How to choose
The differences that matter are practical, not featural. These are the three worth deciding on before you read a setup guide.
Approval card, or text
Claude and ChatGPT render the Sprites card in the conversation, so a budget or a headline is edited where it is proposed. Everywhere else the change arrives written out and the approval waits in Sprites. Neither can apply anything on its own; the difference is how many steps it takes you.
Settings screen, or config file
The workspaces take a URL in a settings panel. The coding agents take a JSON or TOML entry, and the exact key differs per host — httpUrl, serverUrl, url with a type — which is the single most common reason a connector looks configured and shows no tools.
Where the work already is
A performance question belongs where you write; a tracking bug belongs where the code is. The strongest reason to pick a coding agent is not preference, it is that the account and the instrumentation are finally in one window.
Every agent, by the job it does
Grouped by what you would be using it for. Each card names the surface, how it is set up, who it suits and how approvals arrive.
General AI workspaces
Where a marketer already works. These draw the approval card in the thread or answer in plain text, and none of them need a config file.
The only host that draws the Sprites approval card in the chat, so you edit a budget or a headline in place and approve it without leaving the thread.
- Setup
- Settings or config file
- Approvals
- Card in the thread
- Guides
- 15 channel guides
Developer mode gives ChatGPT full read-and-write MCP tools rather than the search-and-fetch pair the built-in connectors are limited to, and the Apps SDK renders the Sprites card in the thread.
- Setup
- Settings screen
- Approvals
- Card in the thread
- Guides
- 15 channel guides
Grok reads the live timeline, so the same thread that spots a trend can brief the campaign that rides it.
- Setup
- Settings screen
- Approvals
- Text, approve in Sprites
- Guides
- 15 channel guides
The account and the analytics are already Google: Gemini reads Google Ads and GA4 alongside Search Console without leaving the family it belongs to.
- Setup
- Settings or config file
- Approvals
- Text, approve in Sprites
- Guides
- 15 channel guides
Coding and IDE agents
For growth engineers and technical marketers. Setup is a JSON or TOML entry, and the value is having the account open beside the code that measures it.
Marketing work that belongs in a commit: Codex can read the landing page it is about to ship and brief Sprites on the campaign pointing at it, in one session.
- Setup
- Config file
- Approvals
- Text, approve in Sprites
- Guides
- 15 channel guides
The campaign and the page it points at live in the same window: change the pricing copy, then have Sprites update the ads that quote it, without a second tool.
- Setup
- Settings or config file
- Approvals
- Text, approve in Sprites
- Guides
- 15 channel guides
Marketing data inside the editor where the tracking code lives, so a conversion event can be checked against what GA4 actually recorded before the pull request is opened.
- Setup
- Settings or config file
- Approvals
- Text, approve in Sprites
- Guides
- 15 channel guides
One MCP configuration shared by the IDE, the CLI, the SDK and Antigravity 2.0, so the connector is added once and every surface has it.
- Setup
- Settings or config file
- Approvals
- Text, approve in Sprites
- Guides
- 15 channel guides
Cascade can hold the marketing tools alongside the codebase, with per-tool toggles so only the ones a task needs are active.
- Setup
- Settings or config file
- Approvals
- Text, approve in Sprites
- Guides
- Hub only
An open client that shows you every tool call before it runs, which suits work against live ad accounts.
- Setup
- Settings or config file
- Approvals
- Text, approve in Sprites
- Guides
- Hub only
Autonomous builders
Agents that build and ship something. Marketing data here shapes what gets built rather than only reporting on it afterwards.
What it builds can be shaped by what actually converts: a landing page written against the search terms that are really paying, rather than against a guess.
- Setup
- Settings screen
- Approvals
- Text, approve in Sprites
- Guides
- 5 channel guides
Long-running work that outlasts a chat: Devin can carry a marketing engineering task across a session and keep the account tools available throughout.
- Setup
- Settings screen
- Approvals
- Text, approve in Sprites
- Guides
- Hub only
From account data to a deployed tool in one place: Replit Agent can build and host the reporting app it just read the numbers for.
- Setup
- Settings screen
- Approvals
- Text, approve in Sprites
- Guides
- Hub only
Watched, not published
Names we are asked about that do not have a page yet, and the honest reason why.
Meta Muse
Reaches the connector through a Custom Connector built in the chat rather than a URL field in its settings, so there is no setup path to verify in the way the thirteen above were verified. The guide covers what is and is not confirmed.
Connect Muse to your ad accountsGrok Bot
A persistent agent rather than a separate connector. Covered in a section on the Grok page; it gets its own URL when its setup diverges or search demand appears.
Roo Code
Real and rising, but below the threshold where a page would be more than a restatement of the Cline and Cursor setup.
Bolt
Watching. Its own MCP support and an MCP server that drives Bolt are two different things, and we will not publish steps until we have verified which one people mean.
Choosing an agent: questions
What is an MCP client?
The application that holds the conversation and does the calling. It speaks the Model Context Protocol to any server you add, reads the tool list that server publishes, and decides which tool to invoke. Claude, ChatGPT, Cursor, Codex, Gemini, GitHub Copilot and Grok are all MCP clients; the server is the thing on the other end that the tools belong to.
Which is the best MCP client?
For a marketer who does not code, Claude — custom connectors exist on every plan and the approval card renders in the thread. For a team standardised on ChatGPT, ChatGPT through developer mode. For anyone who wants the ad account open beside the code that measures it, Cursor or Codex. There is no single answer because the clients differ in setup and presentation rather than in what they can reach.
Do all of these have the same capabilities?
Through Sprites, yes — the same connector and the same channels. What differs is how each host presents the result and how it is configured, not what it can reach.
Which one should a marketer who does not code pick?
Claude or ChatGPT. Both draw the approval card in the thread, both are set up from a settings screen with no file to edit, and both are probably already open.
Which one should a growth engineer pick?
Whichever editor you already use. Cursor, Copilot, Codex, Antigravity, Windsurf and Cline all work; the advantage is having live analytics beside the code that produces it, and that applies equally to all of them.
Can I connect more than one?
Yes. The connector is per host, and the grant is per person, so adding Sprites to Claude and to Cursor gives you both with no interaction between them.
Why is my agent not listed?
Either it does not support custom remote MCP servers yet, or we have not verified its setup. We do not publish steps we have not read in a vendor's own documentation — a guide that is wrong is worse than no guide.
Are these official integrations?
Sprites publishes a standard remote MCP server. Any client that speaks MCP over streamable HTTP can use it, which is the point of the protocol — there is no per-host integration to wait for.
One connector, whichever you pick
Add the Sprites URL to the agent you already use. The channels, the tools and the approvals are the same in all of them.