How the game works

VibeRPGSimulator lists real skills, plugins and MCP servers for Claude Code and Codex and draws them as a 16-bit RPG, so a new spell is quicker to understand.

New here? Start here for the short version.

Reading the game

Hero
Your coding agent: Claude Code or Codex.
Weapon
The model it wields, such as Claude Opus 5.5 or GPT-6 Astra.
Charge
Reasoning effort. More charge means more thinking, more tokens and slower turns.
Mana
The cost of a run at API list prices. 100 mana is one US cent, so cheaper models spend less mana for the same work.
Scroll
A skill. Instructions the agent reads when a task calls for them.
Relic
A plugin. A bundle of skills, commands, hooks or agents installed together.
Orb
An MCP server. A connection that lets the agent act in another app or service.
Summon
A subagent. A helper agent the main agent sends off on a side job.
Class, combo, loadout
A class is a kind of task, like 3D modeling for the Sculptor. A combo is a few spells cast in order. A loadout is a full setup for one quest that you can copy.

Rarity and level

Rarity is the curator's opinion: S is Legendary, A Epic, B Rare and C Uncommon. A question mark means nobody has rated the spell yet, which says nothing about its quality.

Level comes from recorded use in the curator's sessions: 1 session reaches level 2, 3 sessions level 3, 7 sessions level 4 and 15 sessions level 5. Stars and download counts never change a level or a rarity.

A gold crown marks a curator's pick: a spell, combo or loadout the curator uses and recommends.

Loadouts and votes

A loadout names a hero, a weapon, a charge and up to twelve spells in four rows: Scout, Craft, Test and Ship. Copying one gives you shell commands that install each spell and start the hero with that model and effort. Read the commands before you run them; they come from each spell's own documentation.

Anyone can post a loadout without an account. Posts can't contain links, a network can post five an hour, and the curator removes spam. A vote is one per loadout per network, and you can take it back. For both, the server stores a keyed hash of your network address, never the address itself. It drops the hash from a posted loadout after seven days and keeps the one on a vote for as long as the vote stands.

How battles work

Battles play themselves, like an auto battler on a phone. Watch with Auto, or step through with Back and Next. The hero is your agent, and its bar is its context window. Foes are the problems a spell deals with, and they hit back: the Fog Wraith's haze makes swings miss the way outdated docs make an agent guess, a Bug Slime splits when a fix breaks something else, and the Backlog Dragon breathes new asks across the party.

Spells answer them. Fresh docs burn away haze, checks raise a ward, and tests pin a slippery foe in place. When context runs low, the hero compacts it and fights on with fewer details. Every step has a plain line that says what really happens.

Loadout pages replay a whole quest in four waves, one per skill-tree row, with the Backlog Dragon at the end. Each equipped spell fires in its row; a row without a spell falls back on basic attacks. The curator wrote every step from each spell's documentation. Battles show how spells work, not a recording of a real session, and their numbers are not test results.

Sound and music

Battles stay silent until you press Sound, and the site remembers that choice on this device. Browsers also hold sound back until you click, tap or press a key on the page, though some, like Chrome, allow it right away after a click on another page of this site. Until then the Sound button shows as off, and your first click, tap or key press starts the sound. Music plays while Auto runs and fades when you pause or scroll away. On the home page, Theme music plays the title theme.

Every sound is CC0, so no credit is required. These are the sources:

  • 512 Sound Effects (8-bit style) by Juhani Junkala (SubspaceAudio), CC0 1.0. Trimmed the silence at the start, shortened a few sounds to 1.5 s or less, faded the ends, normalized and converted to mono MP3. The report sound is a cloth sound from the RPG Sound Pack followed by a coin from this pack.
  • RPG Sound Pack by artisticdude, CC0 1.0. Trimmed the silence at the start, faded the ends, normalized and converted to mono MP3. The heavy swing and the barrier are pitched down and the miss is pitched up.
  • 4 Chiptunes (Adventure) by Juhani Junkala (SubspaceAudio), CC0 1.0. Stage 1 is the battle theme and Boss Fight the boss theme, each one full loop. Added the loop's first half second after its end so it repeats cleanly, set to about -18 LUFS and converted to MP3.
  • JRPG Pack 2 Towns by Juhani Junkala (SubspaceAudio), CC0 1.0. Town1 - Home Town is the title theme, one full loop. Added the loop's first half second after its end so it repeats cleanly, set to about -18 LUFS and converted to MP3.
  • Music Jingles by Kenney, CC0 1.0. jingles_NES03 from the 8-Bit jingles is the victory jingle. Trimmed the silence at the start, set to about -18 LUFS and converted to MP3.

Summons and orchestration

A summon is a subagent: a helper agent the main agent sends off on a side job. It gets its own context window, does its own model and tool work, and hands back a short summary.

In battle, summons appear beside the hero, tied to it by a thin line, each with its own context bar. A foe that hits a summon drains only that summon's bar, the way noisy searches and logs stay in the subagent's context. When a summon returns, it folds into a scroll: its report, which costs the hero a little context and makes its next strike land.

Claude Code hands codebase searches to its built-in Explore subagent on its own, so Claude Code quests open with Explore scouting ahead. Codex spawns subagents when you ask for them, or when AGENTS.md or a skill tells it to. Either way, subagents use more tokens in total, because each one does its own work.

Example: one lead, three reviewers

One lead agent, three subagents and a messy diff.

Foes

  • Tangle Hydra

Party · Context

  • Lead agentYour agent100/100
  1. Tangle Hydra

    The Tangle Hydra blocks the way.

    What really happensIt stands for a messy diff that needs review.

    Lead agent: Context 100/100.

  2. Regrow

    The Tangle Hydra uses Regrow.

