When you connect an agent — Claude, Codex, Cursor, dip itself — to your tenant, it requests a set of scopes: named capabilities, not roles. A scope is a ceiling on what a kind of client may ever do, never a grant of new access. Holding query_engines:execute_sql never lets a connection run a statement your own roles wouldn’t already let you run yourself. When you connect to https://mcp.<plane-domain>, sign in or sign up at the shared account page that opens. The consent page selects where the connection will act: The workspace address is the short name at the start of your tenant’s URL, such as acme in acme.app.<domain>. The proposed address comes from your email domain, or your email’s local part for personal providers such as Gmail and Outlook. You can edit it before approving: the card shows the resulting URL as you type, and an address it cannot use (anything other than lowercase letters, numbers and hyphens, starting and ending with a letter or number) shows an error under the field and keeps Approve disabled. If the address is taken, correct it and retry; retrying the same approval does not create another tenant. Creating a tenant usually takes under a minute. While it is being set up, the consent card shows Setting up your workspace <slug>… and keeps the page open on its own: it sends the same approval again until the tenant is ready, then returns you to your app. You don’t need to select Approve again. If setup takes longer than five minutes, the card shows “Your workspace is being created. Try again shortly.” and Approve works again; selecting it resumes the same tenant. A consent page stays usable for 10 minutes after it opens. If it has expired, start the connection again from your app. If you join a tenant after opening a consent page that offers to create one, approval asks you to restart the connection. Start again from the app that requested access, then select your tenant. If a tenant for your email domain already exists, the page asks you to contact its administrator for an invitation. You can explicitly choose to create a separate tenant with a numbered slug. An email domain alone never grants membership, and opening the consent page creates nothing. The connection acts in exactly the tenant you selected or created. To connect to another one, start a new connection and choose that tenant. Redeem authorization codes and refresh tokens at the token endpoint of the authorization server that issued them. Grants issued at the shared account address use https://app.<domain>/oauth/token. Grants issued at your tenant’s address use https://<slug>.app.<domain>/oauth/token, including grants for the shared MCP address. Using a different authorization server’s token endpoint returns invalid_grant. Approving a connection shows a card naming the client and your tenant, where the approval goes, and the list of what it’s asking for:
Client name wants to access your tenant. Unverified app Dippr has not verified this app’s name. Approving sends access to claude.ai. It will act as you, with your permissions. Data access is still checked per query.
Any client can register itself and choose its own name, so the name is always marked Unverified app: an app that calls itself “Claude” is not necessarily Claude. What the platform does enforce is the address under it. The approval is only ever delivered to that host, so check it before you approve: claude.ai or chatgpt.com for those assistants, or this computer (a 127.0.0.1 address) for dip and other local tools you just started yourself. If the address is one you don’t recognise, select Cancel. The address is shown so that a lookalike reads differently. It is always shown in full, so read it to the end: claude.ai.evil.example is an evil.example address, not Claude’s. A port the address names is always shown. An address written with characters from another alphabet, such as a Cyrillic “а” in place of a Latin “a”, is shown in the xn--… form your browser uses, with a warning under it. Under that, the screen leads with a few plain-language capabilities. A capability is a short headline, such as “Read your data”, that groups the scopes a client asked for. The default connection shows five: A client that asks for more, such as dip login, also shows headlines like “Connect and change data sources”, “Create and reconfigure query engines”, “Publish data apps” and “Install data apps and change their access”. A publish-only request does not include installation consent.
The consent card marks the client Unverified app, names the address the approval goes to above the permissions, and lists Read your data, Write your own sandbox databases, Draft definitions, Invite people and Stay connected, with a See the exact permissions link below

The consent screen for a new MCP connection: its capabilities, with the exact permissions behind a link.

Select See the exact permissions to see every scope under its capability: the scope’s name and what it means, drawn straight from the same registry that enforces it, so what you’re shown can’t drift from what a token actually carries. Every scope the client asked for is listed there; the headlines never hide one.
The expanded list shows query_engines:execute_sql, query_engines:list and metadata:read under Read your data, workspace_db:write under Write your own sandbox databases, semantic:suggest under Draft definitions, members:invite under Invite people and offline_access under Stay connected

See the exact permissions, expanded: each scope's name and full description under its capability.

The scopes you can see there: Only the scopes a client actually asked for appear on your screen — a client that requests nothing beyond browsing SQL never shows you a row about publishing apps. A new MCP connection that names no scopes of its own asks for the default consent: run SQL, see your engines, search metadata and query history, suggest metrics, write your own workspace databases, see your pipelines and their runs, invite people, publish your pages as read-only apps, and stay connected. Inviting is part of the default so a tenant admin can invite from the chat; for everyone else the permission does nothing, because only an administrator’s invitations go through. See What a new connection asks for by default. Approve grants exactly the list in front of you; Cancel grants nothing. Each consent screen can be answered once. If you click Approve twice, or reload and resubmit after answering, the second attempt says the request is no longer available and grants nothing more. Start the connection again from the app that asked for it if you need a new one.
If the redirect after approving points somewhere other than an address on your own computer, the screen says so and names the host. Only continue if you started this connection yourself.

