Reaktion MCP review: what it does, who it is for and what to check first

Everything below comes from the vendor's documentation at mcp.reaktion.com and the catalog data behind it, read on 27 October 2026. Where the documentation does not state something, such as pricing, this review says so rather than guessing.

Key Takeaways

• Reaktion MCP is a remote, read-only MCP server at https://api.reaktion.com/mcp. The vendor says it works in ChatGPT, Claude, Grok and Gemini, and that it never changes campaigns, budgets or flows.
• It exposes 1,204 metric definitions, 68 dimensions and 467 attributes over the data connected to a Reaktion organisation, plus 44 ready-made prompts. The 1,204 figure counts definitions: there are 457 distinct metric names, and the rest are the same measure on a different basis.
• Four reporting tools run in order for every performance question: inspect_reporting, prepare_report, export_report and read_report. Five inventory tools list the user, organisations, data sources and accounts.
• Store financials are reported excluding VAT and after refunds. Paid-channel ROAS and POAS come from attributed conversions that can include VAT and are before refunds. The two are labelled so they are not compared by mistake.
• It needs a Reaktion account with connected data sources and reports only on data Reaktion holds. Pricing, trial terms, plan requirements and rate limits are not stated in the documentation.

What the Reaktion MCP is

Reaktion (Reaktion.com ApS) is an e-commerce analytics and tracking company. Its MCP server gives an AI assistant read access to the normalized e-commerce, marketing, analytics and messaging data of a Reaktion organisation, with Reaktion's profit definitions applied. You add the server as a connector in the assistant, sign in with your Reaktion user, and ask questions in plain language. The documentation's own framing is that the assistant plans the report, runs it on your connected data and explains every number, so there are no dashboards to build and no exports to stitch together.

The data sources the documentation lists as connected to Reaktion organisations today are: Shopify, WooCommerce, Magento 2 and Dandomain for the store; Google Ads, Meta Ads, Pinterest Ads, TikTok Ads and Google Merchant Center for advertising; Google Analytics 4 and Google Search Console for analytics and search; Klaviyo for email and SMS; and Adtraction, Partner Ads and custom cash flows for affiliate and imported data. Reports combine them under normalized definitions, and a report can span several organisations, converted to each organisation's reporting currency.

What it exposes: metrics, dimensions and attributes

1,204 metric definitions

The catalog lists 1,204 metric definitions in ten families: Store (134), Profit and margins (24), Ads (39), Conversions (580), New customers (372), Analytics (24), Tracking (5), Email and SMS (20), Search (4) and Leads (2). Conversions and new-customer metrics make up most of the count because each measure exists in several bases. In the catalog data there are 457 distinct metric names; a name such as Tracking Uplift appears four times, once per conversion source and channel. Every definition carries tokens that state its basis: where the number comes from, the attribution source, and whether it is net (excluding VAT, after refunds) or gross (including VAT, before refunds).

A few definitions show how the profit metrics are built:

• Gross Profit: total sales after refunds, excluding VAT, minus estimated variable costs after refunds. It does not subtract ad spend or fixed and custom costs.
• Net Profit: total sales after refunds excluding VAT, plus custom cash flows, minus variable costs and ad spend. Recurring custom costs are allocated across the selected period.
• POAS: profit divided by ad spend, as a multiplier, on primary attributed conversions before refunds and including tax.
• New Customer POAS: new-customer profit divided by ad spend, on the same basis.
• Business-wise POAS: store-level gross profit excluding VAT divided by total ad spend across all paid channels. It is the metric the reporting rules point to when you need one total across platforms.
• Tracking Uplift: Reaktion server-side orders divided by pixel-attributed orders, as a percentage. Values above 100 percent mean Reaktion tracked orders the pixel missed.

Two things to check before relying on a number. First, six definitions in the catalog carry a QA flag, and three prompts in the prompt library are on QA hold (two about Google Search impression share and one about Thumb Stop Ratio), which the library states openly. Second, the Business-wise POAS entry describes its gross profit as before refunds while its basis token says after refunds, so for that metric read the basis stated in the report metadata rather than the catalog description.

68 dimensions

The 68 dimensions fall into twelve categories: Data sources (4), Stores and countries (2), Advertising and creatives (12), Attribution (6), Analytics (8), Orders and discounts (4), Products (11), Customers and leads (5), Messaging (3), Search Console (4), Custom finance (1) and Time (8). The attribution dimensions cover first click, last click, linear, journey-based, marketing mix and hybrid channel attribution. Product dimensions include cross-store matched product, variant, brand, type and SKU, and the customer dimensions include an RFM segment. Any compatible metric can be grouped by shop, campaign, product, customer segment, channel or time, and each dimension carries descriptive attributes that can be added as columns.

