Tool overload

What it is

Every MCP tool a client loads puts its name, description and schema into the model's context. Connect a few MCP servers that each map an API one-to-one and the model is choosing between hundreds of similar tools before it has read the question.

Anthropic's tool search documentation names two effects:

• Context bloat. A typical setup with five servers can use about 55,000 tokens on definitions before any work is done.
• Selection accuracy. Claude's ability to pick the right tool degrades once more than 30 to 50 tools are available.

Why it matters

An agent that picks the wrong tool does the wrong thing, or does nothing and reports success. Context spent on definitions is not available for the task, and large tool lists raise the cost of every request.

How it works

The fixes all load less up front:

1. Tool search. The model searches the catalogue and loads only matching definitions. Anthropic's tool search tool works this way.
2. Progressive discovery. The MCP client best-practices guide describes three layers: a catalogue the model can search, an inspect step that loads one tool's full schema, and execution. It suggests switching when definitions pass a threshold such as 1 to 5% of the context window.
3. Fewer, broader tools designed for the model, rather than one per endpoint.
4. Programmatic tool calling, where the model writes code that calls tools in a sandbox and only the result returns.

Example

An agent with six MCP servers attached is asked to update a contact's phone number. It sees three CRMs' update tools, two "patch person" tools and a generic HTTP tool, and calls the wrong CRM. With search, it asks for "update contact phone" in the connected CRM and loads one tool.

Common pitfalls

• Connecting everything "in case".
• Similar names across servers, such as several create_item tools.
• Fixing it by hand, disabling tools per session, which breaks when the task changes.

How actionlake handles it

An agent on actionlake gets eight tools, whether one system is connected or a hundred, and adding a connector never adds a tool. The catalogue, inspect and execute layers are built into the server: find_actions searches 289,000+ actions, get_action returns one action's full schema, and do or ask runs it. Searching and inspecting are free. See the eight tools on the docs page and the platform page.