kodwai is an agentic coding challenge platform for developers. It lets developers solve real coding challenges locally, in their own editor, with their own AI coding agents (Claude Code, Cursor, or Codex), and scores how well developers direct the agent.
Purpose
The core purpose of kodwai is to practice and evaluate "agentic coding" or "AI-native programming" ability, that is, building software by directing AI agents rather than hand-writing code line by line. It does not test memorization of algorithms; instead, it gives ticket-sized project tasks and scores how developers guide AI agents to complete these tasks. It is officially positioned as a complement to LeetCode rather than a replacement: LeetCode is for classic algorithm interview preparation, while kodwai is for practicing the AI-related interview segments used by some companies.
Main Capabilities
Scoring revolves around three axes: Direction (how you direct), Outcome (the final result), and Lift (the improvement relative to AI alone). Direction has the highest weight, covering verification rigor, intent fidelity, edge case coverage, and more; Outcome focuses on the delivered tests and code quality, or the challenge's own scoring criteria; Lift measures the degree to which developers catch hidden traps and do better than AI alone. Under the default weights, the three axes each account for a certain proportion, while challenges with their own scoring criteria use a 45/45/10 weighting.
Each scoring run lists up to 6 moments that affected the score, using a chess-review-like approach, with evidence attached to each moment; if there are mistakes, at least 2 of them will remain on the list.
The challenge library contains 15 challenges: 3 easy, 3 medium, and 9 hard, covering backend, frontend, distributed systems, data, AI, mobile, and other areas. For example, Bookshelf REST API (backend, easy, 60 minutes), Searchable Product Listing UI (frontend, easy, 60 minutes), Idempotent ETL Pipeline with Schema Drift (data, medium, 90 minutes), Production RAG Service with Eval Harness & Cost Guardrails (AI, hard, 120 minutes), Multi-Currency Wallet Ledger with Idempotent Transfers (backend, hard, 120 minutes), Offline-First Notes App (mobile, easy, 60 minutes). Each challenge hides several easy-to-miss requirement traps.
How to use
The process has three steps: first, in the terminal, pick a challenge and run one command; the CLI will download the problem, starter files, and tests, and start the timer: npx @kodwai/cli@latest challenge <slug>; second, solve it in your own editor as usual, planning, prompting, testing, and correcting the agent when it makes mistakes; the CLI will ask which agent to use. Third, in the terminal, submit your code, commit history, test run results, and agent conversation log to get a score out of 100: npx @kodwai/cli@latest submit. The CLI only collects content from the challenge folder.
Scoring and leagues
Developers have an Elo-style rating based on Direction Rating, and are divided into tiers accordingly: Bronze (0 to 999), Silver (1000 to 1149), Gold (1150 to 1299), Platinum (1300 to 1449), Diamond (1450 to 1599), Master (1600 to 1799), Grandmaster (1800 and above). A new league starts every Monday at 00:00 UTC, and the first scoring run will be grouped into a group of up to 30 people; points are based on the best score for each challenge multiplied by a difficulty coefficient (easy 1x, medium 1.5x, hard 2x), the top 5 are promoted and the bottom 5 are demoted. Leagues reset weekly, but ratings and tiers are retained. The profile page is at app.kodwai.com/developers/your-username, showing score, tier, badges, and the agents used; each scoring run also generates a share card containing the total score, the three axis scores, and the best moment.
Team use
Officially, there is also a private interview feature for hiring teams: run interviews based on your own questions, see the complete process of a candidate working on their own machine using Claude Code, score from 0 to 10 according to your own standards, and get the full conversation log for every step.
Limits and costs
kodwai is free for developers. The first 3 submissions are scored by the platform key, after which you need to add your own Anthropic API key in settings to get unlimited runs; the key is encrypted and used only for scoring. Passing the tests is not enough to get a high score: a single prompt with no specification, no checks, and no corrections will get a low Direction score even if the tests are all green. Supported agents are Claude Code, Cursor, and Codex. The platform does not run in a browser sandbox, but on the developer's own machine.
