MCP server

Your form data.
Your AI assistant.
One connection.

Exatom's full form and field optimisation suite is available over MCP. Connect ChatGPT, Claude or any MCP-capable tool, then ask about your forms in plain language — it reads the same analytics, session replays and nudges you work with in Exatom.

Start free trial Book a demo
Endpoint https://olympus-api.exatom.io/mcp
exatom · mcp

You

Why did checkout conversions drop on mobile last week?

Exatom MCP

›resolve-form"checkout"
›form-metrics-tooldevices: ["Mobile"]
›form-metrics-tooldevices: ["Desktop"]
›field-metrics-tooldevices: ["Mobile"]
›field-error-metrics-toolerror-messages

Mobile conversion sat at 18.4% against 31.2% on desktop. The gap is one field: Phone number shows a 34.1% error rate on mobile versus 6.8% on desktop, almost all of it “Enter a valid phone number”. Want me to check the replays for that field?

Example conversation — figures are illustrative

What is MCP

Rainfall and rainbows, technology and applicability

Technology falls like rain: steady, generous, and mostly straight into the ground. Every so often the light catches it and something appears you can actually point at — the moment a clever idea turns into something you can use on a Tuesday afternoon with a conversion target to hit. That arc from rainfall to rainbow is the test we apply to everything we build.

MCP made the arc. The Model Context Protocol is an open standard that lets AI assistants and agents connect to external tools securely. It took flight in 2026, when tools like ChatGPT and Claude built it into their apps and let you connect the software you already work in. In short: it gives your assistant the ability to act inside your apps, on your behalf.

That is the moment it became genuinely useful. A CRO specialist can now ask a question in one place and get an answer built from real form analytics — no export, no dashboard tour, no waiting on someone else's queue.

Exatom's MCP server is free to use on all plans. No add-on, no per-seat fee, no separate contract.

Once connected, your assistant can

Find the friction in your forms

Drop-offs, corrections, validation errors, rage clicks and hesitation — at form level and down to the individual field.

Read session replays as behaviour

Not just the numbers, but what people actually did: the event sequence, in order, with timings — and which field they left on.

Correlate, compare and summarise

Mobile against desktop, autofill against no autofill, campaign against campaign — in the same answer, with the baseline alongside.

Draft Smart Nudges against what it found

Turn a diagnosed friction into a reviewable nudge on the right field, with a rendered preview, before anything is saved.

Run the same analysis on a schedule

Let your assistant repeat a weekly form health check and tell you only what moved, instead of you going to look.

What it can do

One connection, the whole toolkit

Twenty-one tools across five areas. You never call them by name — you ask a question, and your assistant works out which ones to use and how to filter them. Everything is read-only except drafting a nudge, which needs your explicit go-ahead.

Form & field analytics

8 tools

The core of it. Traffic, starts, errors, every kind of abandonment, conversions and conversion value at form level — then drop-offs, corrections, errors, rage clicks, autofill and hesitation for each individual field, including the order people really fill them in. One form or your whole website at once, with the exact error wording your visitors saw.

Session replay

2 tools

Behaviour, not just analytics. Your assistant sizes the cohort worth looking at, then reads what those people actually did — step by step, with timings and the field they gave up on. It hands you a direct link to any replay you want to watch yourself. Privacy-first, every session replay is fully anonymised. Sensitive inputs — names, dates of birth, email addresses, payment details — are automatically masked.

Smart Nudges

4 tools

Close the loop on what you found. Your assistant drafts a nudge on the right field with the right trigger, renders a preview so you see exactly what a visitor would see, and shows you what already runs on the form. Nothing is saved until you say so, and nothing reaches visitors until you activate it through our user-interface.

Account & setup

6 tools

Your assistant finds the right website and the right form from whatever you call it — a domain, a brand name, a form label you half remember. It reads how the form is set up before it reads the numbers, so a figure gets explained by configuration rather than guessed at. When your request could mean two different forms, it asks instead of picking one.

CRO expert

1 tool

Evidence is only half the job. Exatom's in-house form, UX and conversion knowledge interprets what you found: why a pattern is probably happening, what to change, and how to prove it worked — the right field type, autocomplete and mobile keyboard for the field in question, and a way to verify the fix. Grounded in our documented practice rather than improvised, and it only ever sees an anonymised summary of your numbers.

Filter every insight the way you already think about it

Form metrics, field metrics, error metrics and session replays all take the same filters, so any comparison stays apples to apples. Describe the cohort in your own words and your assistant applies it.

Any timeframe

Last week, last quarter, two specific dates, or the same period a year ago.

Device

Mobile, Desktop or Tablet.

Browser

Chrome, Safari, Edge, Firefox, Samsung Browser, and in-app browsers such as Facebook and Instagram.

Country

Zoom in on a specific country of interest.

Validation errors or not

Only the people who ran into an error, or only the ones who never saw one.

Autofill or not

Only the people whose browser filled the form for them, or only the ones who typed it themselves.

Conversion path

A specific point in the path: initial abandoners, submit failed, initial submit but abandoned, or converters.

