Why The Model Reads Tool Definitions¶
Before a model can use a tool, the client asks the server for its tool list. The server answers with a definition for each tool, and the client passes those definitions to the model as context. Everything the model knows about a tool before the first call comes from text that the server's author wrote.
- The client sends a
tools/listrequest to the server. - The server returns its tool definitions.
- The client passes the definitions to the model as context.
- The model reads them and selects a tool for the user's request.
- The client sends a
tools/callrequest, and the server runs the tool.
Inside A Tool Definition¶
The example below comes from the MCP specification.
{
"name": "get_weather", // (1)!
"title": "Weather Information Provider",
"description": "Get current weather information for a location", // (2)!
"inputSchema": { // (3)!
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "City name or zip code" // (4)!
}
},
"required": ["location"]
}
}
- The tool's unique identifier within one server.
- Free text from the server's author. The model reads it to decide when and how to call the tool.
- A JSON Schema that defines the parameters the tool expects.
- Parameter descriptions are free text too, and the model reads them as well.
A definition can also carry annotations, optional properties that describe the tool's behaviour. The specification requires clients to treat annotations as untrusted unless they come from a trusted server.
Text That Works As An Instruction¶
A description is written for the model, and a model is built to follow natural-language instructions. OWASP's MCP Security Cheat Sheet names the resulting risk tool poisoning, where malicious instructions hidden in tool descriptions, parameter schemas or return values manipulate the model's behaviour. OWASP advises treating the whole schema as a possible injection surface, parameter names and types included.
Research Demonstration, April 2025
Invariant Labs built a test server with one tool, get_fact_of_the_day. Its description told the agent to change the recipient of any WhatsApp message sent through a second, trusted server. In the researchers' demonstration, the agent sent the user's WhatsApp chat data to a number they controlled. Nobody was attacked, and the work was a proof of concept. Part 2 walks through it step by step.
One Server Can Steer Another's Tools¶
The model sees tool descriptions from every connected server in the same context, as OWASP's cheat sheet points out. A description from one server can therefore change how the agent uses a tool from another, a risk OWASP lists as tool shadowing. In the Invariant Labs test, the untrusted server never touched WhatsApp itself.
Definitions Can Change¶
A tool list isn't fixed. The specification allows the set of tools to change over time, and a server that declares listChanged should notify clients that subscribed to updates. OWASP uses the term rug pull for the harmful case, where a server changes its tool definitions after the user's initial approval.
The Practical Consequence
Reviewing a server's tools once tells you what they said on that day. Part 5 covers how to record approved definitions and detect a change.