Exports, MCP & Developer Tools

What Is MCP, and Why It Matters

Every other article here tells you how to read a screen. This one is about not opening TrackIQ at all. MCP — the Model Context Protocol — lets an AI assistant you already use query your Amazon data directly. You ask Claude Desktop a question in plain English, it calls TrackIQ, and it answers from your real numbers. No export, no copy-paste, no screenshot. Watch the MCP walkthrough Prefer to watch? Seventy-six seconds on what MCP is and when to reach for it instead of Aria. The problem it removes The usual way to get Amazon data into an AI assistant is to export a CSV, open it, paste it into a chat, and hope the assistant reads the columns the way you meant them. That breaks in three places. The export is a snapshot, so it is stale the moment you take it. It is truncated, because a chat window will not hold a year of daily rows. And the assistant has no idea what the columns mean — it does not know that SP uses a 7-day attribution window and SB uses 14, so it will happily add them together. MCP fixes all three at once. The assistant queries live, asks for exactly the slice it needs, and TrackIQ tells it what each figure means as part of the answer. The MCP access screen in Settings What it looks like in practice You type a question into Claude Desktop. Behind the scenes it picks a tool, fills in your account, date range and filters, and reads back the result. “Which of my campaigns lost money last month?” “Compare my ACoS this quarter against last, by ad type.” “Which products are running out of stock in the next two weeks?” “Which search queries do I convert on that I have no keyword for?” You never name a table, a filter or a date format. That is the whole point — the tools are described well enough that the assistant works out which one to reach for. Four things worth doing with it Ask questions the UI has no screen for. The product has a screen for campaigns and a screen for inventory. It does not have a screen for “campaigns advertising products that will be out of stock before the click converts”. An assistant with access to both can answer that in one turn. Do the analysis where your other work already is. If you write your weekly summary in Claude, pulling the numbers in beats alt-tabbing to fetch them. Let a non-analyst self-serve. A founder or a client can ask “how did we do last week?” without learning where anything lives. There is no dashboard to teach. Build something on top. The same access is available as a plain bearer token, so a script, a scheduled job or an internal tool can read the same data. See Installing TrackIQ MCP in Claude Desktop and Cursor. How this differs from Aria Both answer questions in English. They are not the same tool, and the difference is where the conversation lives. | | Aria | MCP | |---|---|---| | Lives in | TrackIQ, and Slack | your own AI assistant | | Knows about | your Amazon data | your Amazon data plus everything else in that chat | | Can draw | charts, tables, dashboards, Studios | whatever your assistant can render | | Set-up | none | paste one URL, once | The practical rule: use Aria when the question is only about Amazon, because it will give you a chart and let you pin it. Use MCP when the Amazon numbers are one input into a bigger piece of work — a plan, a client email, a model that also involves your costs or your other channels. What it can see, and what it cannot It reads. It does not write. Nothing an assistant does through MCP can change a bid, pause a campaign or alter your account — those actions live in Rule-Based Optimization, under your control. Access is scoped to one brand, chosen when you connect. An assistant connected to one brand cannot read another. And every request is logged. Settings → MCP & API Keys → Call Logs shows the question that was asked, how many tool calls it took, and when — so you can see exactly which AI tools have been talking to your data. The Call Logs view records every question and its tool calls Where to go next - Installing TrackIQ MCP in Claude Desktop and Cursor — the one-URL set-up, and the manual path for clients that need it. - The Full MCP Tool Reference — all 26 tools, what each returns, and the traps in a few of them. Related - Exporting Data to CSV and Excel — the non-AI way to get data out.

