MCP server

DimensionKit runs a remote MCP server, so an AI assistant can build, edit, publish and review your scorecards on your behalf.

https://app.dimensionkit.com/mcp

Paste that into any client that supports remote MCP servers. See Connecting a client to set one up, and Managing connections to see what is connected or cut it off.

What it can do

• Create scorecards, add dimensions and questions, set tier bands, publish
• Edit or remove questions and dimensions on a scorecard that is already live
• Read submissions, including per-dimension scores and the answers people gave
• Check whether a scorecard is ready to publish, and what is weak about it
• Hand you a preview link for a draft before anything goes live

How a scorecard gets built

The tools are shaped around the order the work actually happens in:

1. scorecard_design_guide — the house style, read before writing anything
2. create_scorecard — a draft, with presentation and branding settings
3. add_dimension — the things being measured
4. add_questions — every question for that dimension, in one call
5. set_dimension_tiers — the bands a score falls into, and what each one means
6. check_scorecard — confirms it will publish, and flags what is weak
7. publish_scorecard — makes it public

A draft is not reachable publicly. Tool results include the URL a scorecard will have, flagged urlIsLive: false until you publish, so an assistant knows not to share it yet.

Guidance, not just schemas

scorecard_design_guide carries the things a tool schema cannot say: what to ask you before starting, how to phrase a question so every answer is a plausible thing to admit, how to write a tier insight worth reading, and which of the four visual themes suits which audience. An assistant that reads it first will ask you about your audience and the look you want rather than guessing at both.

Clients that support MCP prompts also get three slash commands — build_scorecard, review_scorecard and choose_theme — which are the interview scripts for those conversations.

Seeing a draft before you publish

Every tool result for a draft includes a previewUrl: a signed link, good for 24 hours, that renders the scorecard exactly as a respondent would see it. It carries a banner saying it is unpublished, and answers submitted through it are not recorded — submissions stay gated on publication. Ask your assistant for the preview link rather than taking its word for what it built.

Two rules the tools enforce

Tier ranges are derived, never authored. A dimension scores between the sum of its lowest answer scores and the sum of its highest, and its tiers tile that span exactly. Adding, editing or removing a question moves the span, so the ranges are recalculated for you. You supply tier labels and insights; the numbers are not yours to set.

Question ids and dimension keys are permanent. Submitted answers are filed under the bare question id, and scores under the dimension key. Renaming either would strand data already recorded, so update_question and update_dimension change everything except those. Remove and re-add if one genuinely has to change.

Editing something that is already live

Safe. Every submission stores its own scores, tier, and the exact wording shown at the time, stamped with a hash of the config that produced it. Editing a scorecard cannot reach back and change results people have already been given.

When a scorecard has submissions, tool results say how many, so an assistant can tell you what is affected before it changes anything.

What it deliberately does not do

• No asking you questions directly. MCP’s elicitation and sampling both need the server to send a request back down an open connection, and this server does not hold one. Anything that needs your input happens in your own conversation with the assistant instead.
• No streaming. The endpoint answers one request at a time and does not hold SSE connections open. Clients that require the older HTTP+SSE transport will not connect.
• No account or billing access. The scopes cover scorecards and their submissions, nothing else.
• No public write API. The REST surface stays read-only plus form submissions; everything that writes goes through MCP or the admin.