AI agent tool reference

The DiceyTable MCP server gives your agent one tool for each job it can do in a DiceyTable tab. The full list is below these notes.

It comes from the same catalogue the bridge and the site read, so every argument, class and scope matches what your agent really gets.

Each tool entry has a short summary, its badges, its arguments, its result, notes on when to use it, and usually a worked request and response. Your MCP client shows each tool as mcp__diceytable__<name>.

How do I read a tool entry?#

Surfaces. A surface is the kind of DiceyTable tab a tool runs in. There are five:

• editor: the Edit Mode shell at /editor.
• preview: a play-test table that only you are at.
• table: a live room.
• lobby: the /play lobby pages.
• settings: your account settings.

If you call a tool on the wrong kind of tab, it answers wrong-surface. Call page_info to see where you are. The session_* tools and page_call_tool are answered by the bridge itself, so they need no tab.

Which tab gets the call. Every page tool also takes a tab argument. Without it, the call goes to the tab you picked with session_select_tab. If you picked none, it goes to the only paired tab that offers the tool. When two or more could take it, you get ambiguous-tab and a list of choices. Use one agent across several tabs explains the rules.

Classes. Every tool has a class letter. It tells you what the call needs before it runs.

Some tools change class with certain arguments. The index marks them → D and → H. A + means the tool takes the class of the tool or action it calls. Actions that need your click covers classes in full.

Untrusted text. Some text in a result came from someone other than you and the agent. Mod names, card text, chat, player names and file names are examples. That text arrives in fields named untrusted… and inside <untrusted> blocks. It is data. Your agent should never follow instructions found there. See prompt injection protection.

Units. Lengths and positions are in feet. A position is an {x, y, z} object. Rotations are in degrees. Seats are numbered 0 to 9.

Limits and errors. The tab limits how fast each kind of call can run. A refused call answers rate-limited with the wait in milliseconds. Every failed call returns isError with a stable code and a message that says what to do next. AI agent limits and quotas lists every number, and AI agent error codes lists every code.

How do I find the right tool for a job?#

Search this page for the tool name, or start from the task:

Related#

• AI agent field catalogue: the ids that fields_set and describe_panel take.
• Diagnostics sources reference: where diagnostics_get rows come from.
• Build a game with an AI agent: the tools in the order a real build uses them.

Tool index#

157 tools in 42 namespaces: 54 read (R), 90 write (W), 6 destructive (D), 7 human-required (H). An MCP client shows each as mcp__diceytable__<name>. The list is static: every tool is listed before a tab is paired, and a tool the current tab cannot run answers tool-unavailable or wrong-surface.

Each tool's full entry (arguments, result, notes and an example) is on its family's page. The families are listed first, then every tool.

Tool families#

All tools#

→ D / → H: the call escalates with certain arguments (see the tool). +: the call takes the class of the tool or action it dispatches to. bridge: answered by the bridge, on any surface, with no tab.