MCP Connections: Verify Tools, Access and Real Results

Connecting an AI assistant to a database or work tool should begin with a question you can verify. “Tell me everything about our business” is too broad for a first test. “List the open issues in this one project and cite each issue” gives you a useful way to check access, retrieval and accuracy.

Brainbaby has an MCP integration page and workspace MCP server settings. This guide explains how to evaluate a connection before relying on it for real work. It does not promise access to every service, a ready-made Academy library, or automatic setup of your database.

What MCP connects

The official architecture overview describes a host application, its clients and connected servers. Servers can expose tools, resources and reusable prompts. The protocol connects those components; it does not guarantee that a particular application supports every server or that a model has read your entire account.

A database connection, for example, may expose a narrow query tool rather than a full copy of the database. Its available operations and access rules determine what the assistant can actually retrieve. An answer without a tool result is not evidence that live records were checked.

Choose one useful first task

Start with a small project containing information you can inspect yourself. A fictional test project might contain three issues: fix a broken signup button, update a help article and review an invoice export. Ask for their titles and statuses, then compare the returned result with the source.

Keep the first task read-only. A successful list operation establishes something about retrieval; it does not establish permission to update records, send messages or run queries against production. Evaluate those operations separately when you need them.

For codebase questions, the GitHub integration may be a more direct starting point. Decide whether you need to inspect repository files, query a work service or build a new application before choosing a connection.

Add a server in the correct workspace

Open Settings and find MCP servers for the active workspace. Workspace owners and administrators can manage these connections; viewers and editors can inspect the list. Make sure you are in the workspace that should own the connection before entering its details.

The current form asks for a name, a server URL and an authentication method. Follow that server provider's current instructions. Do not assume that a normal website URL is an MCP endpoint, or that a personal API key can replace an OAuth flow the provider requires. Enter credentials only in the connection's authentication field, rather than posting them in a chat message.

Brainbaby's form currently supports no authentication, a bearer credential or a custom header. A provider requiring another authorization flow may need a different integration path. Check compatibility before creating or changing credentials.

Inspect discovery and the health result

After adding a connection, inspect the reported result. A saved connection can still have a failed handshake. A discovered tool count confirms that the server returned tool definitions at that moment; it does not mean every operation was exercised successfully.

If discovery fails, check the endpoint, the required authentication method and the account's access to the source project. If you are using a self-hosted server, check that its remote endpoint can be reached by the service. A process running only on your laptop is not automatically available to a hosted web application.

Retry after one targeted correction and record the result. Repeatedly reconnecting without understanding the error can hide the difference between an expired credential, an unsupported endpoint and a permission problem.

Ask for evidence from a bounded request

For the fictional three-issue project, a useful test request is:

Use the connected project tool to list the open issues in the test project. Show each title, status and source link. If the tool cannot access the project, tell me what failed; do not fill in missing records from memory.

Compare the result with the original project. If an issue is absent, check the tool's filters and the account's permissions before assuming that it does not exist. For a database report, establish the table, date range and relevant time zone. A correct-looking summary can still answer a different question.

Separate retrieval from actions

Once reading works, decide which actions you actually need. Updating an issue, inserting a record and sending a message have different consequences. Confirm the target and the proposed change before an important write, then inspect the source afterward. A fluent answer alone cannot prove that a connector performed an action.

If you only want to build an interface around your workflow, explore the web app builder. Creating an interface does not automatically configure its access to production data. Test the integration as a separate part of that project.

Common questions

Does connecting a server give Brainbaby permanent access to everything?

Access depends on the provider, the credential, workspace configuration and the tools the server exposes. Check the current connection and source permissions instead of treating the connection's name as a blanket access guarantee.

Why can a saved connection still fail?

Saving settings and completing a successful discovery request are different steps. Credentials, endpoints and source permissions can change. Review the health result and retest a small request before relying on the connection.

Can I connect a local server directly?

A local process is not the same as a remotely accessible server endpoint. Confirm that your server's transport and deployment are supported. Do not expose a database publicly merely to make a connection work.

Check current plans and allowances before a larger batch of assistant tasks. Keep a small, repeatable test question so you can verify the connection again when its configuration changes.