UTM campaigns

Any UTM source, medium, campaign, content or term — including traffic that arrived with no campaign at all.

Your own dimensions

Whatever you already send us: logged-in state, product category, plan, test variant.

Smart Nudges

Only the people who were shown a specific nudge, so you can see what it changed.

Field level Replays

Only the sessions that touched a particular field — or the ones that gave up on it.

Worth reviewing Replays

Only the sessions our friction score flags for review, or the replays you bookmarked yourself.

In practice

What our Customer Success team asks it

Real working questions, typed as-is. No query language, no field names to memorise — and several of them lean on your assistant's other abilities too: a template, a spreadsheet, your ticketing tool.

Yesterday we changed the mobile version of our checkout. Do an end-to-end pre/post analysis and use session replays to identify the underlying frictions.

Using our brand presentation template, create a quarterly report on our forms that identifies trends and opportunities.

Show me the ten most frictional replay sessions from last week and tell me what they have in common.

Compare mobile against desktop performance for the quote form last month, at form level and field level.

Scan my forms and fields for autofill coverage, then create a ticket for our agency covering every place we deviate from the standards.

Of the people who submitted and then abandoned, where in the form did they actually give up?

Do an in-depth field-level error analysis: which error messages do our customers see, and are they shown at the right moment?

Using the attached marketing spend sheet, analyse which UTM codes have the biggest impact on conversions and how we can optimise for efficiency.

Draft a nudge for the IBAN field based on the errors you found, and show me a preview before anything is saved.

Every Monday, compare last week with the week before and give me a summary plus a to-do list for the week ahead.

Chrome versus Safari on mobile for the registration form, find out why Safari produces more errors.

How to set it up

Connect in under a minute

One URL and one sign-in. There is no API key to generate, copy or rotate.

HTTP transport

Recommended

A direct connection to the endpoint, with no extra software. Everything current — Claude, ChatGPT, Claude Code, Codex, Gemini CLI, Cursor, VS Code — supports it.

stdio via mcp-remote

Fallback

For older clients that only speak stdio, mcp-remote proxies the same endpoint. Requires Node.js on the machine running the client.

01

Have your Exatom login ready

The same email and password you use for the app. Your assistant inherits your permissions exactly — it sees the companies, websites and forms you see, and nothing beyond them.

02

Add the endpoint to your AI tool

https://olympus-api.exatom.io/mcp

Pick your integration below for the exact steps.

03

Sign in when prompted

We use OAuth, the industry standard for connecting apps securely. You sign in on Exatom's own screen and approve the connection there; no credential is ever pasted into your AI tool.

Revoke a connection any time under Connected apps in Exatom.

Integrations

Claudeapp & web
Endpointhttps://olympus-api.exatom.io/mcp
  1. Open Settings → Customise → Connectors.
  2. Click + Add custom connector.
  3. Name it Exatom and paste the endpoint above and click Continue.
  4. Click Connect and sign in with your Exatom account.
  5. In a chat, make a reference to Exatom in your prompt and your assistant will use Exatom in its research.

Team and Enterprise: an owner adds it once under Organization settings → Connectors → Add → Custom → Web; members then find it under their own Connectors and click Connect.

ChatGPTapp
Endpointhttps://olympus-api.exatom.io/mcp
  1. Open Settings → Integrations → Plugins.
  2. Choose Add MCP server and name it Exatom.
  3. Select Streamable HTTP as the transport.
  4. Enter the endpoint above, then save and restart if asked.
  5. Click Authenticate and sign in with your Exatom account.
  6. In a chat, make a reference to Exatom in your prompt and your assistant will use Exatom in its research.
Claude CodeCLI

Terminal

claude mcp add exatom --transport http https://olympus-api.exatom.io/mcp

# then, inside Claude Code:
/mcp        # select exatom → Authenticate

# verify:
claude mcp list

The browser opens Exatom's sign-in once; the connection is remembered afterwards.

CodexCLI

Terminal

codex mcp add exatom --url https://olympus-api.exatom.io/mcp
codex mcp login exatom

Run codex mcp list to confirm, then /mcp in the TUI to see the tools.

GeminiCLI

Terminal

gemini mcp add --transport http exatom https://olympus-api.exatom.io/mcp
Cursoreditor

~/.cursor/mcp.json (global) or .cursor/mcp.json (project)

{
  "mcpServers": {
    "exatom": {
      "url": "https://olympus-api.exatom.io/mcp"
    }
  }
}

Open Settings → MCP, then sign in on the Exatom entry to complete the connection.

VS Codeeditor

.vscode/mcp.json

{
  "servers": {
    "exatom": {
      "type": "http",
      "url": "https://olympus-api.exatom.io/mcp"
    }
  }
}

Or run MCP: Add Server from the Command Palette and pick HTTP. Start the server from the file's inline action and sign in when prompted.

Any stdio clientmcp-remote

For clients that cannot speak HTTP yet, proxy the same endpoint through mcp-remote. Requires Node.js.

Client config

{
  "mcpServers": {
    "exatom": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://olympus-api.exatom.io/mcp"]
    }
  }
}

The OAuth sign-in still happens in your browser; mcp-remote only relays the connection.

