AI assistant access: ask Claude about your own data

What this screen is

Every other screen in Newsletrix is a fixed view of your collected newsletters. This one removes the view. It publishes a single address that an AI assistant connects to over MCP, the open standard these products use to reach outside data, and from then on the assistant can list, search and read your tracked newsletters, your brand aggregates, your topics and your recommendations while you ask questions in plain language.

The practical effect is that questions the app has no dedicated screen for become answerable. Newsletrix ships no "summarise every newsletter I received this week, grouped by sender, with a one-sentence takeaway each" report. A connected assistant will write you one, off the same data the app draws on.

Everything issued here is read-only. That is not a setting you can change. Each credential carries one fixed permission, described in the app as read your newsletters and analysis data, and there is no wider permission to grant. Nothing an assistant can do here will write, delete, or touch your billing, your tracking addresses or your account settings. Access is scoped to the account that owns the credential, so there is no path from your token to anyone else's data.

Two connection styles exist. Claude Desktop and claude.ai use a sign-in handshake, so you paste the server address, approve on a Newsletrix consent page, and never see a token at all. ChatGPT, Perplexity, Gemini and anything scripted use a personal access token you create here and paste into their connector settings. Every plan gets this, the free tier included.

Open it when

• You want an assistant to answer from your real tracked data instead of guessing, for example a Monday digest of everything that arrived last week.
• You are setting up a new laptop or a new assistant and need a fresh credential for it.
• A teammate left, a laptop was lost, or a token leaked, and access has to stop today.
• You are wiring an internal script or bot and need a client identifier and secret rather than a pasted token.
• Your assistant says it cannot reach the connector and you want to re-copy the address or check whether the credential has ever been used.
• You want to keep an assistant restricted to one client or market by naming a workspace in the question.

Screen anatomy

The tab is one column inside the Settings page. The shared Settings header and usage cards sit at the top, the tab strip with AI assistants selected sits under them, and below that come five stacked cards. Numbers on the mock match the walk-through underneath it.

1. Connect an AI assistant The entry point. It holds the read-only server address, a badge saying the same address accepts both a sign-in handshake and a pasted token, a Copy URL button and a Test button. This one string is all most assistants ask for.
2. Active token counter Reads "2 / 3 active" in the card subtitle. Three active personal access tokens is a hard ceiling. At the limit the New token button greys out and an amber notice appears, and you have to revoke one to free a slot.
3. Last used column The single most useful column in the table. It moves every time a credential connects successfully, so "Never" on a token you configured hours ago means the assistant has not reached the server once.
4. OAuth client credentials For server-to-server or scripted integrations that complete the full sign-in handshake themselves. Up to three at a time. The amber "Any callback allowed" badge marks a client with no pinned return address.
5. Set up your AI assistant Four tiles, one per supported product, each with a badge saying which connection style it uses and four numbered steps naming the exact menu path inside that product.
6. Example questions Five tabs of ready-made ones. Click a question to copy it, or use the small Claude button on the row to open it directly in a new Claude conversation.
7. Settings header and tab strip Shared page chrome. The heading and the line "Manage your account, security, billing, and integrations." sit above every Settings tab, and AI assistants is the seventh tab, marked with a violet plug icon.
8. Usage ring and tracking address The two cards above the tab strip belong to Settings as a whole. They matter here indirectly: the weekly analysis allowance on your plan is the ceiling on what any assistant will ever have to read.
9. Credentials gate Before you have created anything, the token and client cards are hidden behind a single card asking whether you need credentials for ChatGPT, Perplexity, Gemini or scripted clients. Pressing Create credentials reveals both, and they stay revealed the next time you open the tab.
10. One-time reveal banners An amber panel appears after you create or regenerate a token, and an emerald one after you create or rotate a client secret. Both state that the value will not be shown again, and both are correct.
11. Empty states With nothing created, the token area reads "No tokens yet. Click 'New token' to create one." and the client area reads "No OAuth clients yet. Most users won't need one - personal access tokens above are simpler." If every token is revoked or expired the table says so and offers a Show them link.
12. Consent page Not part of this tab. When Claude sends you to approve access, Newsletrix shows a separate page headed "Authorize Access" naming the requesting app, the email address you are signed in as and the permissions requested, with Deny and Approve buttons.

