New on the blog: why AI belongs in the feed →
View Categories

MCP Servers And Tools

MCP tools require Premium or higher.

An MCP server gives an AI model tools it can call while it answers: look up an order, search a knowledge base, create a task, send a message. A text feed can let its model use tools from the servers you add, so a form submission can make something happen in another system instead of only producing text.

The AI provider connects to the MCP server and runs the tools. This site sends the provider the server’s address, its access token and the list of tools the feed allows, and receives back the model’s answer and which tools were called.

Before You Start #

  • The server must be reachable from the internet over HTTPS. A server on a private network or on your own computer cannot be used, because the provider cannot reach it.
  • Tools work with Anthropic and OpenAI models. They are not available for OpenAI-compatible provider profiles.
  • You need the server’s address, its access token if it requires one, and the exact names of the tools you want to allow.

Add A Server #

  1. Open GravityOps -> AI Feed -> Credentials and scroll to MCP servers.
  2. Click Add MCP server.
  3. Enter a Name. It is shown where a feed chooses tools, in entry notes and in the ledger.
  4. Enter the Server URL, such as https://mcp.example.com/mcp.
  5. Enter the Access Token if the server needs one. It is encrypted, never shown again, and sent to the server as a Bearer token.
  6. List the Allowed Tools, one name per line, exactly as the server names them.
  7. Leave Enabled checked and click Add Server.

Feeds can only choose from the Allowed Tools list. A tool that is not listed is never offered to a model, even if the server has it. To change the token later, edit the server and enter a new one; leave the field blank to keep the saved token.

Let A Feed Use Tools #

  1. Edit an AI Text feed and find Tools.
  2. Select the tools this feed needs. Select as few as possible.
  3. Describe in the User Prompt what the model should do with them.
  4. Choose the Output Mode:
    • Single target field saves the model’s closing answer to a field, as usual.
    • Actions only saves no field. The model’s closing summary is added to the entry as a note.
  5. Save the feed.

Rules For A Feed With Tools #

A tool can change data in another system, so a feed with tools runs under stricter rules:

  • It runs after form submission only. It cannot use a frontend button or streaming, so a visitor cannot trigger its tools repeatedly.
  • It always runs in the background, whatever Execution Mode says. Tool calls take time and do not hold up the submission or the confirmation.
  • It cannot use JSON multi-field mapping.
  • It cannot have a Fallback Model. A second model would repeat actions the first one already took.
  • Every tool is chosen by name. A feed never gets “all tools on the server”.

The feed’s settings are checked when the feed is saved and again each time it runs. If a server is disabled or deleted, or a tool is removed from its allowlist, feeds that use it fail with a clear message until their Tools setting is changed.

What Is Recorded #

  • An entry note lists the tools that were called, on which server, and how many calls returned an error. It is added whether or not the run succeeded, because the tools have acted either way.
  • The ledger’s Diagnostic column shows the same for each run.
  • Tool inputs and tool results are never stored by this plugin.

Rerunning #

Rerunning a feed with tools repeats its actions: the message is sent again, the record is created again. The entry screen’s rerun button asks for confirmation first. Check the entry’s notes for what already happened before you rerun.

If a run fails after a tool was called, the note still lists the call. The action may have happened even though the feed reports a failure.

Security And Privacy #

  • Submitted values reach the MCP server. Whatever your prompt inserts from the entry goes to the AI provider, and the model can pass it on to a tool. Add only servers you trust.
  • Whoever submits the form can try to steer the model. Text typed into a form field becomes part of the prompt, and a visitor can write instructions into it. The allowlist is the real limit on what the model can do: a feed on a public form should have only tools whose worst-case use you can accept.
  • The access token goes to the AI provider with each request, so the provider can connect to the server. Use a token limited to what the feed needs.
  • Anthropic’s MCP connector is a beta feature and is not covered by Anthropic’s zero-data-retention arrangements. OpenAI sends the model’s tool inputs to the server you name; review its remote-MCP guidance before connecting a server you do not operate.

Stuck? Ask in GravityOps Help #

Search past answers or post your question, tagged "ai-feed".