The MCP tool overload problem

Author

Ikenna Nwaiwu
November 06, 2025

Issue 57

Contents

  • The MCP tool overload problem

  • Interesting content for the week

  • FeedBack & share

  • Upcoming conferences

  • My services: API governance consulting

  • References

The MCP tool overload problem

Model Context Protocol (MCP) tool overload occurs when AI agents are exposed to too many tools advertised by MCP servers. The excessive number of tools, each with its own description of inputs, outputs, and use cases, can make it difficult for a client agent to pick the right tool for the task [1, 2, 3, 4]. This 'tooling bottleneck' problem [5] results in user frustration with the performance of the AI agent, as shown in the comment image below.

The problem is not just due to the limited context window of the agent (i.e context window exhaustion), but due to too many tools being exposed to the agent’s reasoning loop or ranking layer

Why this issue is important from a runtime AI governance perspective

The large context window leads to increased latency and slower responses for the user. Poor tool selection leads to both user dissatisfaction with the AI system and a poor dev experience for the builder. Also, a bloated context window leads directly to higher operational costs, which increases the total cost of ownership of the system.

How MCP tool overload fits into the IGT-AI risk framework.

I created the IGT-AI framework [6] for evaluating AI gateways. The risk model [7] I created for it has four categories: security and privacy risks, content risks, operational and performance risks, and financial risks. I am looking at using this risk model to capture the risks that MCP gateways [8] mitigate, too. In that vein, tool overload will fall in the operational and performance risks category, as shown in the diagram below (although you could argue it overlaps into financial risks as well).

MCP tool overload in the IGT-AI risk model

Approaches for mitigating tool overload

The main approaches involve reducing the cognitive load on the agent in some way, and they are summarised in the table below.

No.

Technique

Description

Comment

1

Selective Tool Activation

Activate and expose only the necessary tools for the given user and task.

This can be done manually, but that is tedious. [4, 9]

2

Tool filtering, tagging, and namespacing

Use metadata tagging (e.g financial, admin) or filtering rules to do a "just-in-time" selection of the appropriate tool based on the task.

Jenova uses this approach for its general AI agent.[5]

3

Use narrower-scoped MCP servers

Instead of large MCP servers, create multiple narrower-scoped servers, based on their domain or function (e.g., "Sales" server, "HR" server).

This approach can also be applied when designing APIs for agents. Don't just convert an API with numerous endpoints into an MCP server. Design the MCP server with the workflows the agents have to execute in mind (which means, using only a select set of endpoints). Arazzo can be helpful here. [14]

4

Hierarchical and semantic tool selection

Sub-agent / tool librarians for tool selection: Deploy specialist AI agents or subsystems (the 'librarians') to intelligently find, assess, and recommend the correct tools. [8, 11, 12]

Embedding-based and graph-phased tool selection: Use a vector embedding or a graph database to select the appropriate tool.

This can also be called 'agentic MCP server configuration' [4] or 'librarian agent' [11]. Klava [12] is an example of this.

A graph-based approach models the tool ecosystem as a graph database. Tools can be represented as nodes, and the relationships between them represented as edges. [15]

5

Context window optimisation

Remove non-essential context, compress tool descriptions, and optimise prompt design.

This is an MCP server design technique that also reduces token usage.

6

Dynamic tool loading

Load and unload tools on demand to dynamically manage and reduce the active context window

This can be done in Spring AI [10], but I am not sure if it requires a configuration reload (I need to test it).

Interesting content for the week

API Production Governance

What Makes a Quality API Contract?: After a recent post I made on LinkedIn, where I discussed why high-quality API contracts are essential for agentic use cases, I received a question about what exactly defines a “high-quality API contract“. I wrote this blog post to answer that question.

Runtime AI Governance

The Agentic AI advantage: Fact, fiction and the future: This report, written by Thoughtworks in collaboration with WIRED, asserts that while AI agents hold the potential to revolutionise enterprise operations, a clear, well-articulated strategy is the difference between success and failure. The report cautions that up to 40% of projects that will be deployed by IT leaders in the coming year may be cancelled by 2027 due to issues surrounding reliability, governance, and security.

The Rapidly Changing Landscape of APIs: Navigating the 2026 API Ecosystem: Kong Inc., analyses the rapid evolution of the API ecosystem and asserts that by 2026, APIs will have firmly transitioned from optional developer tools into regulated, monetised, and essential infrastructure for the digital economy.

AI Governance integration on ML/LLM workflow: Experts in AI governance highlight that to build responsible, compliant, and ethical AI, governance and policy must be integrated into the machine learning workflow from its earliest stages, rather than being left until post-deployment.

Making MCP Servers Production-Ready: Utkarsh Contractor addresses the crucial issue of authorisation necessary to transition Model Context Protocol (MCP) Servers from development to secure production environments.

Prompting agents: What works and why: Nolan Sullivan explores the nuances of prompting Agentic AI, arguing that agents are fundamentally different from chatbots, which merely generate text.

7 MCP Registries Worth Checking Out: J Simpson highlights registries that offer different functionalities, each with a distinct focus, allowing developers to find the right server for their needs.

