Connect and manage a connector

Connecting an account makes it available to projects in the workspace by default. Remove it from a project or agent's connector settings when that work should not use it.

1. Choose the external account

Before opening d5s, decide which account should authorize the integration. Prefer a dedicated service or team account when continuity matters, and grant it only the provider permissions required for the intended work.

d5s cannot use the connector to reach objects the external account itself cannot access. Conversely, broad provider access is not automatically narrowed to one folder or repository unless the provider authorization or connector offers that boundary.

2. Connect in the workspace

Open Connectors to browse providers by what they do, such as communication, customer work, data, or development. Each category initially shows six providers; choose More… to open its shareable full list. Search covers the full catalog from the overview, or filters the selected category when you are viewing one. Popular is a separate starting point, so a provider can appear there and in its functional category. Select a provider to review its detail page. You can share the detail page's URL with a teammate before connecting. Choose Connect, then an available connection method when the provider offers more than one. Depending on the provider, setup uses OAuth sign-in, an access token, a host address with the credential, or the provider's MCP sign-in. After connecting, the same detail page shows the installed connection.

Some providers present multiple methods under one name. Compare each method's access summary and choose based on the work the project must perform, not only on setup convenience. When methods identify different data regions, choose the region where the external account is hosted.

Complete the provider flow and give the connection a recognizable display name when the form allows it. The credential should appear as active. A Needs reauthorization state means its external authorization must be repaired before new runs can rely on it.

For a new Google connection, read the notice about how Google content may be processed by AI providers, then choose Acknowledge and connect to continue to Google authorization.

During reconnect, the provider may reuse an existing signed-in browser session and return to d5s without showing another login page. The successful return confirms the workspace credential only. Check the intended project or agent under Connectors before relying on it in a run.

On a connected provider's detail page, choose one of the three example requests to open the workspace agent picker. It shows agents whose projects can use a connected method for that provider; attach the connection to an existing agent if it is missing. Select an agent, or choose Add an agent for quick creation. d5s opens Chat with that agent and the example as an editable draft; review it before sending. The example does not grant new access to the provider.

To connect the same provider again, return to the directory and choose Add another. Use names that make the account or environment clear, such as Notion — Sales and Notion — Product. Each connection remains independently selectable, rotatable, auditable, and removable. OAuth providers can reject a second connection when it resolves to the same external identity; in that case, reconnect the existing credential instead.

Only workspace roles with connector-management permission can create or archive shared credentials.

Connect Checkly

Select Checkly in workspace Connectors, then choose Connect. Paste a current Checkly User or service API key and give the connection a name that identifies the account. You do not need to enter a server URL. This connection uses Checkly's official MCP server.

User keys start with cu_ and use the accounts and permissions of that user. Service keys start with sv_, belong to one account, and use the role chosen when the key is created. Prefer a service key for shared or scheduled monitoring; Checkly requires a Starter, Team, or Enterprise plan for service keys. Deprecated account keys and older sk_ keys are not accepted by Checkly. OAuth sign-in is not offered for this connection.

Review the available tools before letting an agent use the connection. Allow the check, result, and incident reads it needs. Keep check execution, incident updates, monitor changes, and environment-variable access on Ask or Deny unless you intend to grant them. A Checkly role still limits what the key can do.

Test from the intended agent: "Use Checkly to show the accounts I can access, then show failing checks for our production account." Name the account in later requests if a user key can access several accounts. Checkly can provide check results and available failure assets; missing evidence is not proof that a check passed.

An existing Custom MCP server connection remains available. Attach and test the named Checkly connection before removing an older one. If a request fails for authorization, verify the key's account, role, and revocation status. Rotate or revoke the key in Checkly when its access is no longer needed.

Connect Datadog

Select Datadog, choose Connect, then select your actual Datadog site: US1, US3, US5, EU, AP1, AP2, or UK1. Your Datadog site is independent of your d5s region. Check the site in your Datadog account before choosing. Government sites are not offered.

