The MCP server¶

The same knowledge, asked the way a model asks. Tools take names, not numbers. Answers are far smaller than the HTTP ones, and one that cannot be given says what to do instead of failing. See Demonstrations for worked examples.

Connecting¶

Over stdio, which is what a local client spawns, and which asks for no key:

{"mcpServers": {"2009scape-wiki": {"type": "stdio",
  "command": "scape2009-wiki-mcp"}}}

Over HTTP, behind the same token as the wiki contract:

claude mcp add --transport http 2009scape-wiki http://127.0.0.1:8000/mcp/ \
  --header "authorization: Bearer $TOKEN"

Deployment has the command and args for every other way of installing it, and both transports in full.

Tools¶

Eight are written by hand:

tool	what it answers
`get_thing`	One thing by name, and a count of everything it is joined to.
`search`	Find things by words from their name.
`list_things`	Everything of one sort, a page at a time.
`list_sorts`	The sorts this build knows about, and how many of each.
`compare_by_number`	One sort by a number it records, above or below a threshold.
`how_the_price_moved`	Which way one thing’s worth has gone, and by how far.
`find_close_names`	Given a name that answered to nothing, the closest real names.
`about`	Which build is being answered from.

The rest are generated: one tool per relationship per direction, named from the registry’s own labels (drops and dropped_by, sells and sold_in). A relationship declared later becomes two new tools with nobody writing them; one with no edges in this build is not offered. 32 tools on a current build.

Call get_thing first. Its counts name the tool that reads each one.

Answers¶

Every answer is wrapped in an outcome, so a model never reads an exception.

`found`	`result` holds the answer.
`renamed`	That name is retired. The note gives the one to ask with.
`withheld`	In this build, but not published.
`unknown`	Nothing answers to that name. The note says how to settle it.
`ambiguous`	Several sorts answer to it; which was meant is for whoever asked.

Answers are paged, report the total, and declare a result-size ceiling, so a long drop table cannot swallow a context window.

Two rules the server states up front¶

A ref is for calling with, not for saying. Every answer carries a ref such as item:4587, which is how a model names one exact thing back to a tool. A person reading the answer has never seen one, so refs, ids, coordinates and region numbers never belong in what the model says.

A misspelling is not the model’s to settle. find_close_names answers with names and identities alone: it exists to be shown to a person, not chosen from. Same for things sharing a name. If nothing in the answer tells two apart, they are the same thing to whoever asked.

{"outcome": "found",
 "result": {"of": "Dragon scimitar", "label": "Dropped by", "total": 1,
            "neighbours": [{"name": "King Black Dragon", "type": "npc", "id": 50,
                            "facts": {"Chance": "1/512"}}]}}

Edit this page