Effective access is three things intersected

Holding a scope is necessary, never sufficient. What a connection can actually do is
  • Scope narrows by capability. A connection without query_engines:write can’t create an engine, no matter what your roles say.
  • Your roles and grants narrow by what you personally may do. A connection holding query_engines:write still can’t create an engine for someone who holds no create grant on engines — see Roles and permissions matrix.
  • The endpoint narrows by which door the connection authorized against. A token minted for your tenant’s MCP endpoint can’t be used against an engine’s own endpoint, and vice versa — see Addresses and hostnames.
Two connections holding identical scopes can still see different data, because the roles behind them differ. Never read a wide scope as wide access — check the person or service user behind it.

Scopes are fixed at the moment you approve

A connection’s scopes don’t change after you approve them, even if your tenant later turns on a capability the client would now like to use. If you connected an agent before a feature was enabled, or before your roles changed, ask it to authenticate again so the new authorization can request the wider set. The consent screen then names the permissions the connection did not have before. A connection can gain only permissions in the default set this way. See Reconnecting after a capability was added.

Revoking a connection

Settings → Connected apps lists every connection you’ve approved and revokes any of them. The connection’s own access stops within about a minute; a query-engine, app or registry sign-in it already obtained can last up to 15 minutes (see that page). Removing the server from your agent (claude mcp remove, dip mcp uninstall, deleting the entry from Cursor’s config) only stops your agent using it. Even without a revoke, the connection’s access ends on its own once its offline_access grant runs out, at most 30 days after you approved it. Revoking a service user’s alt_ key is different and does live in the console: go to Settings → Service users, find the key, and revoke it. Every connection using that key is refused within a few seconds on the data plane; the console checks on every request.

Management scopes: why dip login asks for more, and you never see it on this screen

dip login requests a wider set than any MCP connection ever does: the default consent every connection gets, plus the scopes its own commands need:
To approve less, pass dip login --scope "<names>" with exactly the scopes you want. You approve these capabilities for your tenant’s management address. Registry, engine and MCP connections authorize their own addresses; a token for one cannot substitute for another. dip requests publication and installation by name. Another public OAuth client can request them and ask for your consent too, but an ordinary MCP connection does not request them by default. Installation also requires your current permission and ownership, plus a separate confirmation of the selected app’s access and code trust. See Install an app. Headless dip pipeline commands can use service-user API keys with pipelines:read for listing pipelines and run results, pipelines:manage for operating them, or pipelines:publish for publishing pipeline images. A hosted image push also needs apps:publish to sign in to the registry. dip pipeline push and register also work on your own dip login: it asks you once, in the browser, for pipelines:publish. The reserved pipelines:report scope is for an external reporting flow that is not available yet. A local dip pipeline dev run needs no API key: signed in with dip login, it connects to the engine as you, with your own access (the query_engines:execute_sql capability above). With $DIPPR_API_KEY set, it connects as that key’s service user instead, which is how CI runs it. These control-plane capabilities do not appear when you approve an MCP connection. See OAuth scopes for their limits. A saved sign-in from before pipelines:read joined the default (or one narrowed with dip login --scope) lacks it. dip pipeline list, runs and events say so before calling the workspace, and the fix is one plain action: sign in again with the command the message prints, dip login followed by your workspace address. dip cannot tell such a sign-in whether the workspace has pipelines at all, because that check needs the same permission, and the message says so. A workspace whose deployment does not include pipelines answers pipelines are not available in this workspace instead, and signing in again does not change that. A service-user API key without the capability gets its own message naming pipelines:read as the one to grant. A client registered to request documents:write by name can ask you to approve Add documents to Documents. This lets it upload files as you into folders you may add to. Files are parsed and indexed and remain private unless you separately approve automatic sharing or manually share them in the console. Existing approvals and refreshed credentials do not silently gain sharing authority. A separate documents:share_uploads approval, Share new document uploads with folder readers, lets new uploads follow the sharing preference on tenants using Documents access to manage folder permissions. When that preference is on, eligible readers are selected at admission and checked again before publication. When off, new uploads remain private. On the support-managed Documents setup, new agent uploads stay private to you while you retain folder and classification access, even with sharing approved. The upload returns a note explaining that automatic sharing is unavailable. You can manually share the file in Documents using Share… from its row menu. Existing files are unchanged. dip doc upload asks for both approvals, including renewed approval for older logins. Older clients with only documents:write keep uploader-only uploads. See agent upload sharing. Documents must be enabled for your tenant. These permissions grant neither reads nor ACL changes and are separate from the scopes a standard MCP connection requests. API keys cannot hold them.

OAuth scopes reference

Every scope in one table, who can hold it, and how implication works.

Connect an AI agent

Where this consent screen shows up in each client’s setup flow.

Routine access

For routine investigations, register your client requesting agents:run and routines:read explicitly. They appear on the consent screen when requested; they are not part of the default MCP scope set. Reconnecting a client that does not request them will not add them. agents:run starts manual investigations, proposes pending routines, pauses future runs, and dismisses or restores your own never-activated routine drafts. Restoring a draft still requires fresh console confirmation before it can run. routines:read reads metadata and results within your current access. These are separate permissions. Neither approves a write a routine or task proposes. That is agents:review. It is never part of the default or discovered consent and never carried into an exchanged token, so a connection holds it only if the client explicitly requested it and you approved that line on the consent screen. Otherwise, approve in the console. routines:write is reserved and does not authorize confirmation. Confirmation requires the owner’s recent console sign-in and task update permission. A data plane cannot acquire this scope through token exchange. Saving from chat without explicit start creates a pending proposal. Explicit personal starts can begin without console confirmation. Confirm in a fresh console session. A scope alone never authorizes another person’s access.