Installing TrackIQ MCP in Claude Desktop and Cursor

For Claude Desktop this takes under a minute and involves pasting one URL. There is no JSON file to edit and no credential to copy, because the connector registers itself. Everything else on the page exists for clients that cannot do that. Most people never need it. Watch the set-up walkthrough Prefer to watch? Sixty-eight seconds from copying the URL to your first question. Find the screen Settings → MCP & API Keys. It sits under Workspace, below Connections. The MCP access screen, with the server URL and the four steps The MCP server URL is the same for everyone: https://app.trackiq.com/mcp Use the Copy button rather than retyping it. Claude Desktop: the one-URL path The four connect steps 1. Open the custom connector dialog. In Claude Desktop, Settings → Connectors, then Add custom connector. 2. Paste the URL and click Connect. 3. Pick a brand and approve. A TrackIQ tab opens in your browser. Log in if you are not already, choose which brand the connector should be able to query, and click Approve. 4. Ask something. Back in Claude Desktop, try “List my marketplaces” as a smoke test, then something real like “Top 10 SP campaigns by spend last 30 days”. Step 3 is the one that matters for security. The brand you pick is the only brand that connector can ever read. If you manage several and want an assistant that can see all of them, you connect once per brand. Cursor, scripts and everything else Claude Desktop supports OAuth Dynamic Client Registration, which is why it can register itself. Clients that do not yet support it need credentials issued in advance, and that is what the Advanced section is for. The Advanced section, expanded The screen says it plainly: most customers do not need anything below. Open it only if the one-URL path does not apply to your client. Manual OAuth clients issue a pre-made client_id / client_secret pair. The brand is still chosen at connect time on the consent screen, exactly as above. | Use it for | Do not use it for | |---|---| | Cursor (current versions) | Claude Desktop — paste the URL instead | | In-house OAuth-aware integrations | anything that already self-registers | | OAuth proxies | | Per-brand API keys are long-lived bearer tokens locked to a single brand, sent as an Authorization: Bearer header on every request. Per-brand API keys, one section per connected brand Reach for a key when the thing calling TrackIQ does not speak OAuth at all: curl or Postman, claude mcp add from the command line, the older mcp-remote bridge, or a server-to-server script on a schedule. Each connected brand gets its own block showing active and revoked counts, with New Key to issue one. The Setup instructions for Cursor, Claude Code, & direct API button on the same panel opens step-by-step instructions for those clients specifically. Treat a key like a password An API key is a standing credential. Anyone holding it can read that brand's data until it is revoked. - Issue one key per thing that uses it, not one shared key. Revoking a shared key breaks everything at once; revoking a named one breaks only what you meant to break. - Never paste a key into a chat, a ticket or a shared doc. Environment variables and secret stores exist for this. - Revoke rather than delete-and-forget. The revoked count on each brand is your audit trail. If you only need Claude Desktop, you never issue a key at all — which is the safest state to be in. Checking it worked Two independent confirmations: In your assistant, ask “List my marketplaces”. It should come back with your connected marketplaces and account types. In TrackIQ, open Call Logs from the top of the MCP screen. Every request appears with the question that prompted it, the number of tool calls it took, and when it was posted. Call Logs, empty until the first request arrives An empty log after a successful-looking connection means the assistant registered but has not actually asked anything yet — connecting and querying are separate events. If it does not connect | Symptom | Cause | |---|---| | Browser tab opens but never returns | you were not logged into TrackIQ — log in, then retry | | No brand appears on the consent screen | that brand is not connected yet — see How to Connect Your Amazon Accounts | | Assistant sees the connector but every tool errors | the brand's Amazon connection has expired — see Fixing a Broken or Expired Amazon Connection | | Some tools work, others say unavailable | those tools need a higher plan tier | That last row is expected rather than broken. See MCP Tiers: What Each Plan Unlocks. Related - The MCP Web Console — the page you install from. - MCP Call Logs: Seeing What Your AI Tools Asked — checking what your client asked. - The Full MCP Tool Reference — what it can query.

