Connect Supportbench to AI: a practical MCP setup and operations guide

Use Supportbench MCP when you want an approved AI assistant to look up real support records and help you work with their context. For example, a support lead can prepare an account review across several contacts, or an agent can assemble a specialist handoff from a long case history. This guide starts with the connection itself, explains the unfamiliar terms, and takes you through a first verified result before moving on to daily operations. You are connecting to Supportbench’s existing service; you do not need to build an MCP server.

What MCP does—and what you need to know first  ·  Choose the right starting point  ·  Before you connect: agree identity, access, and a test record  ·  Create the connection in Supportbench  ·  Connect Codex in the ChatGPT desktop app: no terminal required  ·  How do I open a terminal and run a CLI command?  ·  Connect with Codex CLI, step by step  ·  Connect from ChatGPT in a browser  ·  Connect Claude using the remote connector  ·  Connect a Microsoft Copilot Studio agent  ·  First test: read one case and verify the answer  ·  Worked workflow: prepare a weekly B2B account review  ·  Worked workflow: hand a complex case to a specialist  ·  Worked workflow: turn a repeated issue into reusable knowledge  ·  When Co-Pilot or the connected AI cannot find an answer  ·  Move from reading to changes deliberately  ·  Troubleshoot setup and daily use  ·  Maintain a dependable team connection

What MCP does—and what you need to know first

MCP means Model Context Protocol. It is a connection standard: an AI application asks a named Supportbench tool to do a specific job, receives the result, then explains it to you. A tool might search for a company, read a case, or retrieve an article. Other tools can make changes, so a connection is more than an extra search box.

MCP does not automatically make every answer accurate. The AI can only use tools available to it and records permitted by the connected agent. A convincing paragraph without record links is not evidence that it read your customer’s current cases.

Choose the right starting point

A Supportbench agent and a customer contact are different identities. This setup binds the connection to an agent. Do not give a customer contact your agent token so they can experiment. A customer-facing knowledge article can explain the feature without granting the reader permission to configure it.

Before you connect: agree identity, access, and a test record

1. Ask your Supportbench administrator whether you may create an MCP connection. You need Administer API configurations and an editable MCP configuration page. If New Token is unavailable, ask your administrator. If the menu is absent, ask for help rather than borrowing a colleague’s login.
2. Choose the Supportbench agent whose work this connection represents. Its role, division, team, and record visibility determine the accessible support context. The binding cannot be changed later: another identity needs an intentional new connection.
3. Confirm the AI client and workspace are approved for the customer records you intend to use. Sign in to that client under your intended work identity.
4. Pick one existing case that this agent can open in Supportbench. Note its exact case ID, company, subject, and current owner. You will compare these values after connecting. A harmless training case is a good first choice.
5. Keep the Supportbench wizard and the AI client available side by side. You will copy the server URL and complete authorization, then return to the client to test.

Create the connection in Supportbench

1. Open Settings, expand API, and choose MCP AI Tokens.
2. Choose New Token. The Connect an AI client wizard opens.
3. At Choose AI, select the application you intend to use. For the Codex procedures below, select Codex.
4. At Choose agent, choose the intended Supportbench agent. Recheck the identity before selecting Create token and continue. This action creates a credential.
5. At Connect, keep the newly created agent token private and copy the MCP server URL from the client card. Copy the complete value; do not substitute the public documentation URL.
6. Complete one client procedure below. The URL tells the client where to connect; the agent token is used separately to authorize it.
7. Return to the wizard and select I completed setup. The Verify step is a checklist, not a connectivity test. Run the read-only test below in the actual AI client, then select Done.

If your team already has a working bearer-token or OAuth connection, retain it unless the administrator intends to replace it. Do not revoke a working token while another person or scheduled workflow still depends on it.

Connect Codex in the ChatGPT desktop app: no terminal required

Use this route for Codex in the ChatGPT desktop app when your installation exposes the MCP server settings described here. Current OpenAI desktop documentation uses these labels; if your managed installation does not offer them, use the supported CLI route with IT assistance or ask the workspace administrator about availability. See OpenAI’s MCP setup instructions.