Good to know

Where the edges are

Two things worth reading before you connect: what the server cannot do yet, and how to treat what it tells you.

Current limitations

These are on the roadmap and being worked on actively. For now they stay in the Exatom app.

Managing forms and fields

You can read a form's full configuration, but not create, rename or reconfigure forms and fields over MCP.

Activating and modifying Smart Nudges

Drafting and previewing work today. Activating a nudge, or changing an existing one including its status, happens in Exatom. A 50/50 A/B test is set up with your Customer Success Manager.

Managing Autofixes

Autofixes are still proposed and approved inside Exatom; they are not exposed over MCP yet.

Screenshots of Exatom views

A screenshot of an Exatom view is often exactly what a report or a presentation needs. You can't request one yet — we're working on making the Conversion Path, Field Metrics and other views available as images.

Missing something that would change how you work? Tell us, or raise it with your Customer Success Manager.

AI use

How AI is involved, and what that means for the answers you get.

Our CRO expert

Our CRO expert uses OpenAI to answer from Exatom's documented corpus, so Exatom's AI terms and conditions apply to it. Every other tool returns your data unchanged, with no model in the path.

AI can make mistakes

Review every answer for accuracy before you act on it, and especially before you ship changes.

Association is not causation

We deliberately keep observed evidence, hypotheses, recommendations and verification steps apart, and ask the assistant to report them that way. A metric that moves together with another is something to investigate, not a proven cause.

Your MCP access is the same as using our User Interface. Treat an MCP connection like a login, and review your Connected apps periodically.

Under the hood

Re-engineered so AI knows Exatom inside-out

We rebuilt our MCP infrastructure from the ground up for AI UX — the experience the model has of our product. A connected assistant should not have to guess how form optimisation works.

Next steps, not dead ends

Every tool says what usually follows it

Your assistant follows a real investigation instead of calling tools speculatively. We applies AI User Experience to each tool so it knows what is next — form metrics point at field metrics, field errors point at replays.

Best practices and don'ts

The rules travel with the data

Which filters can be combined, why a rate on a small denominator is noise, when an association is a signal rather than a cause, and when to keep looking instead of concluding. Written down, so it is not improvised.

Typed in, typed out

Schemas name every column

Every tool publishes both an input and an output schema, so each metric is defined rather than inferred from its name. Invalid arguments are rejected with a message that says how to fix them.

FAQ

Everything you need to know about Exatom over MCP

Can't find the answer? Talk to our team — we're happy to help.

What is MCP, and what is an MCP server?

MCP — the Model Context Protocol — is an open standard that lets AI assistants and agents connect to external tools securely. An MCP server is the piece an application publishes so assistants can reach it.

Exatom's MCP server exposes our form and field optimisation suite as tools your assistant can call: analytics, session replays, Smart Nudges and CRO guidance. You ask in plain language; the assistant picks the right tools, applies the right filters and reports back.

Is Exatom's MCP server free?

Yes. It is included on every Exatom plan at no extra cost — no add-on, no per-seat fee and no separate contract. You need an Exatom account and the data you already collect.

Which AI tools work with it?

Any MCP-capable client. We document the exact steps for Claude, ChatGPT, Claude Code, Codex, Gemini CLI, Cursor and VS Code, and anything else that speaks MCP connects the same way — with HTTP transport directly, or through mcp-remote if it only supports stdio.

How does authentication work, and what can my assistant see?

We use OAuth, the industry standard for connecting apps securely. Your AI tool sends you to Exatom's own sign-in screen and only receives access after you approve it there. No API key or password is ever handled by the AI tool.

The connection carries your permissions. Your assistant can reach exactly the companies, websites and forms your Exatom roles allow, and nothing else. Revoke it any time under Connected apps, which invalidates the token immediately.

Can my assistant change or delete anything?

Currently, no destructive actions are possible through the MCP. We are working on making that possible in a secure way.

The single exception is drafting a Smart Nudge, it will always be inactive and you will need to visit the user interface to test and activate it.

Does my assistant see visitor data?

Only what you ask for. Form and field analytics are aggregate analytics — no individual visitor in sight.

Session replay timelines are different: a replay can contain page URLs, field labels taken from your pages and the exact error text a visitor saw, so treat it as you would review a session recording.

How is this different from the Exatom API?

Same data, different way of accessing it. The API is for developers building an integration. MCP is for the people doing the optimising: you ask a question and the assistant works out which calls to make.

Use MCP for the ad-hoc analysis that used to mean an export and an afternoon.

Can it run on a schedule?

That depends on your AI tool rather than on us. Any client that can run a saved prompt on a schedule — or an automation platform pointed at one of our endpoints — can turn “compare last week to the week before and flag what moved” into a standing weekly check. The server itself is stateless per call, so a scheduled run behaves exactly like a manual one.

Connect Exatom to your assistant today

Free on every plan. One URL, one sign-in, no API key to manage — and your form analytics answer questions in the tool you already have open.

Start free trial Book a demo