MCP server reference

Krafa serves a board's taste profile over a hosted MCP server. It's stateless: every request authenticates independently and is scoped to the single board its token belongs to.

Connection

Tokens are minted per board from ⌘K → Connect AI Agent and look like krafa_ followed by a random suffix. They're read-only: nothing exposed over MCP can modify a board.

Self-hosted instances serve the same endpoint at {your-origin}/api/mcp.

get_taste_profile

The designer's full curated design-taste profile for this board, as markdown. Read this first and follow it when designing anything for them.

Parameters: none.

Returns: one text block containing the profile as Markdown. Active memories only; anything superseded by a conflict resolution is excluded.

The document is titled with the board name, followed by a count, then one section per category in a fixed order: Typography, Color, Spacing & Density, Layout, Imagery, Motion, Voice & Tone, Components. Empty categories are omitted. Each memory is a bullet with its evidence indented beneath.

# Design Taste Profile — Product Surfaces

14 memories across 6 categories.

## Typography

- Numeric data is set in tabular figures at a heavier weight than surrounding copy
  - font-variant-numeric: tabular-nums
  - weight: 500 against 400 body

## Color

- Accents are a single saturated warm hue against near-neutral surfaces
  - accent: #ff5c00
  - surface: #0a0a0a

A board with nothing learned yet returns _No memories yet — the board hasn't been curated._

list_memories

Structured list of the designer's taste memories (statement + concrete evidence values), optionally filtered by category.

Parameters:

Returns: a JSON array of active memories.

[
  {
    "id": "1f4a...",
    "category": "color",
    "statement": "Accents are a single saturated warm hue against near-neutral surfaces",
    "evidence": ["accent: #ff5c00", "surface: #0a0a0a"]
  }
]

Use this over get_taste_profile when the agent needs to reason about individual memories or cite one by id. For "just tell me how to design this," the Markdown profile is better: it's shorter and already organized.

get_exemplars

Up to five of the actual reference images the designer curated their taste from. Look at them to absorb the visual style directly.

Parameters: none.

Returns: a text header followed by up to five base64 image blocks, in board order.

Exemplars aren't a separate collection you manage. They're computed at read time: Krafa takes the nodes that produced the board's active memories, keeps the images among them, and returns the first five. Images over 3 MB and any that fail to load are skipped silently.

Boards with no image-backed memories return No image exemplars on this board yet.

Usage and billing

Every tool call counts as one generation against your plan, whichever of the three it is. Connecting and calling tools/list cost nothing, so discovery is free and an idle client never spends.

Plans include 2,000 generations a period on Pro and 3,000 per seat on Teams, pooled across the workspace. Past that Krafa spends prepaid credits, then falls to overage at $0.02 each, and only refuses once the spend cap is reached. See Plans and billing for the full order.

A refusal is a tool error, not a transport error. The HTTP request still succeeds and the response carries isError with a plain-sentence explanation, which is what lets the agent relay the reason instead of failing opaquely:

{
  "isError": true,
  "content": [
    {
      "type": "text",
      "text": "This workspace has reached its monthly spend cap, so no more generations can be billed."
    }
  ]
}

Authentication failures are the exception and come back as 401 before any tool runs, as above.

Calling it directly

Useful for debugging a connection outside a client:

curl -sS https://app.krafa.ai/api/mcp \
  -H "Authorization: Bearer krafa_YOUR_TOKEN" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Streamable HTTP requires both media types in Accept; omitting text/event-stream is the most common reason a hand-rolled request fails against a client that works fine.

Swap the body for a tool call:

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/call",
  "params": { "name": "list_memories", "arguments": { "category": "typography" } }
}