Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhancing Model Context Protocol (MCP) with Context-Aware Server Collaboration

Meenakshi Amulya Jayanti, X. Y. Han

TL;DR

CA-MCP reimagines MCP by moving execution from a central LLM to specialized MCP servers that co‑ordinate through a Shared Context Store. The central LLM serves as high‑level planner and final summarizer, while the SCS enables real‑time, event‑driven collaboration among stateless servers, reducing central bottlenecks. Empirical evaluation on TravelPlanner and REALM‑Bench demonstrates substantial latency reductions (up to ~74%) and improved constraint satisfaction, with robust coordination and fewer LLM invocations. This hybrid architecture offers a scalable path toward more efficient, resilient LLM‑driven multi‑agent systems by decoupling planning from execution and externalizing shared state for continuous coherence.

Abstract

The Model Context Protocol (MCP) has emerged as a widely used framework for enabling LLM-based agents to communicate with external tools and services. The most common implementation of MCP, proposed by Anthropic, heavily relies on a Large Language Model (LLM) to decompose tasks and issue instructions to servers, which act as stateless executors. In particular, the agents, models, and servers are stateless and do not have access to a global context. However, in tasks involving LLM-driven coordination, it is natural that a Shared Context Store (SCS) could improve the efficiency and coherence of multi-agent workflows by reducing redundancy and enabling knowledge transfer between servers. Thus, in this work, we design and assess the performance of a Context-Aware MCP (CA-MCP) that offloads execution logic to specialized MCP servers that read from and write to a shared context memory, allowing them to coordinate more autonomously in real time. In this design, context management serves as the central mechanism that maintains continuity across task executions by tracking intermediate states and shared variables, thereby enabling persistent collaboration among agents without repeated prompting. We present experiments showing that the CA-MCP can outperform the traditional MCP by reducing the number of LLM calls required for complex tasks and decreasing the frequency of response failures when task conditions are not satisfied, thereby improving overall efficiency and responsiveness. In particular, we conducted experiments on the TravelPlanner and REALM-Bench benchmark datasets and observed statistically significant results indicating the potential advantages of incorporating a shared context store via CA-MCP in LLM-driven multi-agent systems.

Enhancing Model Context Protocol (MCP) with Context-Aware Server Collaboration

TL;DR

CA-MCP reimagines MCP by moving execution from a central LLM to specialized MCP servers that co‑ordinate through a Shared Context Store. The central LLM serves as high‑level planner and final summarizer, while the SCS enables real‑time, event‑driven collaboration among stateless servers, reducing central bottlenecks. Empirical evaluation on TravelPlanner and REALM‑Bench demonstrates substantial latency reductions (up to ~74%) and improved constraint satisfaction, with robust coordination and fewer LLM invocations. This hybrid architecture offers a scalable path toward more efficient, resilient LLM‑driven multi‑agent systems by decoupling planning from execution and externalizing shared state for continuous coherence.

Abstract

The Model Context Protocol (MCP) has emerged as a widely used framework for enabling LLM-based agents to communicate with external tools and services. The most common implementation of MCP, proposed by Anthropic, heavily relies on a Large Language Model (LLM) to decompose tasks and issue instructions to servers, which act as stateless executors. In particular, the agents, models, and servers are stateless and do not have access to a global context. However, in tasks involving LLM-driven coordination, it is natural that a Shared Context Store (SCS) could improve the efficiency and coherence of multi-agent workflows by reducing redundancy and enabling knowledge transfer between servers. Thus, in this work, we design and assess the performance of a Context-Aware MCP (CA-MCP) that offloads execution logic to specialized MCP servers that read from and write to a shared context memory, allowing them to coordinate more autonomously in real time. In this design, context management serves as the central mechanism that maintains continuity across task executions by tracking intermediate states and shared variables, thereby enabling persistent collaboration among agents without repeated prompting. We present experiments showing that the CA-MCP can outperform the traditional MCP by reducing the number of LLM calls required for complex tasks and decreasing the frequency of response failures when task conditions are not satisfied, thereby improving overall efficiency and responsiveness. In particular, we conducted experiments on the TravelPlanner and REALM-Bench benchmark datasets and observed statistically significant results indicating the potential advantages of incorporating a shared context store via CA-MCP in LLM-driven multi-agent systems.
Paper Structure (32 sections, 6 figures, 3 tables)

This paper contains 32 sections, 6 figures, 3 tables.

Table of Contents

  1. Reproducible Code
  2. Introduction
  3. Organization of the Paper.
  4. Limitations and Promising Future Work.
  5. Detailed Architecture of CA-MCP
  6. Key Components of CA-MCP
  7. Central LLM (Long-Term Planner).
  8. MCP Servers (Short-Term Reactors).
  9. Shared Context Store.
  10. Benefits and Conceptual Shift
  11. Related Work
  12. Centralized Orchestration.
  13. Distributed and Hybrid Coordination.
  14. Blackboard Architecture.
  15. SagaLLM.
  16. ...and 17 more sections

Figures (6)

  • Figure 1: Traditional Model Context Protocol (MCP) architecture anthropic_mcp. The Central LLM acts as a centralized orchestrator (A), issuing step-by-step tool calls to stateless MCP servers (B) via synchronous interactions (C), tightly coupling execution to continuous LLM involvement.
  • Figure 2: Our proposed CA-MCP architecture. A new Shared Context Store (SCS) (B') enables autonomous server coordination without continuous LLM involvement. The Central LLM functions mainly as planner and summarizer (A'), while servers operate as stateful reactors over shared state (C'), shifting execution from centralized orchestration to distributed coordination (D').
  • Figure 3: Execution Time (sec) on TravelPlannertravelplanner (500 queries). (a) Overlaid histograms for Traditional (blue) and Context-Aware (green); dashed lines show means: Traditional = 41.989, Context-Aware = 13.523. (b) Pairwise differences (Traditional -- Context-Aware); positive values indicate Context-Aware is faster. Paired t-test:mean_diff = 28.465s, $\mathtt{sd} = 18.059 s$, $n = 500$, $\mathtt{pval} = 1.60e{-137}$. Takeaway: CA-MCP markedly reduces latency.
  • Figure 4: Completeness Score on TravelPlannertravelplanner (500 queries). (a) Overlaid histograms for Traditional (blue) and Context-Aware (green); dashed lines show means: Traditional = 0.764, Context-Aware = 1.000. (b) Pairwise differences (Context-Aware -- Traditional); positive values indicate higher completeness for Context-Aware. Paired t-test:mean_diff = 0.236, $\mathtt{sd} = 0.2408$, $n = 500$, $\mathtt{pval} = 4.39e{-75}$. Takeaway: CA-MCP consistently attains higher completeness.
  • Figure 5: BERTScore (F1) on TravelPlannertravelplanner (500 queries). (a) Overlaid histograms for Traditional (blue) and Context-Aware (green); dashed lines show means: Traditional = 0.745, Context-Aware = 0.757. (b) Pairwise differences (Context-Aware -- Traditional). Paired t-test:mean_diff = 0.012, $\mathtt{sd} = 0.04728$, $n = 500$, $\mathtt{pval} = 2.35e{-08}$. Takeaway: CA-MCP yields a modest but consistent lift in semantic accuracy.
  • ...and 1 more figures