Connections that let an AI assistant reach your real tools: CMS, analytics, hosting.

Webflow

11 Aug

Lets an AI assistant work on a real Webflow site, from editing CMS items to building pages and changing styles, and it reads the site's traffic numbers too. Version 2 does most of that with the Designer closed.

Remote MCP server, mcp.webflow.com/mcp, OAuth, v2.0

FitsOne connection covers the content and the design, so you do not run a second setup for the CMS. Snapshots of elements and anything about the current selection still need the Designer open with the Bridge App running.

Switch on inClaudeconnectorChatGPTapp

Sanity

11 Aug

Gives an assistant your actual content and the shape it has to keep. It queries documents, patches them, prepares a release and knows your schema without being told.

Remote MCP server, mcp.sanity.io, OAuth, 40+ tools

FitsContent only. The site around it still lives in your framework and host, so this sits next to a Vercel or GitHub setup rather than replacing one.

Switch on inClaudeconnector

Storyblok

11 Aug

Lets an assistant work inside your space: find stories, write and update them, publish, and handle assets, using the components your space already has.

Hosted MCP server, mcp.labs.storyblok.com/mcp, personal access token

FitsIt reaches the Management API, which is everything a space admin can do, and it runs on Storyblok's labs host while it settles. Their CLI is the better tool for repeatable bulk work.

Vercel

11 Aug

Your assistant reads the deploy instead of you pasting logs into a chat. It says why a build failed, checks a project's settings and domains, and answers from your traffic numbers.

Remote MCP server, mcp.vercel.com, OAuth, reviewed clients only

FitsThe connection carries your own Vercel access, so anything you can change, it can change. Claude lists it under Connectors; ChatGPT needs developer mode and a paid plan.

Switch on inClaudeconnector, pluginChatGPTadd by URL

GitHub

11 Aug

Puts your repositories, issues and pull requests within reach, so an assistant can open a PR, read a review thread or check what a workflow did without you switching tabs.

Remote MCP server, api.githubcopilot.com/mcp/, OAuth, toolsets per URL

FitsAdd /readonly to the URL for a version that can look and not touch, which is the one to hand an assistant that only reports on work.

Switch on inClaudeconnector

Plausible

11 Aug

Lets your AI assistant read your traffic numbers, so you can ask which pages lost visitors last month instead of building a report.

Community MCP server by the Sentry team, Stats API key, self-hosted or remote

FitsTraffic only, and nothing about individual people. Pair it with Fullstory when the question moves from how many to why.

Amplitude

11 Aug

Ask your product numbers in a sentence, like which signup step people drop at or how last week's release moved retention. It can build the chart or cohort and save it back into Amplitude.

Remote MCP server, mcp.amplitude.com/mcp, OAuth, EU endpoint available

FitsReads and writes, and the two are separate permissions, so an account can be limited to reading. Data on EU residency connects to mcp.eu.amplitude.com/mcp instead, and ChatGPT needs developer mode turned on before it accepts the URL.

Switch on inClaudeconnector, pluginChatGPTadd by URL

Fullstory

11 Aug

Answers why people struggle on a page. It reads recorded sessions for rage clicks, dead ends, errors and abandoned forms, and points at the replays behind every answer.

MCP server inside Fullstory's own Claude Code plugin, OAuth or API key

FitsNeeds a paid Fullstory plan with StoryAI switched on by an admin. Without it the connection succeeds and no tools show up, which looks like a broken setup but is not one.

Switch on inClaudeplugin

Optimizely

16 Aug

The one that actually builds the test. Create the experiment, the flag and the audience from your editor, write the variation's code into it, and read which variant won.

Remote MCP server, exp.mcp.opal.optimizely.com/mcp, OAuth

FitsIt writes as well as reads: flags, experiments and audiences can all be created and changed on a live project, so connect an account you would hand that job to. Authentication runs through your Opal instance and the server sees exactly what that account sees. Past building the test it searches the Feature Experimentation SDK docs, generates integration code, and cross-references your repo to find flags nobody is using any more.

AI reachWritesAn assistant can change things here, not only read them.MCP serverClaudeadd by URLChatGPTadd by URL

VWO

16 Aug

Describe the test in a sentence and it creates the campaign, starts it, and reads the result back, along with the heatmaps and recordings that tell you what to test next.

Remote MCP server, mcp.wingify.ai/mcp, OAuth with dynamic client registration

FitsFour of its tools write: create an A/B or split URL test, clone one, change a campaign's configuration, and start, pause or stop it. So an assistant can move a live campaign - sign in with an account you would give a junior, not an admin. The tool list names only a read tool for variation code, and the docs do not say whether the description it takes also produces the variation itself, so plan on writing the change by hand until VWO says otherwise. Data held in the EU or Asia connects to mcp-eu.wingify.ai or mcp-as.wingify.ai instead, and ChatGPT needs developer mode before it accepts the URL.

AI reachWritesAn assistant can change things here, not only read them.MCP serverClaudeadd by URLChatGPTadd by URL

AB Tasty

16 Aug

The odd one out: it reads rather than builds. See which flags and campaigns a given visitor would get, and turn a written brief into the configuration for one.

Local MCP server, npx @abtasty/mcp-server, environment ID and API key

FitsThe only one of the four that cannot create an experiment. Its four API tools read what a visitor would be served, and the two resource-loader tools plus their prompts turn a written brief into configuration you still apply yourself. Local only, so there is no connector to switch on in Claude or ChatGPT: it needs a config file and a key. It covers Feature Experimentation and Rollouts, so it answers what a visitor gets rather than which variant won.

Kameleoon

04 Sept

It builds the experiment, not just the report: create the test, set its goal and audience, start and stop it, and read which variation won with the confidence attached.

Remote MCP server, mcp.kameleoon.com/mcp, OAuth, 37 tools

FitsThe widest write surface of the four: 37 tools covering experiments, feature flags, delivery rules, variations and typed variables, goals, segments and targeting, with one lifecycle tool that starts, pauses, resumes, stops or deletes a live test. Two ways in, for two different people. In Claude chat on web or desktop, Customize > Connectors > Add custom connector takes the server URL: sign in to Kameleoon, Authorize, then start a new conversation for the tools to show up. No editor and no terminal, so checking whether a test is running or flipping a flag stops being an engineering request. Claude does not list Kameleoon in its connector directory yet, which is why you paste the URL rather than search for it. Claude Code takes the same server through the mcp-remote helper, so that route needs Node, and the browser sign-in means a headless or cloud-hosted agent cannot finish it. Two limits to plan around: nothing writes a variation's code back into a running test, and experiment_create accepts MVT and SDK_HYBRID but fails on both every time, which the docs say outright.

AI reachWritesAn assistant can change things here, not only read them.MCP serverClaudeadd by URL