Connectors
Give agents access to services your workspace already uses, such as Neon, Supabase, PostHog or Sentry, through each provider's own MCP server.
A connector gives Sinatra's agents a service your workspace already uses: your database, your analytics, your error tracker. Once a service is connected, every agent run in the workspace gets that provider's own hosted MCP server as a set of tools, so the agent can check a schema, look up an error, or read the numbers behind a bug report instead of guessing.
Sinatra tells agents to use connected services for reading only and to treat what they return as untrusted data. That is an instruction, not a limit. Sinatra does not enforce read-only access. What an agent can do through a connection depends on the token you paste or what you approve at the provider. Each connection is labelled with how far its read-only access goes, and you accept a warning before connecting.
Supported services
| Service | How you connect | What limits the agent | Reference |
|---|---|---|---|
| Neon | Sign in, or paste an API key | Sign-in: read-only on one project, enforced by Neon. Token: read-only requested, not guaranteed. | Project ID, required |
| Supabase | Sign in, or paste an access token | Read-only requested, not guaranteed, either way. | Project ref, required |
| PostHog | Sign in, or paste a personal API key | Sign-in: read-only, enforced by PostHog. Token: the key's own permissions. | Project ID, optional |
| Sentry | Sign in, or paste a personal token | The skills you leave ticked, or the token's scopes. | Organization slug, optional |
| PlanetScale | Sign in, or paste a service token | The tools and databases you tick, or the token's permissions. | None |
| Cloudflare | Sign in, or paste an API token | The access you approve, or the token's permissions. | None |
| Notion | Sign in, or paste an MCP token | Your full Notion permissions. Notion has no read-only option. | None |
| Vercel | Sign in only | Your Vercel access in the teams you share. Vercel has no read-only option. | None |
Sign in sends you to the provider to approve access, and the provider sends you back. Paste a token takes a token you create at the provider. The connect dialog links to the page where each token is created and says which permissions to give it.
Access labels
Every connection shows one of three labels in the dashboard, and the API returns it as grantLabel:
| Label | grantLabel | Meaning |
|---|---|---|
| Read-only, enforced by the provider | enforced | The provider refuses changes for this connection. Only Neon's and PostHog's sign-ins get this label. |
| Read-only not guaranteed | not_guaranteed | Sinatra asks for read-only access with a flag in the MCP address, and an agent can remove that flag. Neon tokens and Supabase connections get this label. |
| May be able to make changes | token_decides | Sinatra adds no limit. Whatever the token or sign-in allows, an agent can do. |
What to know at each provider
- Neon. After you approve read-only access to one project, Neon shows a second screen asking for broad access to your account. That is how Neon's MCP server issues its tokens. Sinatra keeps the token encrypted and sends it only to Neon's MCP server, which limits the connection to read-only on that project. For a pasted key, an organization admin can create a project-scoped key. Neon recommends connecting agents to development and testing data, not production.
- Supabase. Signing in asks for database read and write access, because Supabase's SQL tool needs it, plus read access to your organizations, projects and logs. Scope a pasted token to the one project.
- PostHog. Signing in asks only for PostHog's read scopes, and PostHog refuses any change an agent tries. You can narrow it to some projects on PostHog's page; the project ID in Sinatra only picks the default, and agents reach every project you approve. PostHog shows Sinatra as an unverified application. For a pasted key, use the Read-only access preset limited to one project, not the MCP Server preset, which can write to every project. EU accounts create keys on eu.posthog.com.
- Sentry. Sentry's sign-in ticks every skill by default, including ones that change issues, projects and alerts. Untick what agents shouldn't use. Sentry ends the sign-in 30 days after you approve it. A pasted token must be a personal token with read scopes only (
org:read,project:read,team:read,event:read). Organization tokens don't work, and organizations in Sentry's EU region haven't been tested. - PlanetScale. On PlanetScale's page you tick which tools the connection may use and which organizations and databases they apply to. Reading data needs Run read-only queries, which is off until you tick it. Leave Run write queries and the billing tools unticked. For a service token, grant
read_organization,read_databaseandread_branch, plusconnect_production_read_only_branchif agents should run read queries. Never grantconnect_production_branch. Paste the token value that starts withpscale_tkn_, not the token ID. - Cloudflare. Signing in preselects read access to your user, accounts, zones, DNS, Workers and their storage, logs and analytics. Anything you add on Cloudflare's page, write access included, agents can use. Cloudflare ends the sign-in 30 days after you approve it. For a pasted token, create an API token with only Read permissions, for only the accounts and zones you want; an account token also needs Account Resources: Read. Sinatra refuses your Global API Key, and tokens with client IP filtering don't work.
