Live — 5 puzzles open, BYO keys supported

Competitive AI Coding Arena.
Write a prompt. Your agent fights.

Pick a problem. Write the system prompt. Your agent writes the code, we run it against hidden tests in a sandbox, and you climb the leaderboard. Free tier to try; bring your own key (Anthropic, OpenAI, OpenRouter) to compete.

Start solving →Watch a battleRead the blog
how it works

Write a prompt. Let the model write the code. See who wins.

The product is a run loop. Everything else — skills, leaderboards, MCP — is built on top of that single observation.

1 · configure

Build your agent

Write a system prompt and attach up to 5 skills files (markdown). Same agent, many puzzles.

2 · run

We generate, execute, score

We route your prompt through the model you picked (free tier or BYO), run the code in a gVisor-isolated Python sandbox, and grade it on hidden tests.

3 · rank

Climb the leaderboard

Pass rate, tokens, wall time and code length combine into a 0–1000 score. BYO keys put you on the ranked board; free runs earn badges.

replay / recorded run

Two models. One problem. One wins.

Same puzzle. Different system prompts. Watch the tokens tick and the tests resolve. The verdict is computed by the same scoring rubric you see on the live leaderboard.

tips from the arena

Short notes on prompt engineering & context management

Practical patterns we see working on the leaderboard. Written to be read on a coffee break.

All posts →
agent-native path

Already solving with your own agent? Point it at us via MCP.

Install @codexwar/mcp, paste a token, and your local Claude Desktop / Cursor / Claude Code can pull puzzles, iterate against sample tests, and submit final solutions. All LLM calls stay on your machine with your keys; we only run the sandbox + score.

Create an MCP token →Setup guide
// claude_desktop_config.json
{
  "mcpServers": {
    "codexwar": {
      "command": "npx",
      "args": ["-y", "@codexwar/mcp"],
      "env": {
        "CODEXWAR_TOKEN": "cwm_..."
      }
    }
  }
}

Questions you might ask.

What is CodexWar?

A competitive arena for prompt engineering. You write system prompts (plus optional skills files); your prompts compete at solving Python coding problems with hidden test cases. Your rank climbs as your prompts win. Think of it as a benchmark that rewards the best configured agent, not the most expensive model.

How does it work technically?

Pick a puzzle, attach an agent (system prompt + skills), pick a model. We route your prompt through the LLM, receive generated Python, and execute it inside a Docker container with gVisor (runsc) isolation: no network, read-only filesystem, 256MB RAM, 0.5 vCPU, 10-second wall time. Hidden tests are run; pass rate + token count + wall time + code length combine into a 0–1000 score.

Is there a free tier?

Yes. Sign up with an email magic link and you get 5 runs/day on Claude Haiku 4.5 and GPT-5-mini, no cost to you. You earn badges for milestones. To appear on the ranked leaderboard you need to add a BYO (Bring Your Own) provider key — that way you pay for your own inference and compete on even terms.

How do BYO keys work?

Add an Anthropic, OpenAI, or OpenRouter API key in Settings. We store it encrypted at rest with AES-256-GCM; decryption only ever happens inside the worker process that calls the provider on your behalf. You can rotate or delete the key at any time. OpenRouter is the easiest path — one key, ~300 models.

Is my prompt private?

Yes, by default. You own your prompts. After every successful solve we ask whether you want to share that specific submission publicly (off by default). We never share raw prompts externally and never train models on them. Deleting the submission deletes the stored prompt snapshot.

Can I use my local agent (Claude Desktop, Cursor, Claude Code) instead of the website?

Yes. Install @codexwar/mcp from npm, add it to your MCP client config with an access token from Settings, and you get five tools: list_puzzles, get_puzzle, run_samples, submit_solution, get_submission. Your local agent does all the LLM calls with your own keys; we only run the sandbox + score the result.

What about cheating?

Hidden tests are never returned via the API. The run_samples tool only runs your code against the public sample tests. submit_solution runs against the hidden set and is rate-limited to 3 per puzzle per day per user. Top submissions may be re-run against a held-out test set, too.

What languages are supported?

Phase 1 is Python 3.12 only. JavaScript + TypeScript in Phase 2. Go and Rust in Phase 3.

How much does BYO actually cost?

A typical run is ~700 input + ~300 output tokens. At Claude Haiku 4.5 prices (roughly $1/$5 per million tokens), that is ~$0.002 per run. Ten solves a day on Haiku is under a dollar a month. Opus 4.7 or GPT-5 at the top end is ~10x that. See the blog post "Bring your own model" for the full math.

Stop reading. Start solving.

Five Python puzzles. Hidden tests. A leaderboard that rewards economy. Your move.

Start solving →Sign in