Every panel explained

MCP server URL

A read-only field holding the one address your assistant connects to, badged OAuth + Bearer, with a Copy URL button and a Test button beside it. Underneath sits a collapsed section holding a second address, rarely needed, that some assistants read to work out the sign-in steps for themselves.

How it is calculated. The address is the same for every account and ends in /mcp. Nothing about it is personal: what identifies you is the credential, never the URL. The badge means the address accepts either a credential issued through the sign-in handshake or a token pasted as an authorisation header.

How to read it. Copy it verbatim, including the path at the end. If an assistant asks for a "connector URL", a "custom MCP server" or a "remote server", this is the value.

Personal access tokens

The card where you mint credentials by hand, for any assistant that wants a pasted token. The subtitle carries the live count against the ceiling, and the New token button opens a short form: a name and an expiry.

How it is calculated. You may hold three active tokens at once, where active means not revoked and either without an expiry or with one still in the future. Names are free text, up to 100 characters. The expiry list offers 30, 90 and 180 days, 1 year and 5 years, with 1 year selected by default. The value appears once, at the moment you press Create token, and cannot be retrieved afterwards.

How to read it. Name the token after the device or product that will hold it, because that name is the only thing distinguishing rows later. Match the expiry to the purpose: a short window for a test or a contractor, a year for your own laptop.

The one-time reveal

Immediately after creation, an amber banner shows the value twice: once raw, and once already formatted as a complete authorisation header for the products that want the whole line pasted in.

How it is calculated. The header field is simply the word Bearer, a space, and the token. The "Expires in" line is the lifetime you chose, rounded to whole days.

How to read it. Paste the raw value where a client asks for a token or a key, and the second field where it asks for a complete header line. Dismissing the banner or reloading the page loses the value for good.

The token table

Every credential ever issued for your account, newest first. The columns are name, source, scope, created, last used, expires, status and actions.

How it is calculated. Source reads "Personal access token" for anything you created by hand, and otherwise names the app that obtained it through the sign-in handshake. A token you revoked shows Revoked, whatever its expiry date says. A token past its expiry date shows Expired. Everything else shows Active. Scope always shows the same single read-only permission. Last used gives the time of the most recent successful connection, in UTC, and reads "Never" until the first one. Revoked and expired rows are hidden by default, with a count of what is hidden printed under the table.

How to read it. Work from the Last used column outwards. Never plus an Active badge is a configuration problem in the assistant. A stale date plus an Active badge is a credential nobody uses, and a good candidate for revoking. A recent date plus a complaining assistant means authentication works and the question is about the data.

OAuth client credentials

A second, more technical route: a client identifier and secret pair for a script or internal bot that completes the sign-in handshake itself rather than carrying a pasted token. Most people never need one, and the card says so.

How it is calculated. You may register three clients. Each carries a name of up to 255 characters and up to ten return addresses, each at most 500 characters. Preset chips fill in the known callback addresses for Claude, ChatGPT, Perplexity and Gemini, described in the app as best-effort values that may shift if a vendor changes their flow. The secret is shown once and cannot be retrieved afterwards, the same as a token. Return addresses match exactly, apart from a single asterisk wildcard, so a partial prefix will not match.

How to read it. Pin the return addresses whenever you can. The alternative, the "Allow any URI (advanced)" switch, leaves the client accepting any callback, which the table then flags in amber as "Any callback allowed".

Set up your AI assistant

One tile per supported product, badged with the connection style it uses: a sign-in handshake for Claude Desktop and claude.ai, a pasted header for ChatGPT, Perplexity and Gemini.

How it is calculated. The tiles hold written instructions and nothing more; no tile connects anything on your behalf. Each carries four numbered steps naming the exact menu path in that vendor's settings. The button copies the server address and opens the vendor's connector page in a new tab, and on the pasted-header tiles it also opens the create-token form with a suggested name already filled in.