The Full MCP Tool Reference

There are 26 tools. You will never call one by hand — your assistant picks them for you — so read this the way you would read a menu, not a manual: to know what is possible to ask for. The useful sections are the last two. The grain of each tool explains why some answers come back per day and others per week, and Six traps covers the places an assistant can produce a confident wrong number. Watch the tool reference walkthrough Prefer to watch? Sixty-seven seconds on how the tools group, and the traps worth knowing. The MCP access screen — one URL, and the tools come with it Start here | Tool | What it does | |---|---| | list_marketplaces | lists your connected marketplaces and whether each is a Seller, Vendor or KDP account | | set_context | records the question being asked, so the call is grouped in your Call Logs | list_marketplaces is the first call in almost every conversation, because everything else needs to know which marketplace you mean. It is also why the account type matters: sellers identify products by SKU, vendors by ASIN, and the same question takes a different shape on each. Sales, products and stock | Tool | Grain | Notes | |---|---|---| | get_account_overview | per day, or one rollup | sales_trend for a chart; advertising_summary for ACoS, TACoS and ROAS in one row | | get_product_performance | per product, or per day | the traffic funnel: sessions → orders → units → revenue | | get_product_categories_performance | per category | how your categories compare | | get_inventory_snapshot | point in time | on hand, inbound, reserved, unfulfillable; not a time series | | get_bsr | per product | Best Seller Rank | | show_products | — | renders product cards or a table with images | show_products is the only tool that exists purely for presentation, and assistants are told not to call it unless you actually ask to see pictures — so if you want thumbnails, say so. Sponsored Ads | Tool | Grain | |---|---| | get_campaigns | per campaign, or per day | | get_campaign_groups | per campaign group | | get_portfolios | per portfolio | | get_ad_groups | per ad group | | get_targets | per keyword or product target | | get_product_ads | per advertised ASIN | | get_search_terms | per customer search term | Two behaviours are worth knowing because they change the answer. Only enabled entities are counted by default. Ask “how many campaigns do I have?” and you get the active ones. If you want everything, say “including paused and archived”. One call covers all three ad types. get_campaigns ranks Sponsored Products, Brands and Display together in a single combined list, tagging each row with its type — so “my top 10 campaigns” genuinely means top 10 across the account, not top 10 of one channel. Amazon DSP get_dsp_performance is one tool with seven lenses. The metric set is identical each time; only what a row represents changes. | Lens | One row per | |---|---| | campaign | DSP campaign | | ad_group | DSP ad group | | creative | creative asset | | audience | target and segment | | geography | country | | technology | device | | inventory | supply source | DSP uses ACTIVE / INACTIVE, not enabled/paused, and defaults to ACTIVE. Every lens reports on a 14-day attribution window and includes the display-specific columns — detail page views, add-to-cart and new-to-brand purchases. See DSP Metrics Explained. AMC Cloud | Tool | Answers | |---|---| | get_amc_ntb_asins | which products bring in new customers | | get_amc_ntb_purchases | which campaigns bring in new customers | | get_amc_time_to_conversion | how long people take to buy after seeing an ad | | get_amc_attribution_paths | which sequences of ads actually convert | All four are monthly, and each month is a separate cohort of people. They are compared, never added. Search and rank | Tool | Grain | |---|---| | get_search_query_performance | per query, per week | | get_search_query_cart_analysis | per query — why it does not convert | | get_keyword_rank | per keyword, per product, per day | get_search_query_performance has a by_product switch that changes the question entirely: off, it answers “how does the market behave on this query?”; on, it answers “which of my products owns which search?” Vendor accounts | Tool | Answers | |---|---| | get_vendor_product_sales | shipped and ordered revenue | | get_vendor_inventory_health | sellable, unsellable, aged inventory | | get_vendor_forecasting | Amazon's demand forecast, with p70 / p80 / p90 bands | These return nothing on a Seller account, and the sales tools above return nothing on a Vendor account. That is the account type doing its job, not a fault. Call Logs records the question and how many tool calls it took Call Logs is the quickest way to see which tools an assistant actually reached for — the TOOL CALLS column counts them per question. The grain of each tool This is the single most useful thing to internalise, because it explains almost every “why can it not just tell me X” moment. | Grain | Which data | Consequence | |---|---|---| | Daily | sales, ads, rank | daily breakdowns are capped at the last 30 days | | Weekly | Search Query Performance | pinned to one week; never blended across weeks | | Monthly | everything AMC | each month is its own cohort | | Point in time | inventory | there is no “stock last Tuesday” | So “show me daily spend for the last quarter” comes back as the last 30 days with a note. That is the cap, not a bug — ask for a total or a weekly roll-up instead. Six traps Worth knowing, because a wrong answer here looks entirely plausible. Ranks are positions, so lower is better. “Improve rank” means the number goes down. A blank rank is not rank zero. It means the product was not found within the depth that was checked — different from ranking badly, and very different from zero. Combining ad types blends attribution windows. SP is credited over 7 days, SB and SD over 14. A combined ranking is right for comparing size, wrong for comparing efficiency. SQP shares do not add across weeks. 20% one week and 20% the next is not 40%. See SQP Is Weekly. Vendor revenue is the manufacturing view. Sourcing is a second lens on the same business, never added on top. Summed forecast percentiles are a planning heuristic. A single day's p80 means what it says; ten days of p80 added together does not, because percentiles are not additive. Ask for the mean when you need an exact figure. Not every tool is on every plan Some are gated by tier — advertising tools need the Advertising tier or above, and AMC needs a plan that includes AMC. An assistant told a tool is unavailable will say so rather than guess. See MCP Tiers: What Each Plan Unlocks. Related - Feed TrackIQ Data into Your Own AI Workflow — what to actually do with them. - MCP Call Logs: Seeing What Your AI Tools Asked — seeing which tools got used.

