mancode adds workflow discipline to AI coding agents like Claude Code and Cursor

The problem it targets: agents that over-build

The README's before-and-after example is the whole pitch in one image. Ask an AI agent to add a logout button and, unrestrained, it will create a new component, a new style file and new color variables. With mancode installed, the agent first sees that your project already has a Button component and design tokens, and the README shows the restrained result: a single Button element with an onClick handler.

mancode is a workflow layer for coding agents, installed from npm and wired into each platform you use. Its answer to over-engineering is procedural: before writing code, the default workflow pushes the agent through six questions, from what problem this solves and whether an existing implementation can be reused, down to what the model is unsure about. The audience is teams that already run AI agents and want reuse-first habits, reviewable plans and fewer surprises in the diff.

Five modes, from daily training to playoffs

The mode system is the product's spine. Plain solo is the default: lightweight, no required identity, session or formal plan, with the model choosing its implementation approach and running verification proportional to risk. The playoffs-grade modes add ceremony where it pays. The /man workflow researches the project, pushes clarification questions that would change the plan, proposes options and produces a persistent plan that does not auto-start implementation.

Around those two poles sit the rest: /manba, /mansolo for a managed handoff where a solo session inherits the original plan, scope, review and acceptance gates, /manteam for shared typed context through a .mancode/shared directory, and /manps, a project health scan that detects stale TODOs, unused dependencies, risky dependencies, mixed icon systems and hardcoded design values. The metaphor is locker-room bravado, but the mechanics underneath are ordinary engineering governance.

Installing and wiring your agent

Requirements are Node.js 22.5.0 or newer; Git is optional, and without it team auto-detection simply degrades to solo. Installation and platform binding follow the README's block:

npm install -g mancode
cd your-project
mancode init                      # 交互式选择平台
mancode init --platform cursor    # 或显式指定一个/多个平台
mancode init --platform all

The init command is interactive and treats detected agents as hints, so nothing installs silently. After that, each platform keeps its native entry point and receives a static bootstrap: Cursor gets rules and commands under .cursor, Codex and several others get a managed AGENTS.md block plus project skills, Copilot gets instruction blocks and prompts. If you add Git or sources later, mancode refresh-project refreshes the project facts. Updating is a numbered menu via mancode upgrade as of v0.6.9.

Continuity: context that survives the session

The Continuity runtime is the less advertised half and arguably the more useful one. Workflows persist their artifacts under .mancode/<namespace>/workflows/<ULID>/, including research, plans, review reports and summaries. TaskRef, Context Packs and explicit sessions carry decisions forward, and a user-confirmed glossary plus shared decisions reduce semantic drift across sessions and agents.

Interrupted work is treated as a first-class case: delivery records, checkpoints, reframe and operation recovery exist so a half-finished state is not mistaken for a final one. Version 0.6.7 added a per-project progress page generated at init, a kanban of tasks, decisions and timelines. For teams, /manteam shares confirmed information as typed entities, and the README is explicit that authority over a task is coordinated rather than assumed.

Privacy and redaction before sharing

mancode's privacy story is local-first. Before you paste logs, config snippets or client material into a share, the tool can inspect locally for credentials and personal information and generate a redacted copy. Shared content gates can block sensitive writes, and a local model gateway can substitute values in supported model request fields before they leave for the upstream provider.

The README also states the CLI ships no telemetry. For teams operating under client confidentiality or compliance habits, this section is the difference between an interesting experiment and something that can actually be adopted; redaction happens on your machine, and the gates are explicit rather than ambient. It is not a DLP suite, and the README does not claim to be, but the default posture is conservative in the right direction.

The discipline mechanics underneath

Several rules in the README read like scar tissue, in a good way. A finding must carry concrete evidence and a user impact. A reviewer process exiting cleanly does not mean the review ledger passed; a verification command returning zero does not mean acceptance criteria are met; the structured ledger and completion gate decide. Models may choose tools freely but cannot rewrite approved goals and acceptance criteria.

Version 0.6.8 aligned man and manba review coverage with the actual diff, added a read-only review inspect inventory, and introduced optional delivery gates declaring TDD scenarios, fix budgets and required GitHub CI checks. Those new gates report themselves unavailable on Windows before starting, a limitation stated in the release notes. The last push was on 2026-09-16, with three releases in four days before it, so the mechanics are still landing fast.

How it compares with native plan modes and spec kits

Every serious agent now ships some planning discipline: Claude Code has plan mode, Cursor has rules, Codex has its own instruction files. Those are native and friction-free, but they are per-platform, and their memory rarely extends past the current session. Spec-first toolkits such as GitHub's spec-kit push in the same direction from the other side: define what to build before code exists.

mancode's differentiator is the combination: one governance layer across eight platforms, evidence rules attached to review, continuity artifacts on disk, and redaction before anything is shared. The cost is also real: another CLI, another directory of state, and a vocabulary of modes your team has to learn. If you work in a single agent and trust its native plan mode, you may not need it. Across several agents and a team, the common layer starts to pay.

Editorial conclusion

mancode fits teams juggling multiple AI coding agents that need reuse-first habits, reviewable plans and context that survives sessions, and who accept a Node 22.5+ CLI and AGPL-3.0 obligations in exchange. Skip it if you work in one agent with a plan mode you trust, or if adding a state directory to every repo is a nonstarter. First verification: run npm install -g mancode, mancode init for your platform, then give the agent a small UI task and check whether it reuses your existing components instead of inventing new ones.

Frequently asked questions

Official sources

1. License: AGPL-3.0
2. Project website
3. README
4. Releases
5. whitelonng/mancode on GitHub