Paste a Personal or service access token and name the connection for its account and environment. Personal tokens start with ddpat_; service tokens start with ddsat_. Prefer a service account for shared or scheduled work. This form accepts one token; an API/application-key pair is not supported. OAuth sign-in is not offered for this connection.

Ask your Datadog administrator to enable MCP Access for the organization. Use a scoped token and an account with MCP Read plus the resource permissions needed for monitors, metrics, logs, or traces. Data access restrictions and network allowlists can further limit access. See Datadog MCP setup and Datadog access tokens.

Review the available tools. Keep monitor changes, muting, and other writes on Ask or Deny. Test from the intended agent by asking for alerting production monitors and related logs. Missing telemetry can mean the account or token lacks access; verify permissions before treating it as healthy.

Personal access tokens expire. Create a replacement before expiry and use Rotate secret on the connection. Revoke old or unused tokens in Datadog. Existing custom connections remain available; attach and test the named connection before removing an older one.

Connect PagerDuty

Select PagerDuty, choose Connect, then API token (US) or API token (EU) for the service region of your PagerDuty account. Paste a user or account REST API token and give the connection a recognizable name. You do not need a server URL. Events API routing keys are not REST credentials and do not work in this form.

Both user and account keys are supported. Some user and team filters require a user key. PagerDuty requires an account with Advanced Permissions for its MCP server; individual tools also depend on provider roles and resource access. Choose the least access needed for your work. See PagerDuty MCP setup and API access keys.

Allow incident context, on-call, and change-event reads as needed. Keep incident updates, responder requests, schedule changes, and other writes on Ask or Deny. Test from the intended agent by asking for open incidents, the current on-call responder, and change events around a deployment.

New OAuth connections are not offered. Existing OAuth and custom connections remain available. Attach and test the named token connection before removing an older one. If access fails, check the selected site, key validity, account, and provider role. Rotate the saved secret when replacing a key, then revoke the old key in PagerDuty. Removing a d5s connection does not revoke its provider credential.

Connect HubSpot

Choose HubSpot and authorize the intended account once for curated CRM actions such as search, record updates, archive, company merge, and association changes. Lead association removal is intentionally unavailable because removing a lead's final primary association can delete the lead in HubSpot.

Review HubSpot's consent screen and the CRM scope set it grants. The app requests a broad set of optional CRM scopes; a portal without a particular HubSpot product may omit the corresponding scope. d5s action policies and per-write approval narrow what agents may do after consent. Reconnect that exact credential after its HubSpot permissions change. A former HubSpot MCP connection is not converted; disconnect it, connect HubSpot again, and reattach the new connection to any project or agent that used the old one.

Connect Jira Cloud

Choose Jira, then sign in to Atlassian with the account whose projects the agent should use. Review the requested access and complete consent. Check that the Jira connection is active and available to the intended project or agent. Jira Cloud uses OAuth sign-in; there is no server URL to enter.

Name the connected Atlassian account in your request. Ask the agent to list its accessible Jira sites before making changes. When there is one site, it is selected automatically for issue work. When several sites are accessible, choose the intended site explicitly. After you choose a site, the agent works only on that site. Choosing a site does not add Jira access that the connected account lacks.

Jira actions use the same Tool permissions and conversation approval settings described below. Reading starts on Allow; creating comments, changing issue status, creating or updating issues, changing issue links, and uploading attachments start on Ask. Approve a change after reviewing the account, site, target issue or project, and intended content. A Jira permission error requires repairing the external account's access, not another d5s approval.

For an attachment upload, identify a file already available in the current conversation's files and the exact Jira issue. Each file must be no larger than 50 MiB. Attachment reads use the same size limit. The agent can open the full attachment (text, CSV, PDF, DOCX, XLSX, image or WAV) from the saved result without repeating the Jira action.

Start with a connection-info read, then read a known issue on the selected site. For a write check, use a test issue, keep Ask for approval enabled, and read the issue back after approving. If an upload or issue change times out or its outcome is uncertain, inspect Jira before trying again: the change may already have succeeded.

Select a social publishing destination

For X, completing OAuth selects that X account automatically. Give the connection a name that makes the public account clear, especially when the workspace connects more than one account.