1. In the ChatGPT desktop app, open Settings → MCP servers → Add server.
2. Name the connection supportbench. Choose Streamable HTTP and paste the wizard’s MCP server URL.
3. Save, then select Restart. If the server requires sign-in, select Authenticate.
4. On Supportbench’s authorization page, paste your own agent token into the requested token field and approve the connection. The token does not belong in the server URL or a chat message.
5. Return to the desktop app. Use /mcp in the composer to inspect the connection, then perform the read-only case test below.

Keep STDIO for servers that must be started as a local program; the Supportbench URL supplied by this wizard is a remote connection. If a form asks for a program to execute instead of a URL, check the selected connection type.

How do I open a terminal and run a CLI command?

A command is a line of text that tells a program what to do. The terminal displays a prompt, waits for your input, and shows the result after you press Enter. You do not need to understand programming to enter the commands below, but the named program must be installed first.

1. Click beside the terminal’s prompt. Copy one command line at a time, paste it, then press Enter.
2. Do not copy a leading prompt such as PS C:\…>, explanatory text, or Markdown backticks. The command examples here contain only the command.
3. Read the result before entering the next line. An error means that step has not completed successfully.
4. When a command opens a browser for authorization, complete that browser step, then return to the terminal.

Connect with Codex CLI, step by step

Install Codex CLI using the current official Codex installation instructions for your operating system, or ask IT to install it. Installing the desktop app does not establish that the codex command is available in the terminal. If a fresh terminal reports “not recognized” or “command not found,” finish CLI installation before continuing.

1. Inspect existing connections. In the terminal, enter:

codex mcp list

If an existing Supportbench connection already works, use it. If the command is not recognized, see troubleshooting below. The list establishes configuration; it does not prove a case can be retrieved.

2. Add the server. Copy the full server URL from the Supportbench wizard. In the following command, replace PASTE-THE-WIZARD-URL-HERE with that value and keep the quotation marks:

codex mcp add supportbench --url "PASTE-THE-WIZARD-URL-HERE"

The word supportbench is the local connection name. --url tells Codex that the next value is the server address. No agent token appears in this command. Do not run it with the placeholder still present. If the name already exists, inspect the existing connection instead of repeatedly adding another one.

3. Authorize the connection. Enter:

codex mcp login supportbench

The browser opens the Supportbench authorization flow. Paste the agent token from the wizard into the requested field there and approve. Do not fill a manual OAuth client-ID or client-secret field with that token. Return to the terminal and read the result. If it reports failure, the connection is not ready even if a browser page opened.

4. Start or return to Codex. If you are still at the ordinary terminal prompt, run codex and sign in to Codex if prompted. Inside the active Codex conversation, enter /mcp to inspect active servers. These are different sign-ins: your Codex account opens the AI application; Supportbench authorization grants the specific support connection.

5. Ask for a known record. Use the test in the next section. Official command reference: Codex MCP configuration and commands.

Connect from ChatGPT in a browser

Open Plugins → + → Add custom MCP server when available in your workspace. Workspace policy can control custom connections. The official connection instructions describe the current interface; the Supportbench wizard provides the values specific to your account.

1. In Supportbench, choose the ChatGPT card and obtain its server URL and your agent token.
2. In ChatGPT’s custom MCP connection form, enter the name Supportbench and the complete server URL.
3. Choose the OAuth automatic-registration route supported by the wizard. Leave manual client ID and secret empty. If the client says automatic registration is unsupported, stop and send the error to your administrator; the agent token is not a replacement OAuth client secret.
4. Choose the connection/sign-in action. On the Supportbench authorization page, enter your agent token and approve.
5. Return to ChatGPT, select the Supportbench connection for the conversation when required, and perform the read-only test below.

A connection configured only on your computer is not proof that ChatGPT in another browser or a hosted conversation can use it. Test from the actual application and workspace where the support team will work.

Connect Claude using the remote connector

First select Claude in the Supportbench wizard and obtain its URL and your agent token. The current official Claude instructions distinguish personal setup from organization-managed setup.