- Notion. Agents act with your full Notion permissions. On Notion's Enterprise plan a workspace owner may first have to allow Sinatra. Notion ends the sign-in 180 days after you approve it. Pasted MCP tokens need a Business or Enterprise plan; a workspace owner creates one in Notion's Developer portal.
- Vercel. You choose the teams to share. Vercel says connecting grants the same access as your Vercel account.
Who can connect
Only a workspace owner or admin can connect or disconnect a service. Every member can see the workspace's connections, and every task started by a workspace member uses them.
A connection keeps working only while the person who made it is an owner or admin of the workspace. If they leave or become a member, Sinatra disconnects the connections they made. It doesn't revoke them at the provider, so revoke those tokens or sign-ins there too.
A workspace can have up to 20 connections, and more than one to the same service, for example one per project. Sign-ins that haven't finished count toward the 20.
Connect from the dashboard
- Open Settings → Connectors. Under Available, select Connect next to the service. For a service that's already connected, the button reads Connect another.
- For a service that offers both, choose Sign in or Paste a token.
- Fill in the provider's reference if the service takes one, such as Neon's Project ID. The help under the field says where to find it.
- Optionally give the connection a Name, up to 64 characters, such as
production. It defaults to the reference, or todefaultwithout one. Agents see the connection as an MCP server namedsinatra-<service>-<name>, shown under each connection. - For a token, paste it into Token. It's sent once and never shown again.
- Read Before you connect, then tick each confirmation (see below).
- Select Sign in to the service, or Connect the service. A sign-in takes you to the provider and back to Settings → Connectors, which says whether it worked.
Agents can use the connection from their next run.
The warning and confirmations
Before connecting, the dialog shows what the connection means, and you tick each confirmation separately. The text depends on the service, how you connect and the access label.
For a connection labelled Read-only not guaranteed or May be able to make changes, a warning comes first. Its title is Read-only is requested, not guaranteed, This token decides what agents can do, or, for a sign-in, This sign-in gives agents your service access. It lists what an agent could do if the token or sign-in allows it. For Supabase, for example, that's running SQL that changes or deletes data, applying migrations and deploying Edge Functions, and reading every row, your logs and your project's keys.
The text that follows says:
- Agents can read from the service on every task in the workspace, for every member.
- What they read is sent to the AI model that runs the task, and can include stored tokens and keys.
- Content the agent reads can influence what it does.
- Granting only the access you intend is your responsibility.
Then you confirm:
- "I understand agents can use this on every task in this workspace, for every member, and that what they read is sent to the AI model."
- Unless the provider enforces read-only, that an agent may change or delete things. For a token: "I understand an agent can do anything this token allows in" the service, "including changing or deleting things."
- That you are responsible for what the token can access, or what you approve at the provider, and for anything agents do with it.
Neon's and PostHog's sign-ins have no warning and only the first and last confirmations. Sinatra records who accepted, when, and which version of the text. If the text changes while the dialog is open, the connect is refused and you review it again.
Connect from the CLI, the API or MCP
None of these takes a token as a command-line argument or a tool argument, and none completes a sign-in. Signing in and accepting the warning happen in an owner's or admin's browser.
CLI.
sinatra connectors # this workspace's connections, and what can be connected
sinatra connectors connect neon --project <project-id> # a dashboard link to sign in
sinatra connectors connect sentry --name prod --token # paste a token at a hidden prompt
sinatra connectors disconnect <connection-id>For a service you can sign in to, connect prints a link to the dashboard's connect dialog, filled in with --project and --name. At a terminal it also opens the link and waits up to 10 minutes for your new connection to appear; with --json or outside a terminal it only prints the link. To paste a token instead, pass --token: the CLI shows the warning and confirmations, asks you to accept, then reads the token at a hidden prompt or from stdin. --accept accepts without the prompt, for scripts that pipe the token in. See CLI, MCP and API.
