CLAUDE LABJP
2.1.274 — This round is aimed at people running Claude Code unattended: a warning when memory runs critical, an environment variable that bounds how long the first turn waits for MCP servers, and an end to sessions retrying a 400 forever10/07 — The old management-settings spelling is accepted until noon PT on October 7, nineteen days from now. Deprecation warnings have been showing since September 10WINUPD — Reports keep coming in of a Windows update leaving Cowork unable to mount a single host folder. Removing the update is still the only workaround anyone has foundNEW — Stopping before you hit the limit: a record of rebuilding the day around the five-hour windowBING — Seven in ten of the people who actually read these pages arrive from Bing. Search has more than one front doorEXCEL — Before handing over a spreadsheet, decide which columns it may read and which it may not2.1.274 — This round is aimed at people running Claude Code unattended: a warning when memory runs critical, an environment variable that bounds how long the first turn waits for MCP servers, and an end to sessions retrying a 400 forever10/07 — The old management-settings spelling is accepted until noon PT on October 7, nineteen days from now. Deprecation warnings have been showing since September 10WINUPD — Reports keep coming in of a Windows update leaving Cowork unable to mount a single host folder. Removing the update is still the only workaround anyone has foundNEW — Stopping before you hit the limit: a record of rebuilding the day around the five-hour windowBING — Seven in ten of the people who actually read these pages arrive from Bing. Search has more than one front doorEXCEL — Before handing over a spreadsheet, decide which columns it may read and which it may not
Articles/Cowork
Cowork/2026-03-09Beginner

Canva × Claude in Practice — Saving Generated Candidates, Picking a design_type, Committing Edits

The parts of the Canva AI connector that actually trip people up in Claude: why generated designs are not saved yet, how to pick a design_type, why editing is a transaction, and what to check before exporting.

Canva4Claude48MCP53ConnectorDesign6Cowork41

The Design in Your Chat Is Not Yours Yet

You ask Claude to make a Canva design, it appears in the conversation, and days later you cannot find it anywhere in your Canva account. My first assumption was that I had bookmarked the wrong thing.

I had not. What design generation returns is a set of candidates, not a saved design. Only when you pick one and take the extra step of saving it does the design land in your account.

The URLs shown alongside those candidates are the second trap. The IDs inside them are not design IDs. That is why a follow-up like "now export that one as a PDF" sometimes fails to find anything.

This connector does a lot, and the marketing copy makes it sound like anything goes once you are in chat. In practice, the gap between a request that works and one that quietly does not comes down to knowing the plumbing underneath. What follows is that plumbing, ordered roughly by where people get stuck.

Setup, and Clearing Up "When Did This Launch?"

Connecting

In Claude.ai, click the connector icon in the message box, choose Canva, sign in, and grant access. Cowork works the same way from its settings screen. Canva documents the steps on its AI connector setup page.

Connecting takes under a minute. The awkward part is what happens when the connection lapses. Once authorization expires, the tools still appear in the list but every call fails. Deciding whether to wait or hand the run back to a person when that happens mid-automation is something I worked through in Not Every Unreachable Connector Means the Same Thing in an Unattended Run.

About the launch date

You will see claims that Canva shipped its Claude connector in January 2026. That is not quite right. The January announcement covered an expansion — on-brand generation using your Brand Kit — and the Canva × Claude integration itself predates it. Canva's own newsroom says the new capabilities build on the existing Claude integration (Create on-brand Canva designs directly inside Claude).

A small distinction, but "brand new feature from last month" and "extension of something already in production" are very different pitches when you are proposing this to a team.

Your Choice of design_type Changes What You Get

Design generation takes a design_type. It has more influence on the result than anything else you say, and the differences are hard to see from outside.

The document-style types are the easiest ones to get wrong.

ValueWhat you getBest for
docA Canva Doc — web-first, flexible layoutMemos, articles, requirements docs, plans, internal announcements
documentA traditional fixed-page templatePrint-oriented forms and paperwork
proposalA visually designed proposal templateProposals that lead with graphics
reportA report template with chartsReports built around numbers

For most business writing, doc is the right answer. Pick document and your text gets poured into a fixed layout, which falls apart the moment the length differs from what the template expected.

Presentations take a different path

presentation is not accepted by ordinary design generation. Presentations route through an outline step first, and only after that outline is settled does structured generation run. So when you ask for a ten-slide deck and get asked to review an outline instead, that is the system working as designed.

When you are in a hurry, the extra step feels like friction. But fixing the structure at the outline stage removes the slide-by-slide cleanup afterwards, and I have come around to letting it stop me there.

When you want your own text, unchanged

If you already have a draft and want it placed as written, use the verbatim mode. It comes with limits:

  • Verbatim only applies to doc. Every other type falls back to normal AI generation.
  • Supported formatting stops at headings (# / ##), paragraphs, bold, italics, and bulleted or numbered lists.
  • Tables, links, code blocks, and images do not come through.

Try to paste a technical article and that last line is where you stop. If your draft contains tables and code, plan to rasterize them beforehand or rebuild the piece in a different type.

Editing Is Read, Touch, Then Commit

Editing a design is not a single round trip. Miss this and you end up staring at a design that stubbornly refuses to show your change.

  1. Read the design with a transaction open. You get the current page state, a locator_id for each element, and a transaction_id.
  2. Send your edit operations using that transaction_id. At this point nothing is saved yet.
  3. Call once more with no operations to commit or cancel.