How MCP and AI are Modernizing Legacy Systems: Saqib Jan, argues that enterprises can finally modernise rigid, monolithic legacy systems without undertaking risky, massive rewrite projects by using a pragmatic new strategy centred on Agentic AI and the Model Context Protocol (MCP).

Update on the Next MCP Protocol Release: David Soria Parra outlines five key protocol improvements prioritised for the upcoming release on November 25th, 2025.

Feedback & Share

What do you think of this newsletter issue?

Login or Subscribe to participate in polls.

Upcoming Conferences

Apidays Paris: Apidays Paris sparks essential conversations on data security, digital sovereignty, and sustainable innovation in the age of intelligent systems. Date: 9 - 11 December 2025 Location: CNIT Forest, Paris.

My Services: API Governance Consulting

Is poor API governance slowing down your delivery? Do you experience API sprawl, API drift and poor API developer satisfaction? I'll provide expert guidance and a tailored roadmap to transform your API practices.

Ikenna® Delivery Assessment → Identify your biggest API delivery pain points.

Ikenna® Delivery Canvas (IDC) & API Transformation Plan → Get a unified, data-driven view of your API delivery and governance process.

Ikenna® Improvement Cycles → Instil a culture of scientific, measurable progress towards API governance.

Ikenna® Governance Team Model → Set up and improve your governance team to sustain progress.

Ikenna® Delivery Automation Guidance → Reduce lead time and improve API quality through automation

Schedule a consultation by emailing: [email protected].

References

[1] S. Blanchfield, “The MCP Tool Trap.” Jentic Blog. https://jentic.com/blog/the-mcp-tool-trap (Accessed: Nov. 6, 2025).

[2] V. Paramanayakam, A. Karatzas, I. Anagnostopoulos, and D. Stamoulis, “Less is More: Optimizing Function Calling for LLM Execution on Edge Devices.” arXiv. https://arxiv.org/abs/2411.15399 (Accessed: Nov. 6, 2025).

[3] J. Flores, “Context as the New Currency: Designing Effective MCP Servers for AI.” Itential Blog. https://www.itential.com/blog/company/ai-networking/context-as-the-new-currency-designing-effective-mcp-servers-for-ai/ (Accessed: Nov. 6, 2025).

[4] T. Antanavicius, “Agentic MCP Configuration.” PulseMCP Blog. https://www.pulsemcp.com/posts/agentic-mcp-configuration (Accessed: Nov. 6, 2025).

[5] Jenova AI, “The Tooling Bottleneck: How Reliability and Scalability Challenges Are Stalling the Future of MCP and Agentic AI.” Jenova AI Resources. https://www.jenova.ai/en/resources/the-tooling-bottleneck-how-reliability-and-scalability-challenges-are-stalling-the-future-of-mcp-and-agentic-ai (Accessed: Nov. 6, 2025).

[6] Ikenna Consulting, “igt-ai.” GitHub Repository. https://github.com/IkennaConsulting/igt-ai/ (Accessed: Nov. 6, 2025).

[7] Ikenna Consulting, “Risks.” igt-ai GitHub Repository. https://github.com/IkennaConsulting/igt-ai/blob/main/igtai/risks.md (Accessed: Nov. 6, 2025).

[8] H. Sharma, “The MCP Gateway: Enabling Secure and Scalable Enterprise AI Integration.” InfraCloud Blog. https://www.infracloud.io/blogs/mcp-gateway/ (Accessed: Nov. 6, 2025).

[9] S. Gawas, “How to fix MCP tool overload.” The AI Stack. https://www.theaistack.dev/p/managing-mcp-tools (Accessed: Nov. 6, 2025).

[10] C. Tzolov, “Spring AI: Dynamic Tool Updates with MCP.” Spring Blog. https://spring.io/blog/2025/05/04/spring-ai-dynamic-tool-updates-with-mcp (Accessed: Nov. 6, 2025).

[11] J. Chong, “Avoid the MCP Server Overload.” TIBCO Blog. https://www.tibco.com/blog/2025/10/16/tibco-ai-avoid-the-mcp-server-overload/ (Accessed: Nov. 6, 2025).

[12] Klavis AI, “Core Concepts: Strata.” Klavis Documentation. https://www.klavis.ai/docs/concepts/strata (Accessed: Nov. 6, 2025).

[13] T. Mclaughlin, “Beyond API Wrappers: Workflow-Based MCP Servers with Elicitation and Sampling.” YouTube. https://www.youtube.com/watch?v=JuPoDQwmYMU (Accessed: Nov. 6, 2025).

[14] OpenAPI Initiative, “Arazzo Specification.” OpenAPI.org. https://www.openapis.org/arazzo-specification (Accessed: Nov. 6, 2025).

[15] M. Lenhard, “From Embeddings to Edges: A Graph Based Approach to Tool Selection,” YouTube, video, 39:12, Jul. 23, 2024. https://www.youtube.com/watch?v=5nFOhwAwVGM (Accessed: Nov. 6, 2025).

Reply

or to participate.