467 attributes

Attributes are the names, IDs, prices, dates, URLs and settings that belong to a dimension: the ad name and thumbnail beside spend and POAS, for example. The catalog deliberately excludes personal data. Customer emails, names, phone numbers and addresses are not available through the MCP; customer-level analysis works on anonymous IDs and aggregated attributes such as market, segment and consent state.

Periods and sources

Every report uses one main period and optional comparisons. The catalog lists 22 main periods, from Today and Yesterday through Last 30 days and Last 90 days to Last 18 months, and four comparison types: previous period, month over month, quarter over quarter and year over year. Relative periods resolve at the time of the report in your timezone, and "Last 30 days" excludes today. Custom ranges take any start and end date; a date-only end includes that whole day, a date-time end is exclusive.

How a question becomes a report

The documentation describes three steps for the user. Connect Reaktion as a connector and sign in; you see only the organisations your Reaktion user can already access. Ask in plain language, naming the store or client, the period and what you want to decide. Read an answer that states its basis: VAT, refunds, attribution source and currency.

Under the hood, four reporting tools run in order every time you ask a performance question:

1. inspect_reporting (discover) finds the right metric, dimension and period in the catalog.
2. prepare_report (plan) builds and checks a reporting plan.
3. export_report (run) executes the plan on your data.
4. read_report (read) returns the data with caveats and metadata.

Five inventory tools sit beside them: get_current_user, list_organisations, list_data_sources, list_accounts and summarize_accounts. All of them are read-only.

The documentation ties this sequence to what it calls three states of data. A metric being in the catalog (found by inspect_reporting) does not mean your organisation has it. It is compatible for you when your connected sources support it at the breakdown you asked for (organisation plus grain). Its values are reliable when data exists for the period, with the caveats stated in the export (read_report metadata). The documentation says a good answer states which state it is in, and the plan step exists so that a report is checked before it runs.

The reporting rules, and why they matter for agencies and CFOs

The documentation sets out five rules that the MCP follows and asks readers to follow when they compare or share results.

1. Store financials are reported excluding VAT and after refunds unless you ask otherwise. For a CFO reconciling an answer against the accounts, this is the basis that matches the books; say it in the prompt to be explicit.
2. Paid-channel ROAS and POAS use a different basis: primary attributed conversions, which can include VAT and are before refunds. They are labelled so they are not compared with store profit by mistake. A media buyer's POAS and a finance team's gross profit are therefore not the same number, and the gap between them is VAT, refunds and attribution, not an error.
3. Never add or average ratios. Percentages, shares, ROAS and POAS are recalculated from their parts; summing or averaging them across rows or days gives the wrong number. This matters in any client report that rolls up campaigns or weeks.
4. Be explicit about time. Refunds land after the order, and each source reports in its own timezone. Compare completed periods and matching weekdays, and expect the last few days to keep updating. A weekly client report built on the last three days will move.
5. Platforms overlap, so do not sum their conversions. Google and Meta can both claim the same order. When one total across channels is needed, the documentation points to store-level metrics such as Business-wise POAS.

A sixth point is stated with the rules: the MCP is read-only and permission-aware. It reads data your Reaktion user can access, agencies see each client separately, and client data is never merged. For an agency that is the governance answer to the question of who can see what: access follows the Reaktion login, not a shared key.

Setup per client

The documentation lists ChatGPT, Claude, Grok and Gemini as the clients the MCP works in, and links Reaktion's own installation guides for Claude, ChatGPT, Grok Bot and Gemini. The step-by-step installation guide on this site covers each client this site supports; the short version is below.

• Claude (claude.ai and Claude Desktop): add Reaktion as a connector with the URL https://api.reaktion.com/mcp and sign in. The vendor says this takes about two minutes and no code.
• ChatGPT: turn on Developer mode, add the Reaktion connector with the same URL and sign in with your Reaktion account.
• Gemini: add Reaktion under Connected Apps as a custom MCP server and sign in with your Reaktion account.
• Grok: the vendor's guide is for Grok Bot. You paste one message in the Grok Bot chat, and it adds the Reaktion MCP and walks you through authorization.
• Claude Code: not documented by the vendor. The standard command for a remote server is claude mcp add --transport http reaktion https://api.reaktion.com/mcp, and the Reaktion sign-in decides whether it works.
• Cursor: not documented by the vendor. The generic method is an mcpServers entry in mcp.json with the url set to the endpoint.
• Microsoft Copilot and Perplexity: not documented by the vendor. Nothing more specific can be said from the documentation.

