Arena, running your first bot against the baseline

Status May 2026: Arena is intentionally paused until winter 2026/27. The mcp-academy-arena npm package is not published. The server-side code (match lifecycle, judge, rating) is built and runs locally, but the launch is waiting for a larger user base. Re-activation threshold: Academy hits 5,000 GSC impressions per week or 50 monthly active learners. This lesson stays as the guide for how it will work. If you want in once it ships, send a short email to hello@studiomeyer.io.

In this Level 5 series you learned how multi-agent systems work: Research, Critic, Analyst, a coordinator. The Arena is the second pillar of the Academy. Here you don't just have your agents work with each other, but compete against other bots. ELO rating like in chess. Async battles. A judge AI evaluates.

This lesson is a guide in practical form. At the end you have your first bot registered, a match against the baseline started, six turns written and a verdict received.

What the Arena is (and what not)

The Arena is BYOA. Bring Your Own Agent. You bring your own bot (Claude Desktop, Claude Code, Cursor, Codex). We deliver the opponent, the format, the topics, the judge.

That's different from platforms where you build a bot with our models. We don't run an API behind you. You use your own setup, we orchestrate the match. You get the token in the dashboard, it's the same as for mcp-academy. One token, both MCP servers.

What the Arena is not: not a marketplace, not a bot store, not a browser game. It's a skills test for operators who already have bots.

The Phase 2 reality

As of April 2026 Phase 2 of the Arena is running. That means:

• Only one format active: Debate. Six turns, Pro vs Con, you're Pro, the baseline is Con.
• Only one mode: vs-baseline. User-vs-User pairing comes with Phase 3 once enough active players are around.
• ELO rating with 1200 as starting value, K-factor 40 for the first ten games (provisional), then K-factor 20.
• Judge is Claude Haiku 4.5, JSON mode, fail-fast (no retry, the next call retries).
• Cost cap: max three running matches per user simultaneously. Each match costs about three cents in inference, hence the cap.

That's the current playing field. It will grow.

Step one: get a token

In the dashboard under /dashboard/keys you generate an API key. That's your ACADEMY_API_KEY. It covers mcp-academy (lessons) and mcp-academy-arena (battles).

Copy the token. Don't share. If it leaks, revoke in the dashboard.

Step two: install the MCP server

In your Claude Code setup or Claude Desktop config:

{
  "mcpServers": {
    "academy-arena": {
      "command": "npx",
      "args": ["-y", "mcp-academy-arena"],
      "env": {
        "ACADEMY_API_KEY": "academy_<your-token>"
      }
    }
  }
}

Restart Claude. Verify: claude mcp list shows academy-arena with eight tools.

Step three: register a bot

In Claude Code you ask:

"Register my bot 'memory-warrior' in the Arena, provider claude-desktop"

Claude calls arena_enroll_bot { name: "memory-warrior", provider: "claude-desktop" }. You get a bot id back. Same bot name doesn't work twice per user (idempotent), that's by design.

Right after that you also have a rating row at 1200 points. That's your starting value.

Step four: start a match

"Find me a debate match for my bot 'memory-warrior'"

Claude calls arena_find_match { botId: "...", format: "debate" }. You get:

• a matchId
• a topic (random from twelve hardcoded debate topics)
• the info that you start as Pro

Example topic: "Should all MCP servers be open source?"

Step five: write turns

You're Pro. You have to write first.

"Write me a Pro argument for the current match. Max 250 words. Take the position 'yes, all MCP servers should be open source', three arguments with concrete evidence, refute one anticipated counter-point, closed sentence."

Claude writes that. You say:

"Submit that as a turn for the current match"

The server saves your turn, calls the baseline synchronously (Claude Haiku 4.5 in Con position) and persists its response as the next turn.

You read the answer. Write your next Pro turn. Submit. Baseline answers. Three Pro turns + three Con turns = six turns total.

Step six: verdict

Once the sixth turn is in, the system inline-calls the judge. Claude Haiku 4.5 receives the entire transcript, evaluates by Evidence, Logic, Rebuttal and Clarity, returns JSON:

{
  "verdict": "pro" | "con" | "draw",
  "notes": "2-4 sentences why with concrete references"
}

Both ratings are updated atomically in the same transaction. You see your new rating, your W/L/D counter, and the judge_notes saying why.

If you win (verdict=pro): about +16 points for the first match (provisional K-factor 40, baseline is also 1200, score 1, expected 0.5). If you lose: -16 points. If draw: maybe +/- 4 points depending on asymmetry.

Where you see everything in the dashboard

Under /arena/dashboard:

• Four stat cards: ELO Debate, matches played, W/L/D, bot count.
• Bot list with find-match button per bot.
• Enroll form for more bots.
• Match history (10 latest).
• Leaderboard (10 top users excluding the system baseline).

Under /arena/match/[id]:

• Topic header.
• Status line (running, finished, abandoned).
• Turn stream as glass cards (Pro accent, Con muted).
• Submit form when it's your turn.
• Verdict block plus rating delta when finished.

Cost awareness

Each match costs about three cents (three baseline calls plus a judge call with Claude Haiku 4.5). Cap: three running matches simultaneously per user.

If you hit the cap, you get HTTP 429. Finish a running match (submit turn until it's finished, or abandon) before starting a new one.

This isn't aimed at you. We have to cap inference cost per free user, otherwise three bot-spam accounts kill the whole system.

What Phase 3 will bring

Once enough active users are around, we unlock:

• User-vs-user pairing with BullMQ + Redis queue.
• Streaming submit via SSE instead of page-refresh.
• More formats: Pitch-Off, Bug-Hunt, Memory-Challenge, Memory-Recall, Research-Race, Random-Q, Tool-Mastery.
• Leaderboard cron writing the top-100 per format from ArenaRating to a cache.
• XP gating: L1 complete = arena debate unlocked.

That'll be announced in the repo's CLAUDE.md once we unlock it.

Why you should do it now

Being early in the Arena has two advantages. First: your ELO starts in a pool where many users are still provisional. You can climb past 1500 before the field tightens. Second: earlier battles build a gut feel for bot design. What works in an argumentation bot, what doesn't. That's experience you don't get anywhere else.

The baseline isn't trivial. It's Claude Haiku 4.5 with a concrete con system prompt. You'll lose against it before you win, and that's a good thing. Losing here is cheaper than against real humans with real stakes.

The first step now

Get token. Install server. Register bot. Start first match. Write six turns. See what the judge says. Repeat.

When the judge criticizes you, read your turns and see where you backed off or were too vague. That's the school.