For LinkedIn Pages, connect with a LinkedIn member who administers the intended company Page. Open the installed LinkedIn connection, find Publishing destinations, and select the Pages agents may draft for. Page selection is a workspace-admin action and can be changed later. If no Page appears, confirm the member's current Page role in LinkedIn, then reconnect and try again.

Activating the credential makes its selected destinations available to that project. It does not authorize publication: every social draft still requires a separate exact-revision approval in its originating conversation.

For Custom MCP server, enter the full HTTPS endpoint and choose OAuth sign-in or a bearer token. OAuth works when the server advertises compatible authorization metadata and supports Client ID Metadata Documents or public dynamic client registration. A server that requires a separately registered OAuth application may need a supported catalog integration or a provider-issued bearer token. Reconnect uses the existing connection's saved URL; add another connection to use a different endpoint.

Connect a Custom API with a workspace skill

For Custom API, enter a name and the trusted host. Choose API token to use the existing header and token setup. Choose Workspace secrets to link existing secrets by name. Use Link secret to select each secret. Values remain hidden. Saving does not contact the provider.

Add SDK scripts, API documentation and usage instructions through workspace Skills. Make the skill and connection available to the same project or agent. Skills do not grant connector access. SDK files are not uploaded in the connector form.

If you need to store a username or password, choose Password or login value in workspace Secrets. Keep values out of skills and conversations. Use Linked secrets in the connection menu to change a connection's references.

Every request through a workspace-secret connection needs Approve once, including reads and login. Review the host, methods, paths, filters and request bodies. Allow all does not remove this requirement. Changed secret links, secret versions or access require a new approval.

These connections support HTTPS request plans with JSON or form bodies and query filters. Login cookies are private to one approved plan. Only its final response is returned. Redirects are not followed. Multipart uploads, binary downloads and extracting a login token from a JSON response are not supported by this request action. An SDK that needs these features requires additional integration work.

Known linked credential and session-cookie values in returned text are replaced with [REDACTED]. The result indicates when redaction occurred. Treat redacted fields as unavailable; do not use them in totals. If a result cannot be returned safely, it stays withheld.

After approval, the agent can import the complete redacted response as a file and run SDK scripts from its workspace skill. Import uses the saved response and does not send the provider request again. If import fails, retry the import instead of repeating the provider operation. If the response format changes, ask the agent to test an SDK working copy against saved responses and propose a reviewed workspace skill update.

A saved connection does not prove that authentication or a workflow works. Start with an approved, narrow read and check its returned source. If a request has an unknown outcome, inspect the provider before making another request.

3. Review where it is available

New connections become available to projects that inherit workspace connectors, including ordinary chats and agents made through advanced creation. Agents made through conversational setup receive only the connections reviewed in their draft; manage later changes in agent settings. For a project, open Context → Connectors. For an agent, open Manage agent → Connectors.

Remove a connector there when that project, chat, or agent should not use it. That opt-out remains in place when the credential is repaired or reconnected. You can add it again later from the same view. The agent's Add connector control lists workspace connectors that were removed from that agent above the full provider catalog. Creating a new workspace credential still requires connector-management permission.

Adding a connection to a project does not reveal its secret. It lets that project request allowed connector actions under the account's current policy.

The first active connection for a provider becomes its project default. Later connections are not default unless you explicitly replace it. If the default is deactivated while another connection remains, select a new default before asking the agent to use that provider without naming an account.

Choose the intended account when asking an agent to use a connected service. Accounts from the same provider remain separate, and an unavailable account never causes a silent switch to another connection. Approvals and activity records stay associated with the selected connection.

If two connection methods conflict on the same project, d5s keeps the existing method rather than making authorization ambiguous. Remove it before adding the other method to that project.

4. Review tool policy

Open a connected account's Tool permissions dialog when available. The dialog has a shareable URL, shows the account's available operations and descriptions, and lets a workspace admin Add account for the same provider. Under Approval, choose Always ask, Allow reading (the default), or Allow all. Allow all requires confirmation because known write operations can run without another prompt; unclassified operations still ask. Choosing another Approval option replaces individual operation settings after confirmation. You can also open the dialog with Manage permissions on the provider page, or from an agent's Manage → Connectors tab.

