whitelonng/mancode

What it solves

Many AI coding agents tend to over-engineer solutions, ignore existing design systems, or skip critical planning and review phases. mancode provides a local workflow orchestration layer that imposes engineering discipline on these agents, ensuring they reuse existing code and follow structured governance processes before implementing changes.

How it works

It acts as a wrapper around existing AI agents (such as Claude Code, Cursor, and GitHub Copilot) rather than replacing them. It installs bootstrap adapters that allow users to trigger specific work modes via slash commands. These modes manage task state, session continuity, and evidence-based reviews in a local .mancode directory, keeping the task context separate from the temporary chat history.

Who it’s for

• Developers using AI coding agents in backend, web, mobile, or CLI projects.
• Teams requiring shared memory, decision records, and coordinated task handoffs across multiple agents or developers.
• UI-heavy projects that want to force AI agents to adhere to existing design tokens and component libraries.
• Projects needing a repeatable, audited code review and verification process for critical changes.

Highlights

• Multi-Agent Support: Integrates with Claude Code, Cursor, Codex, GitHub Copilot, ZCode, Kimi Code, Qoder, and DeepSeek Harness.
• Tiered Work Modes: Offers solo for lightweight tasks, /man for high-governance planning and 9-step execution, /manba for bug diagnosis, and /manteam for shared team memory.
• YAGNI Enforcement: Pushes agents to prioritize reusing existing code and standard libraries over creating new abstractions.
• Continuity: Enables resuming tasks across different chat sessions or clients using stable TaskRef identifiers.
• Evidence-Based Review: Requires actual verification evidence (logs, test outputs) rather than just AI self-assertions to close a task.
• Project Health Scanning: Includes manps to detect stale TODOs, unused dependencies, and hardcoded design values.