Harness Implementations Compared

A living comparison of major AI agent harness implementations. Last updated: April 2026.

Overview

Project Stars License Model Support Memory Model Architecture
OpenClaw 355K Apache-2.0 Any model User-owned files Thin harness + skills
Claude Code Proprietary Claude only Platform-managed Monolithic
Codex Proprietary GPT only Encrypted summaries Cloud HaaS
Cline ~60K Apache-2.0 Multi-model Session-based VS Code extension
Aider ~43K Apache-2.0 Multi-model Git-based CLI tool
Hermes Agent 70K MIT Multi-model Configurable Multi-agent framework
Nexu 2.3K MIT Any model User-owned files Desktop client + harness

Detailed Breakdown

Context Management

Project Max Context Overflow Strategy Context Prioritization
OpenClaw Model limit Summarization + sliding Configurable
Claude Code 200K Built-in compaction Automatic
Codex Model limit Cloud-managed Automatic
Cline Model limit Truncation Manual
Aider Model limit Repository map Automatic

Memory Persistence

Project Memory Type Cross-session Exportable Editable
OpenClaw MEMORY.md + daily files
Claude Code CLAUDE.md + platform Partial
Codex Encrypted summaries
Cline .cline/ directory
Aider Git history
Nexu MEMORY.md + wiki/

Skill / Tool System

Project Skill Model Community Skills Install Method
OpenClaw SKILL.md based 5,400+ openclaw skill install
Claude Code Built-in + MCP Limited MCP config
Codex Built-in No
Cline MCP integration Via MCP MCP config
Aider Built-in Limited
Nexu OpenClaw skills 5,400+ GUI or CLI

Multi-Agent Support

Project Multi-agent Coordination Observability
OpenClaw Sub-agents + sessions Hub-and-spoke Session logs
Claude Code Limited (tool use) Sequential
Codex Codex tasks Parallel tasks Dashboard
Hermes Agent Native multi-agent Configurable Built-in
Nexu Sub-agents + ACP Hub-and-spoke Dashboard

This comparison is community-maintained. If you spot inaccuracies, please open an issue.