MCP Tiers: What Each Plan Unlocks

MCP access is on every plan. You never need to upgrade to connect Claude Desktop — that is not what the tiers are about. What differs is which data your MCP connection can reach, and it mirrors the plan grid exactly. If a data source is not on your plan, the tools that read it are not available either. Watch the walkthrough Prefer to watch? Seventy-three seconds on why MCP access is on every plan, and what actually differs. The rule, in one line MCP reaches the data your plan includes. Nothing more, nothing less. There is no separate MCP entitlement to reason about. Read the Data & Analytics rows of the comparison grid and you have your answer. MCP Access is ticked on all four plans; MCP Web Console is MCP Only Two rows in that section are worth separating, because they look similar and are not: MCP Access — ticked on all four plans. This is the connection itself. MCP Web Console (setup & config) — MCP Only. That plan's entire product is MCP, so it gets a dedicated surface for configuring it. On the other plans you set MCP up from Settings → MCP & API Keys instead, which does the same job. What each plan's MCP can query The Data & Analytics rows decide what MCP can reach | Data | MCP Only | Starter | PRO | PRO+ | |---|---|---|---|---| | Sales & traffic | ✓ | ✓ | ✓ | ✓ | | Inventory | ✓ | ✓ | ✓ | ✓ | | Sponsored Ads (SP/SB/SD) | ✓ | ✓ | ✓ | ✓ | | Organic & Best Seller rank | ✓ | ✓ | ✓ | ✓ | | Category Insights | — | — | ✓ | ✓ | | Search Query Performance | ✓ | — | ✓ | ✓ | | Amazon DSP | ✓ | — | ✓ | ✓ | | AMC Cloud | ✓ | — | ✓ | ✓ | The three bold rows are the whole story, and they produce the counter-intuitive result on the pricing page: MCP Only reaches more data than Starter, at less than half the price. That is not a mistake. MCP Only is sold as Full MCP Access — the complete data surface, with no web application. Starter is the application at entry depth. Neither is a subset of the other. What it looks like when a tool is not on your plan Your assistant will say the tool is unavailable rather than inventing a figure. It does not fail silently, and it does not guess. So on Starter, asking “which queries do I convert on but have no keyword for?” comes back as “I do not have access to Search Query Performance” — not an empty list, which would be far worse, because an empty list looks like an answer. Distinguishing the three “no data” cases matters: | What you see | Cause | |---|---| | Tool unavailable | your plan does not include that data | | Empty result | the data exists but there is none in your range — check brand, dates, filters | | Nothing recent | it has not arrived yet — SQP is weekly, AMC monthly | See My Dashboard Is Empty for the second and third. How many tools you get There are 26 in total. Every plan gets the setup tools and the core sales, inventory, advertising and rank tools. PRO and above add the Search Query Performance, DSP and AMC tools; MCP Only has them too. Vendor-specific tools return nothing on a Seller account and vice versa — that is your account type, not your plan, and it is not something you can upgrade past. For what every tool does, see The Full MCP Tool Reference. Choosing a plan for MCP Work backwards from the questions you want to ask. - “I only want my assistant to reach my Amazon data.” → MCP Only. The full data surface, no web app. - “I want the app and MCP, and I do not need SQP, DSP or AMC.” → Starter. - “I want the app and I want search, display or AMC questions answered.” → PRO. - “More than ten people, or role-based AI agents.” → PRO+. And if you are unsure whether the advanced data is worth it, go up for a week and look. Plan changes take effect at any time and nothing you have built is lost — see Changing Plans: Upgrades and Downgrades. Two things that are not tiered Usage. There are no token limits or usage caps on any plan, so an assistant querying through MCP all day is not metered. See Are There Token Limits or Usage Caps? Write access. MCP is read-only on every plan. No tier lets an assistant change a bid or pause a campaign — that lives in Rule-Based Optimization, under your control. See Security and Data Protection. Related - TrackIQ Plans Explained: MCP Only, Starter, PRO and PRO+ — the plans themselves.

