my-claude

my-claude is a comprehensive agent harness for Claude Code, designed to streamline AI-assisted development through a meta-orchestrator called Boss. It bundles 32 curated agents, 139 skills, 2 named workflows, 2 LSP servers, 8 hooks, and 3 MCP servers into a single plugin installation. Boss operates as a meta-orchestrator that never writes code itself but instead discovers, classifies, matches, delegates, and verifies tasks. It scans available agents, skills, MCP tools, and hooks at runtime to build a live capability registry. When a user submits a request, Boss classifies the intent, matches it against available capabilities using a priority chain (P0 gstack skill → P1 exact skill → P2 specialist agent → P3 multi-agent orchestration → P4 general-purpose fallback), delegates the task to the most suitable specialist with a structured prompt, and then verifies the results by reading changed files, running tests, linting, and building. The harness includes a 3-phase sprint workflow for end-to-end feature implementation: Design (interactive), Execute (autonomous), and Review (interactive). It also provides structured final reports with tables for changes, work summaries, verification results, deliverables, and remaining issues. Named workflows include code-review-fanout (parallel dimension reviewers with adversarial verification) and upstream-audit (per-upstream analysis for pin deltas, overlaps, security signals). Boss discovers agents and skills through hooks: a registry builder scans all loadable agents and skills at session start, a routing summary injects top candidates per intent, a route hint classifies prompts, and an advisor gate ensures advisory consultation for architecture/ambiguity/plan-review intents. Adoption tracking records whether users accept or reject agent/skill outputs per intent, using weighted events to reorder routing candidates within bands. A learning loop detects user corrections and approvals, queuing suggestions that can be applied via `learn approve`. Installation is straightforward via a shell script or as a Claude Code plugin. The installer integrates multiple open-source tools as submodules or CLI dependencies, including oh-my-claudecode, oh-my-openagent, andrej-karpathy-skills, everything-claude-code, anthropic/skills, gstack, superpowers, codeburn, serena, headroom, archify, agency-agents, ast-grep, context7, exa, and grep.app. Model routing assigns different Claude models based on task complexity, from opus for orchestration and deep analysis to haiku for quick lookups. Effort tiers control how hard a model thinks, with precedence from environment variable over frontmatter over session level. The project is MIT-licensed and supports multiple languages including English, Korean, Japanese, Chinese, German, and French.