Use the menu on an operation to set its individual choice:

• Allow — run without pausing;
• Ask first — require a matching approval in the current session;
• Block — block the operation; or
• Use default — follow the Approval choice again.

Sending email and booking or cancelling meetings on Gmail, Outlook, and Google Calendar always wait for an approval card, so those operations offer only Card or Block. A meeting change that notifies guests also waits for a card. On MCP connectors, tools that send email or book meetings show Block only.

For MCP connectors, review actions individually. Actions without a reviewed or explicit policy require approval, even when they appear to be reads. API operation controls appear only for connections that support them.

Operations are grouped as Reading information, Making changes, and Unclassified. The Approval choice applies to current and newly discovered operations without an individual setting. An individual setting takes precedence over the Approval choice, and Reset on a group clears its individual settings. Unclassified operations require approval unless you set an explicit policy for them.

Use Refresh tools when a provider changes its actions. If refresh fails, the previous list remains available.

Each saved permission change is recorded in the organization audit log. Organization Owners and Admins can expand the event in Settings → Audit log to see who changed it, when, and the previous and new setting of each affected operation.

Where operation controls are available, start with write-capable actions set to Ask first. If an API connector does not expose them, use a read-only or narrowly permissioned external account and keep writes interactive. Approve the exact action only after checking its target, arguments, and expected side effect.

The chat composer controls how Ask actions behave for that session. Ask for approval pauses and shows an approval card. Allow all lets Ask actions continue without individual prompts until the session is switched back, except email sends and meeting bookings, which still wait for a card. It does not override Deny, connector permissions, or workspace access restrictions.

Gmail attachments

You can ask the Gmail connector to send attachments totaling up to 25 MiB per email. Review the recipients, text, and attachment details before approving the email. The complete email must also fit Gmail's message limit. If it is too large, reduce the files or share an accessible file link. Your recipient's email service can have a lower limit.

Gmail email style

A workspace admin can set how agents format Gmail drafts and emails from a connected Gmail account. Open the account's settings dialog and choose Email style. Other members see the font, size, signature choice and a preview, but cannot change them.

• Font and Size apply to the text of each draft and email. Default leaves the choice to the reader's mail app.
• Signature set to Gmail uses the signature saved in that account's Gmail settings. If that account has no Gmail signature, emails have none.
• Signature set to Custom uses a signature you paste, for example from Apple Mail. It can be up to 10,000 bytes. Links and https images stay. When you save, other formatting can be removed, and the dialog shows Saved. Some formatting was removed. If nothing in the signature can be kept, the dialog stays on Custom and asks you to paste it again.

Choose Save. Drafts and emails that agents write next use the new style; existing drafts do not change. When an email waits for your approval, its card shows the Signature below the text. If you edit the text, the email keeps its font and signature.

5. Test from the intended project

Run a small read request from the project or agent that will use the connector. Verify the selected source and inspect the tool record. Then test any required write in Ask mode before enabling a schedule. If you later use Allow all, confirm that the session's standing autonomy is appropriate for every manual and autonomous turn that can reuse it.

If the connector is unavailable, check the boundaries in order: public availability, credential state, whether it was removed from this project, tool policy or live approval, and the external account's own permissions.

For HubSpot, start with an account-details read and a record read. Test a create or update on disposable test data in Ask for approval mode. Exercise archive, company merge, or association removal only in a non-production HubSpot portal: confirm both record IDs and the merge primary, approve once, then read the surviving records and associations back before attempting another write.

For a social connector, ask the agent to prepare a short test draft for the named account or Page. Confirm that the preview shows the intended destination, copy, and image before approving. After publication, open the reported provider link. If delivery is uncertain, inspect the external account before creating or retrying any equivalent post.