Exporting Data to CSV and Excel

Export appears on every table screen in TrackIQ, next to Manage Columns. It is available on every plan, MCP Only included. The one thing worth knowing before you click it: an export takes what you are looking at, not the whole table. Watch the walkthrough Prefer to watch? Sixty-nine seconds on what an export actually contains, and when not to run one. Exports respect the view Search, Filters, Manage Columns and Export sit above every table Your filters, your column choices, your date range and your sort all carry into the file. That cuts both ways, and it is the source of nearly every export surprise: The good half. Set the view up properly and the file arrives ready to use — no deleting columns, no filtering in the spreadsheet. The half that catches people. A filter you set last week is still applied, because filters persist between sessions. Active filters show as chips above the table. If there are chips and you did not put them there deliberately, your export is already narrower than you think. Check the row count before exporting. If the table says it is showing 40 of 2,000 entries, that is what you are about to get. Getting a full export 1. Clear the filter chips. 2. Set the date range you actually want. 3. Open Manage Columns and switch on anything missing — a column that is off is a column that will not be in the file. 4. Then export. Step 3 is the one people skip. Manage Columns is also how you get a narrower file — turning columns off before exporting is quicker than deleting them afterwards. The Exports queue Exports are prepared in the background, so nothing blocks while a large one builds. Exports in the sidebar lists them. The exports queue | Column | What it tells you | |---|---| | Export type | which screen it came from | | Account | which connected account | | Marketplace | which marketplace | | Status | whether it is ready | | Requested at | when you asked for it | The queue is empty until you run an export, so a blank list on a new account is expected rather than broken. Account and Marketplace are the useful columns if you manage several. Two exports of the same screen taken minutes apart are indistinguishable by name — those two columns are what tell them apart. The queue has its own Search, Filters, Manage Columns and Export. When not to export Exporting is a habit worth questioning, because three things often do the job better. For a figure or a list — ask Aria with Tables selected. Faster than a round trip through a spreadsheet, and there is no usage cost to asking. For something you check regularly — a Studio or a scheduled Business Report. An export is a snapshot that starts going stale immediately; both of those stay current. For feeding another tool — MCP. A script or an AI assistant reading live data beats a CSV that has to be regenerated every time. See What Is MCP, and Why It Matters. The strong case for an export is a one-off — a reconciliation, an audit, or handing figures to someone who does not use TrackIQ. Two things exports cannot fix They do not extend your plan. An export contains what your plan includes. Search Query Performance is not on Starter, so it is not in a Starter export either. They do not settle attribution. Exporting today's advertising figures captures a period that is still filling in — Sponsored Products is credited over 7 days, Brands and Display over 14. Export a closed period if the numbers need to be final. See Why SP, SB and SD Numbers Don't Add Up. Related - Filters, Manage Columns and Exporting — the controls every table shares. - The Change Log — exportable, and the one export worth keeping before a big change. - What Is MCP, and Why It Matters — the live alternative.