The key constraint: you cannot mix operations and finalizing in the same call. Edit, look at the result, then commit — that order is enforced, not advisory. And a commit cannot be undone.

Elements are addressed by locator_id rather than coordinates. Roughly 28 operation types are available, covering text replacement, find-and-replace, swapping fills, moving and resizing, formatting, inserting shapes, adding and reordering pages, and replacing speaker notes.

⚠️
**Responsive pages behave differently.** On a responsive page, only a restricted set of operations is allowed and no thumbnail comes back. If the same request works on one design but silently does nothing on another, check the page type first.

If you insert a shape from your own SVG path, only the M / L / H / V / C / S / A / Z commands are supported. Q and T are not. Quadratic curves need rewriting as C or S, and arcs as A. When an icon you already have refuses to go in, this is usually why.

Formatting has ceilings too: font size 1–800, line height 0.5–2.5, new page dimensions between 40 and 8000 pixels, and speaker notes capped at 5,000 characters.

Where Exports and Search Stall

Ask which formats are available first

An export will not go through unless you check the formats that specific design supports beforehand, because support varies by design type. Asking for PPTX on a design that cannot produce it simply fails.

FormatMain options
PDFPaper size (a4 / a3 / letter / legal), quality, page selection
PNGWidth and height, transparent background, lossless compression, merge pages into one image
JPGWidth and height, quality (1–100)
MP4Video quality (for example horizontal_1080p)
PPTX / GIF / CSVPage selection, quality

Image width and height accept 40 to 25000 pixels. Worth remembering when you are planning something for print.

Why "find me a template" comes back empty

Design search and template search are separate doors. If you ask for a monthly report template and get a pile of your own existing designs, design search was the one that ran. When you want a template, say plainly that you are looking for a brand template.

Design search has its own quirks. Filtering by keyword requires sorting by relevance, and filtering by kind currently only supports presentations.

If you intend to autofill, check the template's data structure first. An empty result means that template does not support autofill at all — rather than pushing on it, open Canva's create URL for that template and build it by hand.

How I Actually Phrase Requests

Lead with what is already decided

Vague requests produce vague candidates.

Before: "Make me a nice poster."

After: "An A3 portrait poster. Blue gradient background, white text reading 'AI Summit 2026'. Date March 15, venue Tokyo Big Sight. Leave a framed space at the bottom for a QR code."

There is a second, less obvious property here: the generation tool does not remember your earlier turns. When you ask for another pass, you have to restate every condition you gave before. That is why "same thing, just change the color" so rarely lands.

Numbers worth memorizing

  • Up to 10 assets can be inserted at generation time, placed in the order you provide them
  • Comments are capped at 1,000 characters
  • Resize presets cover presentations and whiteboards only — Canva Docs and emails are excluded
  • Design IDs are 11 characters starting with D; brand template IDs start with BTM

That last line looks trivial and is not. Once you know the shape of an ID, you can spot a mix-up mid-conversation without waiting for the failure.

Brand kits remove the cleanup, not the work

Generate with a Brand Kit applied and the colors, fonts, and logo follow your guidelines from the first draft. Canva cited exactly this in explaining the expansion: teams were spending the time AI saved them on pulling outputs back on brand.

That matches my own experience. Working as an indie developer, when I put together promotional images for an App Store listing, having the colors and typefaces already in place removes one whole review pass before publishing. Speed of output matters less, day to day, than not having to fix what comes out. Note that some capabilities, including managing multiple brand kits, assume a Canva Pro plan.

If You Are Trying This Today

Do not start with a document that matters. This connector is much easier to read when the stakes are zero.

Generate a single doc on a throwaway topic, save it properly, then run one full edit-and-commit cycle against it. The difference between a candidate and a saved design, what locator_id looks like in practice, and the fact that committing is its own call — once those three feel familiar, moving to real work is straightforward.

Thank you for reading through a piece with this much fine print in it.

Share

Thank You for Reading

Claude Lab is ad-free, supported entirely by members like you. We publish practical guides daily with implementation code, benchmarks, and production-ready patterns. If you've found it useful, we'd love to have you on board.

  • Copy-paste ready implementation code
  • New advanced guides published daily
  • $5/mo or $15 for lifetime access
View Membership →

If you found this article helpful, a small tip ($1.50) would mean a lot to us. Your support helps keep this site ad-free and covers server and hosting costs.

Related Articles

Cowork2026-03-09
Cowork Connectors Guide — Maximize AI Power with External Service Integration
Learn how to use Cowork's connector features. From integrating with Canva, Figma, Slack, and Asana to interactive tool integration via MCP Apps, discover practical ways to connect external services.
Cowork2026-07-11
Make Your Nightly MCP Connectors' Health Visible — A Lightweight Ledger for Solo Operators
You don't need Enterprise connector observability to see your MCP connectors' error rate and latency. Append one line per tool call, roll it up weekly, and let regressions ring a bell. A working health ledger for anyone running scheduled tasks solo.
Cowork2026-06-04
Where the human still has to press the button — lessons from delegating ad ops to Cowork
From actually delegating repetitive ad operations — AdMob bidding-partner sign-ups and floor tuning — to Cowork, here is the boundary I found between what AI can fully own and the single moves a human still has to make by hand.
📚RECOMMENDED BOOKS
Build a Large Language Model (From Scratch)
Sebastian Raschka
LLM Dev
Prompt Engineering for LLMs
Berryman & Ziegler
Prompting
AI Engineering
Chip Huyen
AI Eng
* Contains affiliate links