Azure DevOps repository access

Connect with the Microsoft work account that has access to your Azure DevOps Services organization. Your company administrator may need to approve the connection. On-premises Azure DevOps Server is not supported by this connection.

Give the agent your organization name or URL and the repository you want it to inspect. It can browse directories, read source files and search code within the connected account's permissions. For example: “In organization Contoso, read the API authentication code in repository Platform and explain how it validates requests. Include the file paths and branch you used.”

This connection is read-only. It does not push code, run pipelines or grant access to Azure infrastructure. It also does not narrow a broadly authorized account to the repository named in your prompt. Use an appropriately scoped external account and review which d5s projects share the connection.

Microsoft 365 SharePoint and OneDrive access

Connect Microsoft 365 with the Microsoft work or school account that has access to the SharePoint sites and OneDrive files you want the agent to read. On work accounts your company administrator may need to approve the connection: reading SharePoint sites requires a permission that many tenants reserve for admin consent. Until it is granted, the sign-in ends on Microsoft's approval page and the connection is not created.

Give the agent the site, library or document you want it to inspect. It can search sites and files, browse document libraries and lists, and read list items, site pages, file metadata and file contents within the connected account's permissions. File downloads support work or school accounts hosted on commercial SharePoint. Personal OneDrive and sovereign-cloud file downloads are not supported. For example: “Read the policy PDF in the HR site's Policies library and summarize its main requirements.”

This connection is read-only. It does not upload, edit or delete files, and it does not change list items or sharing. It also does not narrow a broadly authorized account to the site named in your prompt. Use an appropriately scoped external account and review which d5s projects share the connection. Mail and calendar stay on the separate Outlook connection.

Use connector results in analysis

The agent can use a saved connector result for analysis. When an MCP tool includes an embedded PDF, PNG, JPEG, GIF, WebP image, or WAV audio file (up to eight files per result), the agent can present the original file in its reply, including in a supported chat channel. Preparing the file does not call the provider again.

For supported MCP download actions, d5s can turn a returned PDF link into a file during the same connector call. Moneybird invoice PDF downloads are supported. d5s imports the PDF only when the provider returns it and it is a valid PDF within the size limit. The agent can then attach the imported file without a separate website-access approval. The original connector action still follows its tool-permission setting.

Microsoft 365 can also download supported SharePoint and OneDrive files through the connected account. Other URLs and redirects remain references. For a read-only download or export that is not supported for direct import, the agent can fetch the file in its sandbox. If the URL's host is blocked, the agent can request access to that exact host for the chat. The card needs your approval; keep This website selected for the requested host. Approval grants network access, not the connector's credentials. Temporary links may expire during approval, so after approval the agent may need to call the read-only download action again for a fresh link. d5s never repeats a write action to get a fresh link. Some provider links require authentication beyond a signed URL and cannot be fetched this way. See external API access for the scope of a host grant.

Opening a saved result does not repeat the provider action. A provider action can succeed even when saving its output fails; check the reported outcome before attempting another action.

Remove access

Remove a credential from the project when only one project, chat, or agent should stop using it. Disconnect the workspace credential only when nothing in the workspace should use that connection. The confirmation warns that projects and agents using the connector will lose access immediately.

Disconnecting stops d5s access across the workspace; it does not change data or permissions at the external provider. After reconnecting, check each intended project or agent. Explicit connections must be added again, and a project that previously removed the connection stays opted out.

Google connections and data already retrieved

Google content may be processed by the AI providers used by your workspace to complete the work you request. Read the notice before choosing Acknowledge and connect.

Disconnecting removes the saved credentials; it does not erase earlier tool results, conversations, files or project memory. Deleting a conversation does not delete project memory derived from it. You can remove the provider grant in your Google Account settings and request erasure of stored data through privacy@d5s.tech.

Connect an IMAP / SMTP mailbox

A workspace admin can choose IMAP / SMTP in the web connector directory. Use the web app to add or replace the mailbox login. The mobile app can manage the installed connection and its tool permissions. Enter a connection name, mailbox address, login username or account ID, and password. The login can differ from the email address. When SMTP sending is enabled, the mailbox address and outgoing recipient addresses must use ASCII characters. Message text and attachment names can use other characters.

