Create & connect
Connected apps: let models use your tools safely
Connect an MCP app, enable it for a workspace, review tool arguments, and approve or deny actions before they run.
Before you start
Section titled “Before you start”Connected apps use MCP servers to make external tools available during a model conversation. Manage account-level connections in Settings → Connected apps. A connection being saved does not mean it is active in every workspace: use the MCPs control in the current workspace to enable or disable each available server there.
Connect only services you trust. A tool may read, create, change, or delete information in the external service. Review the displayed tool name and arguments before allowing it. Never paste passwords, API keys, or access tokens into a prompt or share them with a model.
When to use this
Section titled “When to use this”Connect an app when you want a model to retrieve information from a service or perform a supported action there, such as looking up an issue. If you only need to discuss information without granting tool access, paste a safe excerpt or attach a document instead. For connecting Neurothread itself to another AI assistant, use the official guide at neurothread.ai/mcp; this article covers apps used by models inside Neurothread.
- Open Settings → Connected apps. Choose a listed app and select Connect. If it supports sign-in, complete the provider’s authorization in the browser and return to Neurothread.
- For a custom server, choose Add a custom MCP server. Enter a name, an HTTPS Server URL, and the matching Transport. Choose an authentication method only if the server requires it, then provide its credential in the protected field and click Connect.
- Check that the connection loaded its tools. Use Test to reconnect and reload the tool list if necessary. The server card shows the tool count or an error.
- Open the intended workspace and click MCPs. Switch on only the connections needed for that workspace. This workspace setting is separate from the connection’s account-level enabled state in Settings.
- Ask a question that clearly identifies the task and the service. When a tool call is pending, inspect the server, tool name, and Arguments. Select Allow to run it or Deny to refuse it. For several pending calls, decide each one; Allow all approves the pending calls shown together.
- Review the Result displayed after a tool runs. The model can use that result to continue its answer. Turn off a connection in MCPs when you no longer need it in that workspace.
- To edit credentials or server details, return to Settings → Connected apps and use Edit. Use Remove only if you want to delete the saved connection and its stored token.
Worked example
Section titled “Worked example”Connect a project tracker that you trust, enable it in a test workspace, and ask: “List the open tasks assigned to me in the Acme project. Do not change anything.” Before allowing a call, inspect its tool and arguments to confirm it is a read operation on the intended project. Review the result. If the tool proposes changing a task, deny it unless that change is what you intended.
What happens next
Section titled “What happens next”Approved results are shown with the answer and can inform the rest of that response. Tool calls and their statuses remain visible in the conversation. A connection can be switched off per workspace without removing it from your account. For services where access is no longer needed, remove the connection in Settings; then review the external provider’s own access controls as appropriate.
Cost and limitations
Section titled “Cost and limitations”A model answer that uses tools can consume Neurothread credits. The external service may also apply its own rate limits, permissions, or charges. You can have up to 20 connections. Only remote HTTPS MCP servers are supported; internal/private network addresses are blocked. Most calls require your approval. Destructive tools always ask for approval; in Build, read-only tools can run inline without asking. Smart approval is not assumed to be enabled. Tool access does not guarantee that a server exposes the action you need.
Troubleshooting
Section titled “Troubleshooting”- “Enter a valid https:// URL.”: use the complete HTTPS endpoint, including the server path if required.
- “No MCP endpoint at that URL.”: check the endpoint path (often
/mcp) and select the transport the provider supports. - “This server requires signing in.” or “The server rejected the token.”: complete Sign in if offered, or check the credential and reconnect.
- “That address points to a private or internal network.”: Neurothread blocks internal/private server addresses; use a supported public HTTPS endpoint.
- Connection loads no tools: click Test. If the server remains unavailable, confirm it is online and that your account has permission.
- The model never uses the app: check both the account-level connection in Settings and the workspace switch under MCPs, then ask for a specific action supported by the listed tools.