MCP Call Logs: Seeing What Your AI Tools Asked

When you connect an AI tool to your TrackIQ data, something is querying your business on your behalf — and it is worth being able to see what it asked. Call Logs records every MCP request, and it records the question, not just the endpoint. That distinction is what makes it an audit trail rather than a debug log. Watch the walkthrough Prefer to watch? Seventy-six seconds on why recording the question, not the endpoint, makes it an audit trail. Where it lives Settings → MCP & API Keys → Call Logs. The call log, with its own date range and export It is a full table with the same controls as every other table in TrackIQ — its own date range, Search, Filters, Manage Columns and Export. | Column | Records | |---|---| | Question | the natural-language question the AI asked | | Tool Calls | which TrackIQ tools it invoked to answer | | Posted At | when | Why the Question column matters Most API logs record an endpoint and some parameters. That tells you what was fetched and nothing about why. Here the log captures the question in the words it was asked, which changes what the log is good for: You can tell intent from access. "Which campaigns spent over $100 with no sales last week?" is a recognisable piece of work. A row that reads nothing like the work your team does is worth asking about. You can debug an answer. If an AI gave you a figure that looked wrong, the log shows what it actually asked and which tools ran. Most of the time that is enough to see whether the question was imprecise or the data was. You can see how the tools were chained. The Tool Calls column shows how many TrackIQ tools one question needed. A question that fanned out across six tools is doing real synthesis; one that used a single tool is a lookup. If it is empty An empty log with "There is no data for the selected period" means one of three things, in order of likelihood: Nothing has queried yet. The log populates the first time a connected client actually asks something — connecting alone does not create entries. The date range is wrong. It has its own range control, independent of the rest of the app. Nothing is connected. Check the connection itself. See Installing TrackIQ MCP in Claude, ChatGPT and Cursor. What to look for Volume that does not match the work. A quiet week with hundreds of calls is worth understanding. Usually it is an agent looping rather than anything sinister, but you want to know either way. Questions from a brand nobody is working on. Per-brand API keys mean you can tell which integration is pulling what. An unfamiliar brand appearing is a scoping problem. Repeated identical questions. Somebody is asking the same thing daily by hand. That is a candidate for a report template or an alert instead. See Business Report Templates and the AI Report Generator. Questions that returned nothing. Often a sign the asker expected data TrackIQ does not hold — worth correcting before somebody builds a decision on an empty answer. Using it as an audit trail Two habits worth having if MCP is used across a team: Review it monthly, briefly. You are looking for pattern changes, not reading every row. Export it before an access review. The Export button gives you the period as a file, which is the artefact an auditor or a security questionnaire actually wants. The log answers "who queried what, and when" for AI access specifically — a question most tools cannot answer at all. What it does not tell you It does not record the answer. Only the question and the tools used. If you need the answer, you need the AI client's own history. It does not distinguish people from agents beyond which key was used. A key shared across a team logs as one source. It is read access only. Nothing in the log can have changed a bid, a budget or a rule — MCP cannot write. If something changed, the Change Log and a rule's Run History are where to look. It is not a permission system. The log tells you what happened; it does not restrict what can happen. That is what per-brand keys are for. Related - The MCP Web Console — the screen this sits on. - Installing TrackIQ MCP in Claude, ChatGPT and Cursor — connecting a client. - Feed TrackIQ Data into Your Own AI Workflow — what generates these entries. - The Change Log: Auditing What a Rule Changed — the audit trail for writes.