Enter the IMAP host, port, and folder. IMAP requires TLS. Use the settings from your mail provider. Configured hosts must be public mail servers.

Leave Drafts folder empty for automatic detection, or enter its exact server name. If detection finds multiple folders or no folder, reading stays available. Save draft is available after a folder is detected or configured. Set an exact folder before saving drafts. Connecting checks folder access without creating a message; it does not prove write access. Replacing credentials keeps automatic detection unless you enter an exact folder.

Enable sending with SMTP is off for new connections. Reading and saving drafts need only IMAP. To send mail, enable SMTP, enter its host and port, and select TLS or STARTTLS. Replacing an existing SMTP connection keeps sending enabled unless you turn it off.

Choose Connect mailbox. d5s checks IMAP and any enabled SMTP login and sender without reading messages or sending a test email. If the check fails, check the server settings and enter the password again. Other form values remain available.

The connection follows the workspace's sharing and agent access settings. Open Tool permissions to review its tools. Search, read message, and read attachment default to Allow. Save draft supports Ask first, Allow, and Block. Send email is available only when SMTP is configured. It always requires Ask first, including chats set to Allow all. Choose Block to prevent a tool from running.

An authorized agent can read mail on demand. You can also add a scheduled mailbox check to start work only when new messages arrive. Connecting alone starts no checks. Reading does not mark a message as read. Search covers the selected folder and can return a continuation for more results.

For a daily review, specify the first included date and the first excluded date. For example, 20 October through before 21 October covers 20 October. Dates use the mailbox's received-date calendar, which can differ from the sender's date. The agent can continue the same search across result pages. An empty page does not prove the search is finished. If more pages remain or a message cannot be read, treat the result as partial. A complete search covers only its selected folder, dates and search text; it does not establish that every relevant case document was found.

Save draft creates an unsent message in the selected mailbox's Drafts folder. With Ask first, review the From, To, CC, BCC, subject, text, and named attachments, then choose Save. With Allow, the agent can save drafts without asking. Open your mail app to edit and send a saved draft. Saving never sends through SMTP.

For a reply draft, ask the agent to use the original message from this connection. The draft stays in that email conversation where the mail provider supports it. Existing drafts that a person edits are not overwritten.

Sending creates a separate review card. Check every recipient, subject, text, and attachment before approval. The card keeps the attachment contents proposed with it. A send proposal does not save a mailbox draft. An approved send uses the connected address. Replacing the connection requires a new proposal for any waiting email.

You can attach one conversation file of up to 5 MiB to each email. A larger file or a second file returns an error before approval. The complete encoded email must fit within 10 MiB. This limit is checked before approval and after changes to the subject or text. Retrieved messages have a 100 MiB limit. Larger messages return an explicit size error, and search coverage stays partial. A large message can occupy its own search page. Long body previews are truncated; the original message remains available. If another large mailbox read is active, retry after the busy response. Original messages and retrieved attachments remain available with the conversation's retained tool results.

If a draft save shows Unknown, check the Drafts folder before asking for another save. d5s checks the existing operation instead of blindly saving another copy. If the mailbox refused a draft and no Retry is available, check the connection and ask for a new draft.

A successful send means the SMTP server accepted the message. It does not confirm delivery to a recipient. If the result is uncertain, check the mailbox before sending again. d5s does not automatically retry an uncertain send.

A downloaded nested email can have different MIME formatting. Use the retained original message when exact bytes matter.

Certified-email headers and receipt attachments are retained as evidence. d5s does not verify their certificate, signature, delivery, or legal effect.

To update settings, choose Replace credentials for the connection and enter both login fields again. A failed replacement keeps the current connection. One-time and session read approvals expire after a successful replacement. Disconnecting removes its access from agents.

Query PostgreSQL and create figures

If PostgreSQL appears in your connector list, follow Connect PostgreSQL for agent queries. The guide covers verified database TLS, on-demand SSH tunnels, network allowlists, and statistics-only monitoring accounts.