How to read it. Follow only the tile that matches your assistant. The steps and links describe third-party products, so they can drift when a vendor redesigns its settings; if a menu name has moved, the sequence still tells you what to look for.

Example questions

Twenty-four ready-made questions, sorted into five tabs. The tabs are Inbox digest, Compare senders, Subjects and CTAs, Topics and trends, and Metrics and quality. Clicking a question copies it. The small Claude button on the row opens it in a fresh Claude conversation.

How it is calculated. The library is fixed and hand-written. None of it comes from your data, so it does not change as your account fills up. Until you have created a credential, an information box above the tabs points out that the examples assume a connected assistant.

How to read it. Treat the tab names as a map of what the connection can actually answer well. A few of the questions ask for things the assistant has to work out rather than read directly, such as reading time or the number of links in an email, and those answers may be approximate or unavailable.

What the assistant can ask for

No panel shows this, but it is what sits behind the connection. An assistant can make ten kinds of request. Every one is read-only, and every one is answered from your account alone.

How it is calculated. Every request answers for the account that owns the credential and for no other. Most also accept a workspace name, which must match one of your group labels exactly and restricts results to newsletters delivered to addresses carrying it. Only the screenshot comes back as a picture. The other nine come back as text your assistant reads and rewrites for you.

How to read it. The list is the honest ceiling on what an assistant can tell you. If a question needs something outside it, the assistant will either derive it from the creative or the plain text, or it will not answer, and both outcomes are better than a confident guess.

Metric definitions

Filters and controls

Some things are deliberately missing, so you can stop hunting for them. There is no export button, and no log of what an assistant has asked for beyond the Last used timestamp. You cannot narrow a credential to some of the ten requests instead of all ten, and you cannot look up a token value after the moment it was created. Nothing here is held back for a paid plan either. The tab is either configured or it is not.

How to use it

1. Copy the server address first It is the same for every assistant and every device you own. Keep it on the clipboard while you work through the vendor's settings.
2. Pick your route from the tiles Claude needs no token. Paste the address, approve on the consent page, and you are connected. ChatGPT, Perplexity, Gemini and anything scripted do need a token, so start from their tile and let it pre-fill the create form.
3. Name the token after the device, and set the expiry deliberately "ChatGPT on work laptop" beats "token 2" three months later. Shorten the lifetime for anything temporary.
4. Paste it, then run one question from the library Use an Inbox digest example rather than inventing a question, so a blank answer tells you something about the connection rather than your phrasing.
5. Come back and confirm Last used has changed This is the whole verification step. Once the column shows a timestamp instead of Never, authentication is proven and any remaining problem is about data or phrasing.
6. Prune the table when something changes A laptop replaced, a contractor finished, an assistant you stopped using: revoke the row. Three slots fill quickly, and Revoke all active tokens is there for the bad day.

Limits and caveats

• Read-only is absolute and also a ceiling. An assistant can summarise and compare, but it cannot add a tracking address, delete a newsletter, change a setting or start an analysis for you.
• Three active tokens and three clients per account. The counter above the table also counts credentials Claude obtained by signing in, while the ceiling applies only to the tokens you created by hand, so the tab can report the limit before you have really reached it.
• Token and secret values are shown exactly once. Nothing in the app, and nobody at Newsletrix, can recover a lost value after that.
• Every score the assistant returns for open rate, click rate, compliance or effectiveness is an AI estimate made from the content of the email. Newsletrix sees competitor newsletters as an ordinary subscriber and never sees another sender's real opens or clicks.
• Brand and account averages are unrounded, have no minimum sample size, and skip blanks rather than counting them. An average can therefore rest on far fewer newsletters than the send count beside it suggests, and the effectiveness score in particular is often missing.
• The account-wide total counts every newsletter you have collected, including ones that have not been analysed, while the averages next to it only cover analysed ones. The two numbers answer slightly different questions.
• Search matches subject lines and brand names only, not body text, and returns at most 50 sends at a time. A very long piece of raw markup comes back shortened, and the assistant is told that it was. A screenshot comes back only for a send that already has one.

FAQ

Related screens