Canva

Canva — Design Automation via MCP and the Connect API

Overview

Canva exposes two automation surfaces, and picking the wrong one wastes an afternoon:

Use the MCP connector with a human in the loop; use the Connect API for unattended runs, hundred-row fan-outs, or acting for users other than the developer. (A third server, the Dev MCP at canva.dev, only serves docs for building Canva apps — it does not touch designs.)

Instructions

Setup A — AI Connector MCP

claude mcp add --transport http canva https://mcp.canva.com/mcp

Clients without HTTP transport go through the bridge: npx -y mcp-remote@latest https://mcp.canva.com/mcp. The first call opens a browser to authorize the Canva account; the client stores the token. Verify with tools/list.

Tools (rate limit per user per minute in parentheses):

Editing is transactional. start-editing-transaction → perform-editing-operations (repeatable) → commit-editing-transaction. Nothing lands in the design until the commit; cancel explicitly rather than dropping an open transaction.

Setup B — Connect API (OAuth 2.0 + PKCE)

Create an integration at canva.com/developers, set the redirect URI and scopes. The flow is PKCE-mandatory:

# 1. verifier (43–128 chars) + S256 challenge
VERIFIER=$(openssl rand -base64 96 | tr -d '\n=+/' | cut -c1-96)
CHALLENGE=$(printf %s "$VERIFIER" | openssl dgst -binary -sha256 | basenc --base64url | tr -d '=')

# 2. send the user here (scopes space-separated, state = CSRF nonce)
https://www.canva.com/api/oauth/authorize?response_type=code&client_id=$CLIENT_ID\
&code_challenge=$CHALLENGE&code_challenge_method=S256&state=$NONCE\
&scope=design:content:write%20design:meta:read%20asset:write

# 3. exchange the code — Basic auth is base64("client_id:client_secret")
curl -X POST https://api.canva.com/rest/v1/oauth/token \
  -u "$CLIENT_ID:$CLIENT_SECRET" \
  -d grant_type=authorization_code -d code="$CODE" -d code_verifier="$VERIFIER"

Access tokens last 4 hours; refresh with grant_type=refresh_token. Each refresh token is single-use — the response returns a new one, and losing it means re-consent, so persist the rotation atomically before using the new access token.

Scopes — nothing is implied

asset:read · asset:write · design:meta:read · design:content:read · design:content:write · brandtemplate:meta:read · brandtemplate:content:read · brandtemplate:content:write · folder:read · folder:write · folder:permission:write · comment:read · comment:write · collaboration:event · profile:read · openid · profile · email

asset:write does not grant asset:read. Request each one you call.

The async job pattern — exports, autofills, uploads, imports, resizes, merges

Every heavy Connect operation is a job: the POST returns {"job": {"id", "status": "in_progress"|"success"|"failed"}}, and the result exists only once a GET on the job returns success. Never treat the POST response as the result.

Poll on a back-off (1s → 2s → 4s, cap ~5s). Tight loops burn the 100 req/min read budget long before the job finishes.

Endpoint map (Connect API)

• Designs — POST /v1/designs (design_type preset doc|email|presentation|whiteboard or custom 40–8000 px, max 25 MP area; or from design_id / brand_template_id), GET /v1/designs (query, ownership=any|owned|shared, sort_by, limit 1–100, continuation), GET /v1/designs/{id}, /pages, /export-formats, /dataset
• Assets — POST /v1/asset-uploads with Content-Type: application/octet-stream, the raw bytes as the body, and the name in a header: Asset-Upload-Metadata: {"name_base64":"<base64 of a ≤50-char name>"}. GET/PATCH/DELETE /v1/assets/{id}
• Autofill — POST /v1/autofills with brand_template_id + data keyed by the template's field names
• Brand templates — GET /v1/brand-templates, /{id}, /{id}/dataset (the dataset is the field list autofill expects — read it, never guess field names)
• Exports — POST /v1/exports, formats pdf, png, jpg (quality 1–100 required), gif, pptx, mp4 (quality required), csv, html_bundle, html_standalone
• Folders — POST /v1/folders, GET /v1/folders/{id}/items, POST /v1/folders/move
• Comments — threads and replies on a design; Analytics — views, viewers, links
• User — GET /v1/users/me, /me/profile, /me/capabilities (check tier here before calling a gated endpoint)

Paginate with continuation: pass the token from the previous response until it is absent.

