Short answer: KeyRunner can reduce tool-related LLM token usage by returning only the tools an agent is authorized to use. Each agent is assigned a named policy containing an approved group of tools. The agent does not receive the complete enterprise tool catalog.
Why large tool catalogs consume LLM context
An AI tool is more than a short name. Its definition normally includes a description and a parameter schema. Some tools also contain detailed instructions, allowed values, and output information. When a client sends these definitions to a model, they become part of the model's available context.
Both function-calling systems and MCP tools use structured definitions to help a model decide what to call. The OpenAI API reference counts tool definitions as part of conversation instructions, while the MCP tools specification defines tools through names, descriptions, and input schemas.
If an agent needs five tools but receives 500, the model may be given hundreds of definitions it cannot or should not use. That extra context can increase input tokens, weaken prompt-cache efficiency when catalogs change, and give the model more irrelevant choices.
What is AI agent tool curation?
AI agent tool curation is the process of giving an agent a small, relevant set of tools instead of every tool connected to the organization. The selection should be based on the agent's role, purpose, and access policy.
Tool curation is different from asking the model to ignore tools. Tools outside the agent's policy are not returned as available choices. This applies least privilege at discovery time, before the agent attempts an action.
How KeyRunner curates tools using policy
In KeyRunner, a policy is a named group of tools. When an enterprise onboards an agent, it assigns that agent a policy based on its job. When the agent asks KeyRunner for tools, KeyRunner returns only the tools allowed by that policy.
- Onboard the agent. Register its identity, role, and business purpose in KeyRunner.
- Assign a named policy. The policy groups the specific tools that role is allowed to discover and call.
- Request available tools. The agent connects through KeyRunner and requests its tool list.
- Filter before discovery. KeyRunner evaluates the assigned policy and returns only approved tools.
- Choose from a focused catalog. The model receives fewer tool definitions and selects from tools relevant to its task.
- Audit activity live. KeyRunner records the agent, tool, policy decision, execution result, and time.
The MCP specification supports this general pattern. It states that a tool set may vary based on the authorization presented with a request. KeyRunner adds an enterprise policy layer that makes the permitted set explicit and manageable by agent role.
How curated tools can reduce token usage
Tool-related context is broadly driven by the number and size of definitions presented to the model. A useful way to think about it is:
tool context ≈ names + descriptions + parameter schemas + tool instructions
If a complete catalog contains 200 tools and a finance agent's policy contains 12, KeyRunner returns those 12 approved tools. The exact token difference depends on the length of the schemas, the model client, caching, and how often the definitions are sent. KeyRunner does not claim one universal savings percentage. It removes definitions the agent is not allowed to use, which reduces tool-related context whenever those definitions would otherwise be sent to the LLM.
Token efficiency and security reinforce each other
Policy-curated tools are not only a cost optimization. The same smaller catalog also supports least privilege. A customer-support agent can receive tools for reading cases and updating approved fields without seeing payroll, infrastructure, or bulk-export tools.
This can also improve tool selection. A model choosing among a few well-named, relevant tools has fewer unrelated descriptions to compare. Curation does not guarantee correct selection, but it removes choices that should never have been available to that agent.
Live audit logs answer: which agent is doing what?
Enterprises need more than a list of permitted tools. They need evidence of actual behavior. KeyRunner captures live audit logs so security, platform, and compliance teams can trace which agent requested a tool, which policy applied, whether the action was allowed or blocked, and what happened during execution.
This connects tool efficiency with agentic security and AI governance. Policies define the agent's available capabilities. Audit records show how those capabilities are used.
Example: different agents, different tool policies
- SupportAgent policy: get customer, read case, update case status, issue a refund below an approved limit.
- FinanceAgent policy: read invoice, check payment, create payment request, request approval for a transfer.
- DevOpsAgent policy: read deployment status, trigger an approved pipeline, open an incident, roll back a release.
Each agent receives a focused catalog. None of them needs to process the definitions belonging to the other roles. Administrators manage the named policies centrally, while KeyRunner applies the correct tool set when each agent connects.
What to measure when reducing tool-related tokens
- Total tools connected to the enterprise environment.
- Tools returned to each agent after policy filtering.
- Tokens used by the full catalog versus each curated catalog.
- Prompt-cache hit rate for stable tool lists.
- Tool selection errors and calls to irrelevant tools.
- Cost and latency per agent session.
- Allowed, denied, and failed tool executions by policy.
Frequently asked questions
How can enterprises reduce LLM token usage from AI agent tools?
Return only the tools that are relevant and authorized for each agent. A smaller tool catalog means fewer tool names, descriptions, and parameter schemas need to enter the model context. KeyRunner does this through policy-based tool curation.
What is policy-based tool curation?
Policy-based tool curation maps an AI agent to a named policy containing approved tools. When the agent asks for available tools, it receives only that policy-approved set instead of the full enterprise tool catalog.
Does a smaller tool catalog always reduce tokens?
It reduces the amount of tool-definition content presented to the model when the client includes those definitions in context. Actual savings depend on schema length, catalog size, client behavior, caching, and how often tools are sent.
Can curated tools improve AI agent accuracy?
A focused tool set can make selection easier because the model compares fewer irrelevant or similar options. Results still depend on clear names, descriptions, schemas, prompts, and the model being used.
How does KeyRunner audit AI agent tool usage?
KeyRunner captures live audit records for agent tool activity, including the agent identity, requested tool, policy decision, execution status, and time. This helps enterprises see which agent is doing what.
The takeaway
KeyRunner gives each AI agent the smallest useful tool catalog based on policy. Instead of sending every connected tool to every agent, enterprises define named groups of approved tools and assign them by role. This can reduce tool-related LLM context, support better tool selection, enforce least privilege, and provide live visibility into agent activity.
See how KeyRunner governs agent execution, explore the interactive scenarios, or book a product walkthrough.