The MCP Web Console

Settings → MCP & API Keys is where you connect AI tools to your TrackIQ data, and where you check what they have been asking. The page is deliberately lopsided. Almost everyone needs only the first card. Everything below it is labelled Advanced because it genuinely is. Watch the walkthrough Prefer to watch? Ninety-nine seconds on the one card that matters, and why there is no secret to store. The whole page in one screen The MCP access console Three things live here: The guided connect card — one URL, four steps, no credentials. Advanced — manual OAuth clients and per-brand API keys, for clients that cannot self-register. Call Logs — top right. What the connected tools actually asked. See MCP Call Logs: Seeing What Your AI Tools Asked. The one URL https://app.trackiq.com/mcp There is a Copy button beside it. That single URL is the entire configuration for any client that supports OAuth Dynamic Client Registration — Claude Desktop being the common case. The four numbered steps on the card are the whole procedure: 1. Open the custom connector dialog. In Claude Desktop: Settings → Connectors → Add custom connector. 2. Paste the URL. The client registers itself with TrackIQ. Nothing to copy back. 3. Pick a brand and approve. A TrackIQ tab opens in your browser. You choose which brand the connector may query, then approve. 4. Ask a question. "List my marketplaces" is the standard smoke test. Step 3 is the security boundary. The connector is scoped to the brand you pick at that moment, in your browser, on a TrackIQ page. Nothing you paste into the AI tool grants access — approval happens on TrackIQ's side. Why there are no credentials to handle Dynamic Client Registration means the client asks TrackIQ for its own identity, and TrackIQ issues one. No client ID to copy, no secret to store in a config file, no token that leaks if someone screenshots your terminal. That is the reason to prefer the top card over anything in Advanced whenever your tool supports it. Advanced: when the top card is not enough Manual OAuth clients and per-brand API keys The section carries its own warning, and it is accurate: Most customers don't need anything below. Two mechanisms sit here, for two different kinds of client. Manual OAuth clients Pre-issued client_id / client_secret pairs, for OAuth-aware clients that do not yet support Dynamic Client Registration. The brand is still chosen at connect time on the consent screen — a manual client does not pre-bind a brand. Use for: Cursor (current versions), in-house OAuth-aware integrations, OAuth proxies. Do not use for: Claude Desktop. Paste the URL at the top of the page instead. + New Client creates one. Per-brand API keys Long-lived bearer tokens scoped to a single brand, sent as Authorization: Bearer … on every request. Use for: curl, Postman, claude mcp add via the CLI, the older mcp-remote bridge, server-to-server scripts — anything that does not speak OAuth at all. Keys are listed per brand, each with a live count: 0 active · 0 revoked. + New Key generates one for that brand. The scoping is the point. A key cannot reach a brand it was not issued for, so a script that only needs one brand should hold a key for only that brand. Choosing between the three | Your client | Use | |---|---| | Claude Desktop | the URL at the top | | Cursor, OAuth-aware but no DCR | manual OAuth client | | curl, Postman, a script, a cron job | per-brand API key | | Claude Code CLI | API key via claude mcp add | There is also a Setup instructions for Cursor, Claude Code, & direct API button in the Advanced panel, which expands the client-specific detail. See Installing TrackIQ MCP in Claude, ChatGPT and Cursor. Operating it safely Issue the narrowest thing that works. An API key for one brand beats an OAuth client that can reach any brand you approve. Revoke rather than delete-and-forget. The active/revoked counts on each brand card are there so you can see what is outstanding. Check Call Logs after connecting a new client. It shows the questions asked and which TrackIQ tools each one used, which is the quickest confirmation that a connection is working and being used as expected. Nothing here can change your advertising. MCP is read access. A key cannot move a bid, a budget or a rule — so the risk from a leaked key is disclosure, not damage. That is still worth avoiding. Related - What Is MCP, and Why It Matters — the concept. - Installing TrackIQ MCP in Claude, ChatGPT and Cursor — per-client setup. - MCP Call Logs: Seeing What Your AI Tools Asked — the audit trail. - MCP Tiers: What Each Plan Unlocks — which tools your plan exposes. - Feed TrackIQ Data into Your Own AI Workflow — what it is actually good for.