1. For an individual connection, open Customize → Connectors → + Add → Add custom connector. Enter a descriptive name and the Supportbench server URL, then choose Continue.
2. Review authentication and continue to the OAuth client choice. For Supportbench’s documented flow, select Register automatically, then add the connector.
3. On Supportbench’s authorization page, enter your own agent token and approve.
4. For Team/Enterprise, an authorized administrator first adds the remote server under Organization settings → Connectors → Add → Custom → Web. Members then connect using Customize → Connectors with their intended credentials.
5. Enable the connector for the conversation through + → Connectors, then run the known-case test below.

Claude’s remote connector reaches the server from Anthropic’s cloud, including when using Claude Desktop. Being able to reach a URL through your own VPN does not establish that Claude can reach it. Ask the connection administrator to investigate reachability rather than changing the server URL to an unrelated endpoint.

Connect a Microsoft Copilot Studio agent

This route configures a Copilot Studio agent; it is different from Supportbench’s built-in Your Co-Pilot. Have the Power Platform administrator confirm that the environment’s policies allow the connection. Microsoft’s standard-harness MCP instructions describe the current wizard.

1. Select the Microsoft client in Supportbench’s MCP wizard and obtain the server URL and intended agent token.
2. In Copilot Studio, open the target agent’s Tools → Add a tool → New tool → Model Context Protocol.
3. Enter Server name, a clear Server description, and the exact Server URL supplied by Supportbench.
4. Choose API key, location Header, and header name Authorization, then create the tool configuration.
5. Create the connection. Enter the credential value as the word Bearer, one space, and the agent token. Keep this value in the credential field; do not put it in the agent’s instructions or a prompt.
6. Select the new connection and choose Add to agent. In Preview, request a read-only lookup of the known test case and compare the result with Supportbench.

Use the supported Streamable transport. If environment data policies block the tool, ask the administrator to resolve the approved configuration. Test before distributing the agent to colleagues; its connection identity determines the support access being exercised.

First test: read one case and verify the answer

Open your chosen test case in Supportbench first. Replace the example ID below with that case’s actual ID. The example is deliberately read-only: it does not send a message, assign an agent, or close anything.

1. Check that the client actually invokes a Supportbench tool rather than answering from general knowledge. If the tool requires read approval, inspect its target and approve the intended read.
2. Compare the returned ID, subject, company, and owner with the open case. If the ID resolves to the wrong organization or a similarly named record, correct the scope before continuing.
3. Open the returned link and compare the latest relevant activity. A compact tool response may include only recent context; ask for earlier activities when the task requires the whole investigation.
4. If access is denied or the record is absent, check the bound agent and its permissions. Do not grant broader access simply to make a test pass.
5. After this succeeds, try a small company search and a knowledge search. A successful single-case read does not establish that every reporting or writing capability is available.

Record the client, workspace, bound agent, test case, test time, and outcome in your team’s connection handover. Keep credentials out of that record.

Worked workflow: prepare a weekly B2B account review

Imagine Northstar Manufacturing has several contacts, an integration investigation, and equipment cases handled by different specialists. An account review needs a single account perspective without losing the separate commitments on each case. The following is a recommended operating method, not an automatic built-in report.

1. Resolve the customer. Ask the AI to search companies for the exact name and show possible matches with IDs. Choose the intended company before asking for its cases. Similar names and subsidiaries can otherwise be mixed.
2. Define the scope. State open cases versus cases closed during a period, the dates, and what you need included. Ask it to report filters, pages retrieved, and any limits. “Everything important” is too vague for a reliable review.
3. Build an evidence table. Request one row per case with ID/link, customer impact as recorded, owner/queue, last customer update, next action, promised update if explicitly recorded, and missing information. Tell it to label an inferred action as a proposal.
4. Inspect exceptions. Open cases with unclear ownership, conflicting updates, or missing commitments. Confirm facts in the actual history before using the table with the customer.
5. Plan action. Ask for proposed internal follow-up notes or a meeting agenda. Review them before authorizing any save. Customer-facing language needs its own recipient and promise check.
6. Close the review loop. Record agreed actions in their cases using the appropriate task, assignment, or activity workflow. A meeting spreadsheet should not become the only place where the next commitment exists.

