A disciplined development workflow framework for Claude Code, featuring Plan→Work→Review cycles and a native Go guardrail engine.
Claude Code Harness is a development workflow framework designed exclusively for Claude Code CLI, currently at v4.3.3 (codename "Hokage"). Its core is the 5 Verb Workflow: /harness-setup for initialization, /harness-plan for requirements planning, /harness-work for parallel implementation, /harness-review for 4-perspective code review (security/performance/quality/accessibility), and /harness-release for automated releases.
The framework includes a native Go guardrail engine with 13 declarative rules (R01–R13, response < 10ms), protecting against sudo calls, .git/ and .env writes, out-of-project writes, rm -rf, git push --force, --no-verify, test assertion tampering, and more. Rule verification is re-runnable for audit-grade evidence.
It supports Breezing Agent Teams for autonomous agent collaboration (Planner + Critic plan review → batch parallel implementation), and /harness-work all provides one-click full-pipeline automation (experimental). Optional harness-mem integration enables cross-session memory, falling back to local JSONL logs when no external dependency is available. A PreCompact hook prevents long tasks from being interrupted by context compression.
Since v4.0, it is fully Go-based with no Node.js dependency, reducing PreToolUse latency from 40-60ms to < 10ms and SessionStart from 500-800ms to 10-30ms. It deeply adapts to Claude Code official specs (monitors/agents plugin manifests) and supports Opus 4.7's literal instruction following and xhigh effort.
Extended capabilities include OpenAI Codex delegation for parallel implementation (Codex Engine), Cursor 2-Agent collaboration mode (Cursor as PM, Claude Code as implementer), and slide generation (/generate-slide) and video generation (/generate-video).
Supported platforms: macOS (darwin-arm64/amd64), Linux (amd64). Requires Claude Code v2.1+ (recommended v2.1.111+). Windows support status is unconfirmed. MIT License.