kodwai plugin for Claude Code, Codex and Cursor

Start, check and submit kodwai challenges without leaving your agent. The plugin also links your agent session to the challenge, so kodwai submit uploads exactly the right transcript, even when you started the agent in the folder above the challenge.

kodwai is where developers prove they can drive an AI agent. You solve a real challenge on your own machine with your own agent, then submit the session. kodwai scores it on three axes: Direction (how you steer the agent), Outcome (does it work) and Lift (did you catch the traps).

What you get

Everything you'd do on app.kodwai.com, from your agent. Ask in plain words ("where am I in my league?", "why did my last run score low?", "which badges am I close to?") and the agent picks the right command, or type the command yourself.

Command	What it covers
/kodwai:challenges	Browse and search challenges, one challenge's spec and rubric and top 10, the Challenge of the Day, the weekly sprint, events
/kodwai:leaderboard	All-time board (by agent, model, category) with your rank, any challenge's board, your weekly league and its zones
/kodwai:profile	Tier, Direction Elo, level and XP, rank, streak, mastery, badges and progress, quests (and claiming XP), Wrapped, your README rank card, anyone's public profile
/kodwai:runs	Your run history, a run's full result with every signal and the judge's evidence, share links, rating a challenge
/kodwai:account	Username, scoring key and free runs, feedback to the kodwai team and the replies, open any kodwai page
/kodwai:challenge <slug>	Start a challenge in this folder: creates kodwai-<slug>/, starts the clock
/kodwai:status	Time left and files so far, or your score once submitted
/kodwai:submit	Asks once, then submits and shows your score
/kodwai:abandon	Drops the challenge in progress without a score

Plus:

• A rules skill that tells the agent how a kodwai workspace works: you lead, it never submits on its own, it leaves .kodwai/ and the provided tests alone.
• Session hooks (SessionStart, Stop, SessionEnd) that record the session id and transcript path in <workspace>/.kodwai/agent-sessions.json. Only ids and paths are written, never content, and nothing leaves your machine until you submit.

Guardrails:

• /kodwai:challenge, /kodwai:submit and /kodwai:abandon only run when you type them.
• The agent shows you feedback, ratings, profile edits and username changes before sending them.
• Your Anthropic key never goes through the chat. key add opens the settings page (or a hidden prompt in your own terminal), and the agent won't use a key pasted into the conversation.
• Password changes stay on the website.

What this plugin runs and sends

Being exact about it:

• The skills run one program: the kodwai CLI, via npx @kodwai/[email protected] <command> (source, npm). Your agent asks your permission before running it, as it does for any shell command.

• The CLI talks to one service: the kodwai API at https://api.kodwai.com, over HTTPS.

  • To read your challenges, leaderboards, profile, badges, quests and runs, and to make the changes you ask for (claiming quests, sharing, ratings, feedback, profile and username).
  • When you submit a challenge, it uploads that challenge folder's code, git history and test results, plus the transcript of the agent session linked to it.
  • Nothing is uploaded without a command you ran or approved.

• Your sign-in is a token the CLI keeps in ~/.kodwai/config.json, created when you approve "Authorize CLI" in your browser. The CLI sends it only to the kodwai API. The plugin never reads it.

• The hook script (scripts/record-session.mjs) is local only.

  • It reads the hook's JSON input (session id, transcript file path, working folder) and three environment variables: CLAUDE_PROJECT_DIR, CURSOR_PROJECT_DIR and CURSOR_VERSION.
  • It writes the session id and transcript path to .kodwai/agent-sessions.json inside the challenge folder, and only when the folder is a kodwai challenge.
  • It makes no network calls and never reads transcript contents.

• Your Anthropic API key never passes through the plugin or the chat. You enter it on the kodwai settings page or in a hidden prompt in your own terminal.

• Privacy policy: https://www.kodwai.com/privacy

• Terms of service: https://www.kodwai.com/terms

• Support: https://github.com/kodwai/plugin/issues or [email protected]

In Codex, skills are invoked with $ instead of /: $kodwai:leaderboard.

Install

Claude Code

/plugin marketplace add kodwai/plugin
/plugin install kodwai@kodwai

Or from your shell: claude plugin marketplace add kodwai/plugin && claude plugin install kodwai@kodwai.

Codex

codex plugin marketplace add kodwai/plugin
codex plugin add kodwai@kodwai

Codex doesn't run a plugin's hooks until you trust them. Open Codex, run /hooks, and trust the three kodwai hooks. Without that the commands still work, and the CLI still links the session by itself when you start the challenge from inside Codex.

Cursor

In Cursor: Customize → Plugins → From GitHub Repository → https://github.com/kodwai/plugin.

How a challenge goes

1. Open your agent in any folder and run /kodwai:challenge rate-limiter. Pick a challenge at app.kodwai.com/dev/challenges.
   • The first time, the agent shows you kodwai's data collection notice and asks if you accept.
   • Your browser opens to sign in.
2. Work on it with your agent, in the same session. Everything happens inside kodwai-rate-limiter/.
3. Run /kodwai:status whenever you want to see the clock.
4. When you're done, run /kodwai:submit. It runs the tests, uploads the code, git history, test results and this session's transcript, then shows your score.

The CLI underneath is @kodwai/cli, run with npx @kodwai/[email protected]. You can use it directly from any terminal too.

Requirements

• Node.js 20 or newer (the CLI and the hook script both run on Node)
• git
• In sandboxed agents (Codex), the kodwai commands need network access. The agent asks to run them outside the sandbox.

Develop

npm install
npm run check     # recorder tests + manifest/schema validation + claude plugin validate

Try it locally without installing:

• Claude Code: claude --plugin-dir ./plugins/kodwai
• Codex: codex plugin marketplace add ./ then codex plugin add kodwai@kodwai
• Cursor: copy plugins/kodwai to ~/.cursor/plugins/local/kodwai and reload the window.

Layout:

.claude-plugin/marketplace.json     Claude Code marketplace
.agents/plugins/marketplace.json    Codex marketplace
.cursor-plugin/marketplace.json     Cursor marketplace (GitHub import)
plugins/kodwai/
  .claude-plugin/plugin.json        manifests, one per agent
  .codex-plugin/plugin.json
  .cursor-plugin/plugin.json
  skills/                           challenges, leaderboard, profile, runs, account,
                                    challenge, submit, status, abandon, rules (shared by all 3)
  hooks/hooks.json                  Claude Code + Codex hooks
  hooks/hooks-cursor.json           Cursor hooks
  scripts/record-session.mjs        links the session to the workspace
schemas/cursor/                     Cursor's official manifest schemas (for validation)

Keep the version the same in all three plugin.json files; npm run validate checks it.

License

MIT