Mcp MCP Server

Craaft MCP Server

Model Context Protocol server for Craaft - lets AI assistants read and write your Kanban boards.

Deployed as an HTTP service. Each request authenticates via the caller's Authorization: Bearer header - either a cra_* personal access token (same ones used by the Craaft REST API) or a cro_* OAuth access token minted by signing in through the connector flow below.

Two-factor sign-in on a craaft account doesn't affect either kind of token. Account security settings (password, two-factor, API keys) are browser-only on the platform, so no tool changes them.

Deploy

cd mcp
gcloud run deploy craaft-mcp \
  --source . \
  --region us-central1 \
  --allow-unauthenticated \
  --set-env-vars CRAAFT_BASE_URL=https://craaft.io,MCP_PUBLIC_URL=https://ai.craaft.io/mcp

The PORT env var is set automatically by Cloud Run.

Connect from Claude (OAuth)

With MCP_PUBLIC_URL set, the server is an OAuth 2.1 protected resource: no token needs to be copied by hand.

craaft is in the Claude directory: connect it from there and skip to step 3. Or add it as a custom connector:

1. In Claude Desktop or claude.ai, go to Settings > Connectors > Add custom connector.
2. Paste https://ai.craaft.io/mcp.
3. Sign in to craaft and approve the connection.

To revoke access later, go to Craaft Settings > Connected apps.

Configure (Claude Desktop, cra_ token)

1. Sign in to Craaft → Settings → API keys → create a token
2. Add to ~/.config/claude/claude_desktop_config.json:

{
  "mcpServers": {
    "craaft": {
      "url": "https://craaft-mcp-xxxxx.a.run.app/mcp",
      "headers": {
        "Authorization": "Bearer cra_your_token_here"
      }
    }
  }
}

MCP registry

craaft is listed in the official MCP registry as io.craaft/mcp, described by server.json. craaft.io proves it owns the io.craaft namespace by serving a public key at https://craaft.io/.well-known/mcp-registry-auth (platform repo, internal/server/mcpregistry.go). The matching private key is not in this repo.

To publish a change, raise version in server.json (a published version can't be changed or republished), then:

PRIVATE_KEY="$(openssl pkey -in ~/.config/craaft/mcp-registry-key.pem -noout -text | grep -A3 'priv:' | tail -n +2 | tr -d ' :\n')"
mcp-publisher login http --domain craaft.io --private-key "$PRIVATE_KEY"
mcp-publisher publish

Install mcp-publisher from the registry's releases.

Tools

The three bulk_* tools are all-or-nothing: one bad item rolls the whole batch back and the error names the offending index, so a failed call never leaves a partially applied board. They send no notification emails (activity events and realtime updates still fire), and each call spends one rate-limit token against a 60-burst budget - which is why they should be preferred over looping the single-card tools.

Endpoints

Env vars

Attachments

This server runs remotely, so it cannot read the caller's filesystem: upload_attachment takes the contents inline (content_text for text, content_base64 for binary), not a path.

download_attachment returns text-ish content types as readable text and everything else base64-encoded. Attachments can be up to 25 MiB, which no conversation wants inline, so it refuses anything over max_bytes (default 256 KiB) and reports the real size instead. Text is truncated rather than refused, since truncated text is still readable; truncated base64 is not, so binary over the limit is skipped entirely.

Attachments are a paid feature. On a Free workspace the upload returns 402 and the error names the current plan. Check canUploadAttachments on the project first - it reflects the board's workspace plan, not the caller's.

Import

preview_import and import_board take the same arguments: filename, the file inline (content_text or content_base64), and optionally source (trello, jira, asana, csv or craaft; detected when left out), name for the new board, and a CSV mapping from craaft fields to headers. Preview first: it writes nothing and, for a CSV, returns the headers, the suggested mapping and sample rows. A mapping replaces the suggestion, so send every field you want. import_board always creates a new private board and counts toward the workspace's board limit (402 when full). Limits: 5 MB, 2,000 cards, 200 columns; attachments aren't imported.

Known platform quirk: archive then delete

archive_column only archives cards in a column flagged isDone, and returns 200 with archived: 0 (and an empty id list) on any other column rather than erroring - check the count. Its returned ids let you undo the sweep with restore_card on each one.

delete_column is blocked by live cards only. Archived cards no longer block it - they are reassigned to the board's first column and the delete proceeds, so they keep existing and stay archived. (This used to fail: the cards_project_column_fk constraint still saw the archived rows, so the delete returned 409 forever. Fixed on the platform side.)

Archive and restore

archive_card takes a card off its board without deleting it; it stays searchable and shows up in list_archived_cards until restore_card puts it back in the exact column and position it was archived from. Archiving an already-archived card, or restoring one that isn't archived, is an error. list_archived_cards uses the same lean board projection as get_board (description omitted), plus archivedAt, capped at 200 rows (newest first) - older archived cards stay reachable through search_cards.

Board templates

list_board_templates is the same fixed catalogue for every caller, no plan gate. Pass a template's key as create_project's template argument to start a board with that column layout. create_project's response never includes columns at all, template or not - that field is only populated by get_project / get_board. Only the default kanban template seeds the historical todo / doing / done column keys anyway, every other template generates its own, so call get_project on the new board before creating cards on it.

Webhooks and email-to-card (board admin, Pro)

Both surfaces are gated the same way: board not visible → not found; visible but you're not a board admin → forbidden; Free workspace → a 402 naming "integrations" as the limit. delete_webhook and disable_inbound_email are not plan-gated, so a workspace that downgraded after setting either up can still turn them off.

A webhook subscription's secret is returned in full on every read (list, create, update) - it is never redacted after creation. It signs craaft format deliveries via the X-Craaft-Signature header; slack and discord deliveries are unsigned chat messages, authenticated only by the receiving service's secret URL. update_webhook's subscription_id argument is the subscription id, not the workspace-level endpoint_id the response also carries.

Email-to-card mints one intake address per board (<token>@<inbound-domain>). Mail becomes a card only when the sender is a workspace member and SPF/DKIM pass; everything else is dropped silently, no error surfaces to the sender. update_inbound_email's rotate flag mints a new token and invalidates the old address immediately - there's no grace period.

Not exposed, and why

GET /projects/{id}/events is the board's SSE realtime stream, not an activity log - a long-lived text/event-stream that doesn't fit a request/response tool. For a board's history use get_board_activity (/projects/{id}/activity), and for one card's history get_card_activity (/cards/{id}/events).

Also absent: POST /projects/{id}/cards/rebalance (internal position maintenance), the board background-image endpoints, and the unauthenticated /public/projects/{token} read - a token-holding agent should just use get_board.