    What really happensA messy diff hides new problems inside old ones.

    Lead agent takes 8. Lead agent: Context 92/100.

  3. Subagents

    The lead agent splits the review three ways and summons a specialist for each.

    What really happensYou ask for one subagent per point. Each gets its own context window, and they run in parallel.

    Security: Context 40/40. Test gaps: Context 40/40. Maintainability: Context 40/40.

  4. Regrow

    The Tangle Hydra lashes at Security.

    What really happensThe subagent takes the hit in its own context. The main context stays clean.

    Security takes 8. Security: Context 32/40.

  5. Regrow

    The Tangle Hydra lashes at Maintainability.

    Maintainability takes 8. Maintainability: Context 32/40.

  6. Subagents

    Three specialists strike at once.

    What really happensThe subagents work at the same time and keep their noisy output to themselves.

    Tangle Hydra takes 22. Tangle Hydra takes 22. Tangle Hydra takes 22.

  7. Subagents

    The summons return with short reports.

    What really happensThe lead waits for all three, then gets one short summary from each.

    Lead agent takes 2. Lead agent takes 2. Lead agent takes 2. Lead agent: Context 86/100. Lead agent: Insight. Security returns. Test gaps returns. Maintainability returns.

  8. Longsword

    The lead agent strikes where the reports point.

    Tangle Hydra takes 54 (critical hit). Tangle Hydra falls.

  9. Victory

    The Tangle Hydra falls.

After the battle

  • The summons took 16 damage in their own context. The lead spent 6 on reports.
  • Subagents cost more tokens in total, because each one does its own model and tool work.

Looks

Every spell wears one of twenty item designs in one of twenty palettes, like the recolored monsters of classic RPGs: scrolls, tomes and talismans for skills, medallions and crowns for plugins, orbs, lanterns and keys for MCP servers, and familiar animals for subagents. No two spells share a look, and battle effects use the spell's colors. When you craft a spell, you pick its design and palette.

Bestiary

  • Bug Slime

    Stands for a defect in your code.

  • Flaky Wisp

    Stands for a failing or flaky test.

  • Fog Wraith

    Stands for missing facts, outdated docs or thin context.

  • Tangle Hydra

    Stands for a messy diff that needs review.

  • Blank Mimic

    Stands for an empty screen or canvas that needs a design.

  • Clay Golem

    Stands for a model or part that isn't shaped yet.

  • Reel Phantom

    Stands for footage to make, cut, caption or render.

  • Clockwork Imp

    Stands for a chore you keep doing by hand.

  • Paper Golem

    Stands for documents, slides or sheets to produce.

  • Gate Warden

    Stands for the release gate: deploys, app stores, pull requests.

  • Vault Serpent

    Stands for data, accounts and payments.

  • Error Imp

    Stands for a crash report from a live app.

  • Breach Spider

    Stands for a security hole an attacker could use.

  • Echo Banshee

    Stands for audio to record, clean up, voice or transcribe.

  • Riddle Sphinx

    Stands for a question your data hasn't answered yet.

  • Storm Elemental

    Stands for an outage, a failed deploy or a slow pipeline.

  • Mirage Djinn

    Stands for an AI answer that sounds right but isn't.

  • Fickle Harpy

    Stands for an audience whose attention you have to win.

  • Queue Kraken

    Stands for tickets, invoices and contracts waiting on you.

  • Backlog Dragon

    Stands for the whole quest, every task on the list.

Skill-tree rows

Scout
Gather context.
Craft
Make the thing.
Test
Check the work.
Ship
Deliver it.

Proving Grounds

Battles explain; the Proving Grounds measures. It runs spells and combos on fixed quests with hidden checks, and runs every setup Unarmed too: the same quest, hero, weapon and charge with no spells. Steadiness is reliability. Buff is how far the spells moved the clear rate compared with Unarmed.

The Hall of Champions ranks those results on ladders, which never change a rarity or level.

What counts as use

Local, read-only extractors look for skill reads, skill tool calls, MCP calls and custom commands in Codex, Claude Code and OMP session logs. Subagent work rolls into its root session, and support sessions are left out.

Each spell's log matches connect those records to the library. Generic names may have no match, so zero recorded use does not mean a spell was never used.

Read the records and their definitions

Environment and harness versions

Session logs don't record the operating system, so the curator's OS, architecture and installed harness versions are read when the stats are published. Codex and Claude Code logs record the harness version of every session, and each spell shows the version in its first and latest sessions. OMP logs record none, so its installed version is shown instead.

Spell reports from testers carry their own OS and harness versions unless they leave them out.

Privacy

Published stats are aggregate counts for listed spells and combos from the curator's own sessions. Prompts, commands, tool output, session ids and file paths are never published, and the session logs stay on the curator's Mac.

Browsing needs no account. Spell and combo pages are static. The Tavern, loadouts, votes and the report form use the site's API.

Sources and licenses

Every spell links to its source. Follow that project's install steps and license. Some source-available skills have restrictive licenses and are listed as links only.

The spellbook lists the most popular spells first. Popularity is GitHub stars on the spell's repository and, when the spell installs from npm or PyPI, its downloads last week, both read from public APIs on the date shown. When a repository holds several spells, its stars count for all of them. A spell inside a big product's repository shows no stars, because those count the product. Popularity never changes a rarity.

Dates are snapshots from the day a spell was checked. Galleries hold only media the curator added.

Tavern and reports

The Tavern shows what is trending on GitHub, X and Reddit, refreshed once a day. Rumors there never change a rarity.

Tested a skill, plugin or MCP server? Send a spell report with your results and sources. Made or use one that isn't listed? Craft a spell in your browser and send it to the curator. Accepted reports and spells join the library here, with credit.