Skip to main content
The Moda connector is the path for hosts where your agent cannot run a shell — claude.ai in the browser, Claude Desktop, and other chat-first surfaces. It is a hosted MCP server: you connect once, sign in, and your assistant can create, read, edit, and export real Moda canvases. On claude.ai there is nothing to paste. Moda is published in the Claude connector directory, so you find it and press Connect. The connector URL below is still what other hosts need, and what claude.ai needs in the one case the directory does not cover — see Adding it by URL. If your agent does have a terminal, use the CLI instead. It is cheaper in context, faster, and it can see the files on your machine. The connector exists so nobody is stuck without Moda, not as the recommended default.
The connector and the CLI call the same Moda service, so plan entitlements, credits, brand kits, and canvases are identical across both. Nothing is gated on which transport you picked.

​
Connector URL

​
Authentication

The connector URL supports two authentication methods. Pick one per caller — the server accepts both concurrently, so one teammate can be signed in over OAuth while a scheduled job uses a key:
  • OAuth sign-in — the default, and what the setup steps below use. The host holds the tokens, the model never sees a credential, and there is no plan requirement.
  • A Moda API key as a bearer token — send Authorization: Bearer moda_live_... against the same connector URL above.
A host whose custom-connector flow asks for an API key rather than a sign-in does not need a different endpoint. Meta’s Muse is one: it asks for a key even with a person in the chat. Point it at the connector URL and give it a key.

​
Using an API key

Mint the key

In Moda, go to Settings → Developer → REST API → Create API key. Any plan can create one. The key is shown once — copy it into a secret manager.

Keep the default scopes

The connector enforces scopes per tool on key auth, so a narrowed key returns insufficient_scope on individual tools — export, media, drive and site calls fail first — with nothing to indicate the key rather than the endpoint is at fault. OAuth grants skip that check. Accept the picker’s default, which is every scope except Organization admin. Scopes beyond the connector’s allow-list are stripped on connector calls anyway but stay live for the REST API and CLI on the same key, so adding admin gains nothing here and widens it everywhere else. Don’t narrow it further unless you want a deliberately restricted key.

Point the host at the connector URL

Use the connector URL from the top of this page and send the key as Authorization: Bearer moda_live_.... Nothing else changes — the same tool surface, the same workspace, the same credits.
Key calls run as the key’s owner and their team, where an OAuth grant runs as the individual who signed in. Revoke or rotate from the same settings page; revoking is immediate and kills the key for the REST API and CLI too, since it is one credential across all three.

​
Set it up on claude.ai

Open the Moda listing

Go to Moda in the Claude connector directory, or find it under Customize → Connectors by searching for Moda.

Connect and sign in

Click Connect. Moda opens an OAuth sign-in in your browser; approve the scopes. Tokens are held by Claude — the model never sees a credential.

Choose a workspace

If your Moda account belongs to more than one workspace, Moda asks which one this connection should act in. One workspace and the step is skipped. Your agent can still act in another of your teams on an individual call; changing the default means disconnecting and reconnecting.

Turn it on in the chat

In a conversation, click + at the lower left (or type /), hover Connectors, and toggle Moda on.
Claude Desktop reads the same directory: Customize → Connectors, find Moda, Connect.

​
Adding it by URL

Use this when you do not see Moda in your connector directory. It reaches exactly the same server; it is only a longer way in.
On a Team or Enterprise workspace this is not a workaround. Both routes need an Owner to approve Moda for the organization before any member can connect, so adding the URL from your personal settings will not complete either. See Enterprise managed auth.

Open connector settings

Go to Customize → Connectors, click +, and choose Add custom connector.

Add Moda

Name it Moda, paste the connector URL above, and click Add. Then connect and sign in as above.

​
Plan and workspace notes

Every member holds their own grant, so each one only ever sees the canvases, orgs, and credits their own Moda account has access to.

​
Set it up on ChatGPT

ChatGPT reaches the same connector once its developer switch is on:

Turn on developer mode

Settings → Security and login → Developer mode. Once per account.

Create the connector

Open ChatGPT Plugins, click +, and create a developer-mode app for a remote MCP server using the connector URL above.
Two things to expect there. Write actions ask for confirmation on each call by default — that is ChatGPT’s behaviour, not Moda’s, and you can tell it to remember your choice for the conversation. And developer-mode availability differs by plan, so check OpenAI’s own documentation if the switch is missing. Moda is not published as a ChatGPT deep-research connector — that needs a different search/fetch tool shape, which this surface does not expose. Developer mode is the supported route.

​
First calls

The connector expects two calls before real work:
  1. moda_bootstrap — the handshake. Returns your identity and teams, plan tier, scopes, entitlements, and the current media model matrix in one payload, so the assistant can skip a pile of lookups. (Brand kits come from brand_list.)
  2. load_skill — Moda’s full authoring guides, served live by the server. Call it before authoring a format you have not authored yet in this conversation. See Skills.
After that it is the normal loop: create or find a canvas, apply markup or an edit batch, screenshot it, repair, export.

​
What the connector can and cannot reach

The connector runs in the host’s cloud and talks to Moda’s servers. It never touches your machine. It can use: conversation context, attachments the host passes to tools, URLs you paste, and files already in your Moda workspace. It cannot use: local paths, local folders, repositories on your disk, or writing an export to a folder you choose. No connector tool takes a file path. Two consequences worth knowing:
  • Content goes inline. Markup and edit programs are passed as string arguments, not read from a file.
  • Results come back as links. Every mutation returns the canvas’s editor URL, and exports return a download link rather than writing a file. Opening it in the browser is the connector’s equivalent of the CLI writing the file next to your source.
If a host cannot pass an attachment’s bytes to a tool, the upload tool falls back to an in-conversation upload control, and failing that, to uploading the file at moda.app and telling the assistant to look for it.

​
The tool surface

The connector exposes one focused profile rather than everything Moda can do — a tool list an assistant can actually hold in its head: Work that takes a while — a video render, a large export — returns a task handle immediately; the assistant polls task_status rather than blocking the conversation. Authoring is the default: the assistant builds the canvas itself with the tools above. The deterministic tools are free, media_fetch_logo and media_video_frames included. Three consume the workspace’s Moda credits — canvas_edit_image, media_remove_background, and task_delegate. task_delegate is the other route — it hands the whole brief to Moda’s own design agent, which builds the design on a canvas by itself while the assistant waits. Like any long job it returns a task handle and the canvas link straight away. The assistant is told to reach for it only when your own words hand the work over — naming Moda’s agent, making Moda the one doing the designing (“I want Moda to design this deck”), or naming the hand-off itself (“delegate this to Moda”). “Use Moda to make a deck” is not that: there Moda is where the work happens, so the assistant authors it. That judgement is recorded rather than trusted: the tool requires a user_request_quote carrying your own wording that handed the work over, the assistant may not invent one, and the quote is kept on the call so a delegation that spent credits can be audited against what you actually asked for.

​
Troubleshooting

Still stuck? support@moda.app.