API. GET /v1/workspaces/{id}/connectors lists the catalog and connections; any member can call it. Each catalog entry carries its acknowledgement text, its confirmations and a connectUrl to the dashboard dialog. POST to the same path connects a pasted token, with the token in the request body and the confirmations you accepted, exactly as the catalog gave them. Sign-in isn't available over the API: send an owner or admin the entry's connectUrl. See Connector endpoints.
MCP. list_connectors reads the workspace's connections. connect_connector returns a dashboard link that opens the service's connect dialog, filled in with the project and name you give it; an owner or admin opens it, signs in or pastes a token there, and accepts the warning. disconnect_connector removes a connection. No MCP tool accepts a token.
Connect links in agent replies
When a task would go better with a service that isn't connected, the agent can end its reply with a link to that service's connect dialog. It's told to do so only when reading the service would change the result: it finishes what it can, says what it couldn't check, and offers each service at most once in a conversation. Automated fixes, such as reworking a failing CI check, don't offer links.
A link opens the dialog for an owner or admin. A member who opens it is told that connecting needs an owner or admin, and can share the page with one.
After you connect from a task's link, the dashboard takes you back to the task. For a task started from the CLI, MCP or API, the dialog has a box to continue that task after connecting, which names the task and is ticked by default. Leave it ticked and Sinatra posts a message on the task as you, saying the service is now connected, and the agent picks the task back up. For a task started anywhere else, reply on the task to continue.
When a sign-in stops working
Sinatra refreshes a sign-in's access when an agent run starts and every five minutes while it runs. Nothing refreshes it on a schedule while no agent is using it.
A sign-in needs to be done again when the provider ends it, such as Sentry and Cloudflare after 30 days or Notion after 180, or when refreshes keep failing at the provider for about two hours or more. The connection then shows Needs sign-in again, and agents don't get it. An owner or admin selects Sign in again on the connection, which opens the dialog filled in with its project and name. The agent may also offer a link for it, marked as needing to be signed in again. sinatra connectors prints the command that signs it in again, and the API and list_connectors return its status as needs_reconnect.
A pasted token isn't refreshed. If it expires or you revoke it at the provider, the agent's calls fail until you disconnect it and connect a new token.
Disconnect
An owner or admin selects Disconnect on the connection. Agents lose access on their next call, and Sinatra deletes its copy of the token. What's left at the provider depends on how you connected:
- A pasted token still works at the provider until you revoke it there. Sinatra can't revoke it for you.
- A sign-in is revoked at the provider. If that fails, the dashboard says so: remove Sinatra in your account settings at the provider.
- Cloudflare, Sentry and Vercel keep their MCP server listed as an authorized application after a sign-in is disconnected, even when the revocation worked. Remove it yourself:
- Cloudflare: Cloudflare MCP Server, under My Profile, then Access Management, then Connected Applications in the Cloudflare dashboard.
- Sentry: mcp.sentry.dev, under User settings, then Account, then Authorized Apps.
- Vercel: Vercel MCP, under Account Settings, then Sign in with Vercel in the Vercel dashboard.
How the token is protected
The token never enters the sandbox. Sinatra stores it encrypted with a key bound to your workspace. The agent's MCP configuration holds only a placeholder. When the agent calls the provider, Sinatra's egress proxy swaps the placeholder for the real token, and only:
- on a TLS request to that provider's MCP host, such as
mcp.neon.tech, never any other host; - in the request's
Authorizationheader, never in a body or URL; - while an agent run that was given the connection holds it. A run releases it when it ends; one that fails before releasing it holds it until the sandbox's next run ends, the sandbox is removed, or 24 hours pass.
A placeholder sent anywhere else isn't swapped. Responses from the API, the CLI and MCP never include a token or any part of one.