Whichever client you use, the first prompts should be inventory questions: which user you are signed in as, which organisations you can access, and which data sources and accounts are connected with their reporting currencies. That confirms both the connection and the permissions before any report is trusted.

Strengths, as the documentation shows them

• The basis of every number is explicit. Each metric carries its VAT, refund and attribution tokens, and each answer states VAT, refunds, attribution source and currency. This is the part of the documentation most directly useful to anyone who has argued about whether a ROAS figure included tax.
• Profit metrics are first-class: gross profit, net profit, POAS, new-customer profit and Business-wise POAS are defined in the catalog with their formulas, which is possible because Reaktion holds cost data alongside orders and ad spend.
• Breadth of sources in one normalized model: store, five ad platforms, GA4, Search Console, Klaviyo, two affiliate networks and custom cash flows.
• Read-only by design, permission-aware, and no personal data in the catalog. Agency clients are kept separate.
• The prompt library states the dimensions, metrics and requirements for each of its 44 prompts, and marks three of them as on QA hold rather than hiding them.
• Caveats travel with the data: the three states of data and the read_report metadata are built into the tool sequence rather than left to the reader.

Limitations and open questions

• Pricing, trial terms and plan requirements are not stated in the documentation. The page has a free trial button that leads to a Reaktion sign-up page and a demo button that leads to a booking page, and nothing more about cost. Treat the cost of using the MCP as unknown until Reaktion states it.
• It requires Reaktion data. The MCP reports only on sources connected to a Reaktion organisation. It cannot read a store or an ad account that is not connected there, and the documentation says connection status does not certify freshness: the MCP reports what it actually found.
• No write actions, by design. It can recommend budget moves or negative keywords, and you make the changes in the ad platform. Teams that want an assistant to act will need a different tool for that step.
• Custom formulas and account or row filters are unsupported. The documentation says to use dimensions to break results down and ask the assistant to focus on the rows you care about.
• Net profit needs cost settings (product costs, shipping, payment fees and custom cash flows). Without them the metric is labelled unavailable instead of estimated.
• The catalog is larger than the set of measures. 1,204 definitions resolve to 457 distinct names, six definitions carry a QA flag, three prompts are on QA hold, and one entry (Business-wise POAS) describes its refund treatment differently from its basis token.
• Recent days keep updating. Refunds land after the order, and each source reports in its own timezone, so the documentation's advice is to compare completed periods and matching weekdays.
• Rate limits, data retention, lookback limits and the exact authentication mechanism are not stated on the documentation page.
• Client coverage is four assistants: ChatGPT, Claude, Grok and Gemini. Claude Code, Cursor, Microsoft Copilot and Perplexity are not documented. The documentation page also changed while this review was being written (Gemini and the Grok Bot and Gemini guides were added on 27 October 2026), so expect it to keep moving and re-check it before relying on a detail here.

Who should use it, and who should not

It fits merchants on Shopify, WooCommerce, Magento 2 or Dandomain who already run Reaktion with ads, GA4 and Klaviyo connected and want profit questions answered in chat without building dashboards. It fits agencies with several client organisations on Reaktion who want one evidence standard across clients and the guarantee that client data stays separate. It fits finance people who need numbers excluding VAT and after refunds, with the basis stated, rather than platform-reported revenue.

It does not fit anyone who is not a Reaktion customer, because there is nothing for it to read. It does not fit teams that want the assistant to change budgets or flows, teams that need custom formulas or row filters today, or anyone whose cost settings are incomplete and who needs net profit. Merchants on a store platform the documentation does not list should not assume a connection exists.

If the question is only about GA4 or Google Ads, the Google Analytics MCP Server and the Google Ads MCP Server in this directory are read-only and run on your own machine, and they need no Reaktion account. For profit across sources, the Reaktion MCP needs the Reaktion account and the connected data described above.

What to check first

1. Which organisations and data sources the connection can see, through the inventory tools.
2. Whether cost settings are complete enough for gross and net profit, and whether the metrics you plan to use are compatible for your organisation at the grain you need.
3. The basis of each number in the first few reports, against the five reporting rules, before any figure goes to a client or a CFO.
4. The pricing and plan terms, which you will have to get from Reaktion, since the documentation does not state them.