Claude Code and Codex

In a private world, the storyteller can be each player's own Claude Code or Codex, connected to Worldtable over MCP. Worldtable then calls no AI model for that world: every scene, every person you meet and every place that's made is written by your agent, on your own Claude or ChatGPT plan.

Worldtable still keeps the rules, as in any world: health, damage, rewards, money, the map, the clock and who exists. Your agent writes the scene and asks for changes through the world's tools, and the engine decides whether they happen.

The world's owner chooses who writes the world in the Studio, under AI tab, Storyteller, Each player's own agent.

Only private worlds can use it. A player's agent is also the one playing, and it sees the storyteller's notes, so it could go easy on them. That's fine among friends, which is why open worlds always use Worldtable's own storyteller.

Each player connects their own agent, once per world:

• On the Worlds page, open the world's agent page (the robot button on its card).
• Pick Claude Code or Codex, and press Get my setup command.
• Paste the command in your terminal. It connects your agent to this world and installs the storyteller helper. For Claude Code it updates Claude Code first.

Open Claude Code or Codex in any folder and tell it what your character does. It passes your words to the world exactly as you wrote them, then shows you the finished scene.

You can also play on Worldtable in the browser while your agent listens: it waits for your turn and writes the scene, which appears on the page as usual. You make your character in the browser either way, with your agent listening, since it checks your backstory.

The storyteller helper. With the helper on (recommended), each turn goes to a fresh helper on a smaller model (Sonnet 5.5 for Claude Code, GPT-5.6 Luna for Codex), so your chat only ever holds the scenes. With it off, your agent tells the story itself: turns take about as long, but its chat grows by tens of thousands of tokens a turn.

If you own the world or build in it, your connected agent also gets the world's building tools, the same ones Hob uses in the Studio: the premise, regions, peoples, magic, lore, cast, rules, the rule system with its dice and sheets, town and region names, and landmarks. Ask for a change in plain words and it makes it.

Changes save at once and show up in the Studio, and every save can be undone from the Studio's history. See Making a world.

• Turns take a while. About a minute with Claude Code on Sonnet 5.5, 2 minutes with Codex, 5 on Haiku. The world is big, and your agent reads it before it writes.
• It uses your own plan. A turn reads roughly 60,000 to 100,000 tokens, so it uses up your Claude or ChatGPT limits much faster than chatting does.
• Your agent has to be running. If it isn't, a turn waits up to 5 minutes and then fails. Nothing is lost; try again.
• No pictures or voices. Those come from models Worldtable would pay for, so they're off in these worlds.
• No credits. Turns in a world told by players' own agents don't use Worldtable credits.

This is the server the setup command connects. You don't need these details to play; they're here for MCP clients and directories.

• Address: https://worldtable.gg/api/mcp (streamable HTTP, JSON responses).
• Authentication: a bearer token (Authorization: Bearer chr_agent_…) made on the world's agent page. Each token belongs to one player in one world.
• Playing: status, take_turn, last_scene, next_turn, recent_scene, accept_quest and undo_last_turn.
• The helper's tools: work, submit, read_job, search_tools and read_prompt, which carry a turn's storyteller jobs.
• Building (owners and builders): world_outline and the world_* tools.