Control MCP tool definition token overhead in agent workflows
Updated October 22, 2026 · first published September 19, 2026
MCP tool definitions add model-input tokens when the host includes them in a request. The size and cost depend on the exposed schemas, active tool selection and provider cache behavior. Connecting more servers does not imply a fixed number of tokens or a fixed overhead percentage.
Where does tool-schema overhead come from?
A tool definition can contain a name, description and input schema with nested types and parameter guidance. The official MCP architecture separates tool discovery from execution. The host controls which discovered definitions reach the model; MCP does not require every schema to appear on every inference turn.
Measure the serialized input your model provider receives. Counting connected servers or tool names is not enough. A small catalogue with long parameter descriptions may consume more tokens than a larger catalogue of concise definitions.
How do you estimate the cost?
Illustrative example: if your host includes a 4,000-token tool block in six requests, that is 24,000 input-token occurrences. At an assumed uncached rate of $2 per million, the schema component costs $0.048. Reducing each block to 1,000 tokens brings it to $0.012, saving $0.036 if the other request components and outcomes stay unchanged.
Those are assumed rates and counts, not measured production savings. Cache hits or writes change the calculation. Tool selection may also change the number of steps, so compare whole-workflow cost and completion quality, not just the smaller prompt.
Which approaches reduce unnecessary schema input?
- Select relevant tools: expose the tools needed for the current task or discover them on demand. Measure any router call or discovery overhead too.
- Keep stable prefixes reusable: apply the provider's supported caching controls and log actual cache reads and writes. Identical schema text alone does not guarantee a hit.
- Shorten descriptions carefully: remove duplication while retaining units, required fields, constraints and guidance needed to choose the correct tool.
Do not assume a universal schema size, cache lifetime or savings percentage. OpenAI's caching guide, for example, distinguishes cache-write categories from cached reads. Other providers have their own rules.
How should you validate a change?
- Replay a fixed set of tasks with the same model settings and record cache state.
- Compare the full catalogue against a selected subset or revised descriptions.
- Record tool-selection errors, missing required tools, retries, latency and acceptance outcomes.
- Sum model and downstream tool charges across every attempt. Divide by successful tasks.
A shorter tool description is useful only if the agent can still select and call the tool correctly. Keep changes that improve measured task economics at your required quality.
Related guides and sources
The MCP server cost guide separates hosting, downstream services and token costs. The agent workflow architecture guide adds logging, shared budget reservations and retry controls. Official architecture and caching references above were checked on 22 October 2026; this revision updates the English text.
Frequently asked questions
Are all connected MCP tool definitions sent to the model on every turn?
Not necessarily. The host chooses which tools to expose. Discovery, model input and tool execution are separate operations.
How can teams reduce MCP tool-definition overhead?
Select relevant tools, reuse eligible stable prefixes and shorten redundant descriptions. Test tool selection and cost per successful task before claiming savings.
Does caching eliminate tool-schema costs?
No. Cache reads and writes use provider-specific rates and eligibility rules. Measure actual usage rather than assuming every repeated definition is a hit.
Related
- Server tools and separate invoices
- Prompt caching economics
- Agent spend guardrails
- MCP server costs
- Practical agent workflow architecture
Want this applied to your stack? Bring the provider bills, gateway logs, and top workflows; we will map the cost drivers and savings path. Book a free audit →