This is especially useful in B2B support because one organization can have several people reporting connected problems. The aim is to connect the evidence across the relationship while keeping each case’s owner and customer commitments clear.

Worked workflow: hand a complex case to a specialist

A helpful handoff explains what is known, what has already failed, and the decision needed. Asking only for “a case summary” often produces a chronology that leaves the specialist to rediscover the problem.

Review each section. If the version was never captured, keep it as an open question instead of filling it with a plausible value. If a promise appears only in your proposed plan, do not describe it as something the customer already agreed to. Use the case’s assignment, task, mention, or escalation workflow to establish responsibility after the draft is reviewed.

Prepare the customer update separately. The internal handoff can include technical uncertainty and coordination details that are not appropriate for the customer. A concise customer update should say what is confirmed, what happens next, and when the next update is expected.

Worked workflow: turn a repeated issue into reusable knowledge

1. Start with a resolved case whose outcome has been verified. Ask the AI to identify the reusable symptom, applicable environment, cause if established, steps, and expected result.
2. Search existing articles using the customer’s phrase, product terms, and the actual error. Ask for links and meaningful differences before proposing a new article.
3. Draft a correction to the canonical article when it already covers the task. Create a distinct article only when the new task or applicability genuinely differs.
4. Remove customer-specific logs, identities, serial numbers, credentials, and contract details from a public draft. Keep necessary private context in the restricted case.
5. Review the draft through your KCS process. For supported MCP knowledge writes, inspect and approve the exact normalized preview before applying the batch.
6. After publication, ask the original customer-style question again. Open the returned article and check whether its steps answer the question—including the next question a beginner is likely to ask.

An answer that says “run the command” is incomplete if the intended reader does not know what a terminal is. Improve the source procedure, title, summary and relevant search phrases together. Adding keywords without supplying the missing instructions will only make an inadequate answer easier to find.

When Co-Pilot or the connected AI cannot find an answer

A “no matches” message does not establish that the feature is unavailable. It can reflect wording, scope, permissions, the selected assistant, or missing documentation. First make the question self-contained, then test a known record or article to separate connection failure from a search problem.

For a useful report to the knowledge team, include the exact question, the answer/no-match message, the article that should have helped, the user role and selected search mode. Redact customer information and credentials. A documentation fix and an assistant’s retrieval/context fix are separate outcomes; retest both rather than assuming that publication repaired everything.

Move from reading to changes deliberately

Start new workflows with reading and drafting. When a write is useful, state the exact record and intended change. “Improve these cases” leaves too much open; “prepare an internal note on these two case IDs with this reviewed text” is a concrete instruction.

Keep Supportbench’s normal operational accountability. An AI-generated recommendation is not approval from the specialist, evidence that a customer agreed, or proof that a workflow ran. Inspect the saved result after an authorized action.

Troubleshoot setup and daily use

Maintain a dependable team connection

Assign an owner for each operational connection. Record its purpose, bound agent, client/workspace, expiry where configured, and the person who will test it after changes. A shared process should not silently depend on the personal account of someone who has changed roles.

• Review the bound agent’s access when duties, divisions or team membership change. Access changes affect what the connection can return.
• Use Revoke in MCP AI Tokens when the credential should no longer be used. Revocation blocks future requests; it does not erase previously retrieved information or saved activities.
• When replacing a connection, test the replacement with a known case and knowledge article, then retire the old credential through the approved process. Coordinate users and dependent work.
• Keep tokens, authorization codes and secrets out of screenshots, shared notes and prompts. Record non-secret identifiers and error references instead.
• Measure the workflow by evidence completeness, corrections required, and whether agreed actions reached the case record. Do not judge it only by how quickly the assistant produced text.

For configuration ownership and token management, continue with the linked administrative reference. For case, view and KCS practices, use the focused guides below.

Continue learning

• Configure and manage MCP agent access
• Work a case from first response to verified resolution
• Understand company and contact relationships
• Build operational views for account and workload reviews
• Use KCS to maintain trusted reusable knowledge