Batch-Produce On-Brand Designs from a Spreadsheet with Canva

The Problem

Priya runs marketing ops for a franchise with 240 locations. Every seasonal campaign means 240 versions of the same poster: same layout, different city, address, offer price and hero photo. The design team builds one master in Canva, then someone duplicates it 240 times and retypes four fields per copy. It takes three days, and roughly a dozen come out wrong — a stale price, a city misspelled, a logo nudged out of the safe zone.

The workaround everyone reaches for is worse. Exporting the layout to a script that draws text onto a PNG loses the brand: fonts drift, the design team can no longer edit it, and every future change means editing code instead of a design.

The Solution

Keep the master in Canva as a brand template and let an agent fill it. Canva's autofill takes a brand template id and a map of field values and produces a real Canva design — same fonts, same brand kit, still editable by the design team. The agent reads the CSV, uploads each hero image as a Canva asset, fires one autofill job per row, polls the jobs, exports the finished designs, and files them into a folder per region.

Two ways in, same underlying API: the AI Connector MCP (https://mcp.canva.com/mcp) for a human-in-the-loop run, or the Connect REST API for an unattended job in CI.

terminal-skills install canva brand-guidelines social-content csv-parse

Step-by-Step Walkthrough

Step 1: Check the plan before building anything

Autofill and brand kits are Enterprise features; brand templates and resize need Pro. GET /v1/users/me/capabilities answers this in one call, and it is the difference between a working pipeline and a batch of permission_denied responses at row 1. On a Pro account the fallback is create-design-from-brand-template plus manual editing — real, but not unattended.

Step 2: Connect the agent

For the interactive path:

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

The first tool call opens a browser to authorize the account. For the unattended path, register an integration at canva.com/developers and implement the OAuth 2.0 PKCE flow (code_challenge_method=S256), requesting exactly the scopes the job needs: brandtemplate:meta:read, asset:write, design:content:write, folder:write. Canva implies nothing — asset:write does not grant asset:read.

Step 3: Read the template's real field names

The single most common failure is guessing field names: a key the template doesn't have is silently ignored, so the job succeeds and returns a design with empty placeholders.

curl -H "Authorization: Bearer $TOKEN" \
  https://api.canva.com/rest/v1/brand-templates/DAFxyz123/dataset
# → { "headline": {"type":"text"}, "city": {"type":"text"},
#     "price": {"type":"text"}, "hero": {"type":"image"} }

Map CSV columns onto those exact keys once, and validate every row against them before spending a single API call. Use the brand-guidelines skill to settle what actually varies per location versus what must stay locked in the master.

Step 4: Upload the images, then autofill each row

Images enter autofill as asset ids, not URLs, so each hero photo is uploaded first — POST /v1/asset-uploads with the raw bytes and the name in the Asset-Upload-Metadata: {"name_base64": "..."} header (or upload-asset-from-url over MCP, which takes the URL directly).

POST /v1/autofills
{ "brand_template_id": "DAFxyz123",
  "title": "Autumn promo — Leeds",
  "data": { "headline": {"type":"text","text":"30% off through October"},
            "city":     {"type":"text","text":"Leeds"},
            "price":    {"type":"text","text":"£24.99"},
            "hero":     {"type":"image","asset_id":"MSGxyz789"} } }

Step 5: Treat every response as a job, not a result

Autofills, uploads and exports are all async. The POST returns {"job": {"id", "status": "in_progress"}}; the design only exists once a GET on /v1/autofills/{jobId} returns success, with the id at job.result.design.id. Poll on a back-off (1s → 2s → 4s), never in a tight loop.

Rate limits are per user per minute and they are the real pacing constraint: autofill 60/min, creates and exports 20/min, reads 100/min. 240 rows is a throttled queue that finishes in about twenty minutes — not a Promise.all that trips 429 on row 30.

Step 6: Export, file, and hand back editable links

POST /v1/exports with format: {"type":"png"} (or pdf with size: "a4" for print), poll /v1/exports/{exportId}, then download immediately — export URLs die after 24 hours. POST /v1/folders/move files each design into a folder per region so the local teams can find theirs.

What Priya sends the franchisees is the design's edit_url (valid 30 days), not the export link: a location that needs a tweak opens it in Canva and edits it, because it is a real Canva design and not a flattened image. Pair the output with the social-content skill for the caption and posting cadence that go with each variant.

Real-World Example

A 240-location franchise ran its autumn campaign this way. The design team built one master template and signed it off once. The agent validated all 240 CSV rows against the template dataset up front and caught 9 bad rows before spending an API call — two missing prices, seven cities with no hero image — and reported them instead of generating half-empty posters.

The run took 23 minutes wall-clock, paced by the 60/min autofill limit, and produced 231 designs exported as print-ready A4 PDFs plus 1080×1080 PNGs, filed into 12 regional folders. Three locations later asked for a different photo; they opened their edit_url and changed it themselves in Canva. Against three days of manual duplication, the measurable win was less the time than the error rate: zero stale prices, because no human retyped one.

Related Skills

• canva — the integration this workflow runs on: MCP connector, Connect API OAuth, autofill, exports, folders, and the tier gates and rate limits that shape the batch
• brand-guidelines — decide what varies per variant and what stays locked in the master template before you automate the duplication
• social-content — captions, formats and posting cadence for the variants once the designs exist
• csv-parse — validate and stream the source rows, so a malformed row is caught before it becomes an empty poster