Bring documented MCP tool names in line with the live servers

[mattpodwysocki]

Split out of #80, then extended: once the first wrong tool name turned up I audited every *_tool in the repo against tools/list on all three servers in .mcp.json — runtime (31 tools), devkit (22), docs (3).

Four documented names exist on none of them, and two more are documented in a way the server now contradicts. An agent following these docs calls tools that aren't there.

Runtime — mapbox-mcp-runtime-patterns

name	status	fix
point_in_polygon_tool	gone	→ points_within_polygon_tool (37 occurrences, 3 skills)
version_tool	gone	removed from the table, utility-ops list, and examples/README.md
category_list_tool	deprecated	its own description says use resource_reader_tool with mapbox://categories
render_map_tool	undocumented	added — see below

render_map_tool's description opens "PREFERRED MAP TOOL: … ALWAYS use this tool to display maps", and the skill never mentioned it. Added with the distinction from static_map_image_tool: render_map_tool draws an interactive GL JS map and consumes the mapboxRender.ref other tools return in structuredContent; static_map_image_tool is for a static PNG or JPEG.

Also undocumented and now listed: ground_location_tool, place_details_tool, and nine local tools (convex, destination, difference, intersect, length, nearest_point, nearest_point_on_line, points_within_polygon, union). The table claimed 9 local tools; there are 17.

DevKit — mapbox-mcp-devkit-patterns

• create_token_tool does not exist. The skill told readers to run it as step 1 of token management, and an eval expectation required recommending it. Token creation isn't an MCP operation — it happens in the account dashboard or via the Tokens API. list_tokens_tool, which does exist, inspects what's already there.
• get_latest_mapbox_docs_tool does not exist on DevKit. Documentation lookup is a separate server, mcp-docs.mapbox.com, offering search_mapbox_docs_tool, get_document_tool and batch_get_documents_tool. Corrected, and the endpoint named — the skill pointed every MCP example at the DevKit URL.

AGENTS.md also disagreed with its own references

• LangChain used DynamicTool; references/langchain.md uses DynamicStructuredTool with a Zod schema, which is what stops the model passing malformed coordinates.
• Pydantic AI passed a bare model string and gave directions_tool origin/destination parameters. It takes coordinates, a list of {longitude, latitude} objects, and the import is OpenAIChatModel — OpenAIModel does not exist and raises ImportError. The skill's own eval expectations already forbade the form its AGENTS.md demonstrated.
• Two isochrone_tool calls passed coordinates positionally, hiding that it takes a single {longitude, latitude} object.

Measured

--repeats=3: devkit 96.3% skill / 98.1% agents; runtime 93.9% skill. Agents surface for runtime moved 82% → 99% in earlier measurement.

The runtime figure sits below an earlier 98.3% reading on the same content. That eval's runs were 100/67/100, so the gap is its own spread rather than this change — which is the kind of thing the --repeats work in #87 exists to tell apart.

Remaining

Only call_tool and my_tool are unmatched, and both are method names in example code rather than tool references.

Ordering

Touches eval files that #87 relocates, so #87 should land first.

Test plan

• npm run check passes
• Every *_tool name re-audited against live tools/list on all three servers — nothing unmatched but the two method names
• mcp-devkit-patterns and mcp-runtime-patterns re-run at --repeats=3

🤖 Generated with Claude Code