Skip to content
Help Center
Submit a request Sign in

Connect AI clients with MCP access

Connect an AI client to Supportbench under a deliberately selected agent identity. The wizard creates the connection credential and gives client-specific setup instructions; the client can then use only the operations and records allowed by that agent’s effective access. Use a read-only operational task to verify the connection before relying on it.

Applies to: Supportbench Core. Location: Settings → API → MCP & AI Tokens. Requires Administer API configurations.

Before you begin

  • Identify the client, the business task and the person responsible for maintaining the connection. Read the MCP operations guide if Model Context Protocol or AI tools are new to you.
  • Review the intended Supportbench agent’s role, Division and customer access. Choose an identity whose permissions fit the task; do not use an unrestricted administrator merely because setup is easier.
  • Plan one permitted read and one approved excluded-record access check. Keep tokens and any OAuth secret in protected client settings, separate from shared setup notes.
  • Review existing tokens before creating another connection. A prior unfinished wizard may already have created a credential.

Configure and verify

  1. Open Settings → API → MCP & AI Tokens and choose New Token. The wizard stages are Choose AI, Choose agent, Connect and Verify.
  2. In Choose AI, select the client you actually intend to use. The available cards include ChatGPT, Codex, Claude and Copilot Studio; follow the selected card rather than mixing instructions from different clients.
  3. Choose the Supportbench agent whose identity the connection must represent. Recheck its role and Division access before proceeding. The binding cannot be edited after creation.
  4. Choose Create token and continue. This creates saved access before the rest of the client setup is complete; closing the wizard does not revoke it.
  5. In Connect, follow the selected client’s displayed instructions. Copy only the values that card requests into the corresponding client fields. In particular, the current ChatGPT card uses automatic registration and leaves manual client ID and secret blank.
  6. Choose I completed setup after completing the client-side connection. In the AI client, make a read-only request for a known permitted case or article and inspect the returned source. Verify the expected identity and record scope.
  7. Check an excluded record using the approved access test. Do not treat a successful read as proof of permission for every write or every customer record.
  8. Choose Done after verifying the actual client request. Record the client, owner, selected agent and purpose without the secret. Done closes the setup flow; it is not a network or permission test.

Settings reference

Connection setup

SettingWhat it controlsHow to choose
Provider / ClientSetup cards for ChatGPT, Codex, Claude or Copilot Studio where available.Choose the client you actually use and follow its displayed connection method.
AgentActive agent bound to the access; binding cannot be changed after creation.Use the appropriate role and division scope.
Token or OAuth setupProvider-specific credential or authorization instructions.Copy each value into the field named by the client; an environment-variable-name field expects a variable name, not a token.
Connection and audit detailsIdentifies configured access and available audit evidence.Track the client and owner without recording the secret in shared text.
RevokeRevokes the selected access.Verify replacement access before revoking a connection still in use.
Created / ExpiresShows credential lifecycle details. The wizard currently creates a ten-year expiry and has no custom expiry input.Review actual displayed dates and maintain ownership/rotation records.

Behavior to know

  • There is no user-entered expiry in this wizard. Core currently creates a ten-year expiry; inspect the displayed Expires value and maintain an explicit connection review process.
  • Provider interfaces and authorization methods can change. Use the current selected setup card and the focused client guide, and check official client instructions when that provider’s setup changes.
  • A token is not an account-wide permission bypass. Tool availability, tenant scope, role permissions and record visibility still apply. Client behavior also differs; do not promise that every client asks before every mutation.
  • Revoke affects future access for every connection sharing the token. Replacing one client’s settings does not automatically migrate other consumers.

Recommended use

  • Begin with a narrow, read-only B2B workflow, such as reading one permitted account’s open-case context and citing the source. Introduce write operations only when the operator understands their impact and authority.
  • For normal rotation, identify every consumer, connect and verify the replacement, then revoke the superseded token. For a suspected exposed credential, follow the incident process rather than preserving access at the expense of containment.
  • If creation returns an unclear result, inspect the list before retrying. Keep an owner and purpose for each token so unfinished or duplicate setup is detectable.
  • Separate four checkpoints: credential created, provider authorization completed, expected tools available, and permitted request successfully executed. These are different observations.

Worked example

  1. A support lead connects an AI client for a specialist who works on a limited Division. They review that agent’s access and create the token for that identity.
  2. The specialist follows the selected client card, then asks for a known permitted case and reads the returned source details.
  3. An approved excluded-Division test remains unavailable. The team records the connection owner and purpose, without copying the secret into the case or article.
  4. Only after these checks does the team design additional authorized workflows. A successful read is not treated as approval for unattended updates.

Troubleshooting

ProblemWhat to check
The wizard was closed before completionInspect the token list. Create token and continue already saves access; either finish the intended client setup or deliberately revoke unused access.
Done was selected but the client cannot read SupportbenchCheck provider authorization, server/connection values, available tools and the selected agent separately. Done is not a connectivity test.
The client can read some records but not othersCompare the selected agent’s actual role, Division and record access. Do not broaden the identity merely to make an excluded example succeed.
The wrong agent was selectedThe binding cannot be changed. Prepare a correctly scoped replacement, verify it and revoke the obsolete connection as appropriate.
I cannot find an expiry fieldThis wizard does not accept a custom expiry. Review the displayed Expires value and use your credential review/rotation process.
A shared token was revokedEvery consumer using it can lose future access. Reconnect affected consumers with authorized replacement access and verify each one.
A credential may have been exposedUse the organization’s incident process. Keep secrets out of diagnostics and shared messages; describe the connection by owner, purpose and non-secret identifiers.
  • API → Create and manage API tokens
  • Organization → Configure agent roles and division access
  • Knowledge Base → Configure preferred terminology for AI and search

Was this article helpful?