Examples

Example 1: 240 localized social posts from one brand template

"Fill our 'Q3 Promo' brand template with every row in offers.csv and export PNGs."

# 1. Read the template's real field names — never guess them
curl -H "Authorization: Bearer $TOKEN" \
  https://api.canva.com/rest/v1/brand-templates/DAFxyz123/dataset
# → { "dataset": { "headline": {"type":"text"}, "price": {"type":"text"},
#                  "hero": {"type":"image"} } }

# 2. Product shot → asset (poll the job, capture asset.id)
curl -X POST https://api.canva.com/rest/v1/asset-uploads \
  -H "Authorization: Bearer $TOKEN" \
  -H 'Asset-Upload-Metadata: {"name_base64":"aGVyby1hdXR1bW4ucG5n"}' \
  -H "Content-Type: application/octet-stream" --data-binary @hero-autumn.png

# 3. One autofill job per row (60/min — throttle the loop)
curl -X POST https://api.canva.com/rest/v1/autofills \
  -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
  -d '{"brand_template_id":"DAFxyz123","title":"Autumn promo — Berlin",
       "data":{"headline":{"type":"text","text":"30% off through October"},
               "price":{"type":"text","text":"€24.99"},
               "hero":{"type":"image","asset_id":"MSGxyz789"}}}'
# → {"job":{"id":"af_01H...","status":"in_progress"}}

# 4. Poll → success gives job.result.design.id → POST /v1/exports → poll → job.urls[]

Scopes needed: brandtemplate:meta:read, asset:write, design:content:write. Autofill is Enterprise-only — on a lower tier this returns permission_denied, and the fallback is create-design-from-brand-template (Pro) followed by manual editing.

Example 2: Same job, one prompt, over MCP

"Find our Q3 Promo brand template, make a version for the Berlin offer with this product shot, and export it as a PNG."

search-brand-templates("Q3 Promo")        → template id
get-brand-template-dataset(id)            → headline / price / hero  (Enterprise)
upload-asset-from-url("https://cdn.acme.dev/hero-autumn.png")  → asset id
autofill-design(id, { headline, price, hero })                 → design id + edit_url
export-design(design_id, format: "png")   → download URL (valid 24 h)

Give the user the edit_url — it lasts 30 days and opens in Canva, while the export URL dies in 24 hours. If autofill-design is missing from tools/list, the account is not Enterprise: say so instead of retrying.

Guidelines

• Every URL expires, at a different rate. Thumbnails ~15 min, export downloads 24 h, edit_url / view_url 30 days. Download export bytes at once; hand humans the edit_url, never a signed export link.
• Read the dataset before autofilling. GET /v1/brand-templates/{id}/dataset (or get-brand-template-dataset) returns the exact field names and types. A guessed key is silently ignored, producing a design with empty placeholders rather than an error.
• Tier gates are hard. Resize and brand templates need Pro; autofill and brand kits need Enterprise. Check GET /v1/users/me/capabilities before a batch and read permission_denied as "wrong plan", not "wrong request".
• Limits are per user, per minute: creates and exports 20, autofill 60, reads 100. A 240-row batch is a throttled queue, not a Promise.all. Back off on 429.
• Jobs are not results. A 200 on POST /v1/exports means accepted. Poll until success, and handle failed by reading job.error.code rather than re-submitting.
• Blank designs vanish. A design created via the API and never edited is deleted after 7 days. Autofill or edit it in the same run if it needs to persist.
• Refresh tokens are single-use. Store the rotated token before the next call; a lost rotation means the user must re-consent through the full PKCE flow.
• PKCE is mandatory; the code_verifier and client_secret stay server-side. A Canva integration cannot be driven safely from client-side code.
• Design ids (DAF…), asset ids (MSG…) and job ids are scoped to the authorized account; a 404 on a valid-looking id usually means it belongs to a different team.