2025 | multi-harness plugin marketplace (Claude Code, Codex, Cursor, OpenCode…) |
| omnigent-ai/omnigent | 11.1K | Jun 2026 | meta-harness that orchestrates other harnesses |
| mvschwarz/openrig | 8.4K | 2026 | persistent agent teams from Claude Code, Codex and Pi |
| HarnessRouter/harnessrouter | 3.8K | Aug 2026 | self-hosted unified interface across harnesses (Apache-2.0) |
If 2024–2025 was the harness explosion, 2026 is the meta-harness turn — the open ecosystem is now composing harnesses the way harnesses composed models. ACP-style interop (agent-to-agent protocols) is the connective tissue to watch.