ACL Findings2026
ProMCP: Profiling Token Flows and Latency Costs in Model Context Protocol-Based LLM Agents
Sumera Anjum, Weijian Zheng, Rajkumar Kettimuthu, Heng Fan, Yunhe Feng
Abstract
The Model Context Protocol (MCP) aims to standardize the integration of Large Language Models (LLMs) with external tools, yet existing research primarily evaluates functional capabilities while treating the underlying protocol as an opaque black box. This oversight obscures critical inefficiencies in token flows and latency distributed across MCP's decoupled Host-Client-Server architecture. In this paper, we introduce ProMCP, an end-to-end profiling and instrumentation framework that decomposes the MCP workflow into a sixstage communication pipeline, enabling granular attribution of computational costs. We evaluate widely varying deployment topologies, from local LLM models to commercial off-the-shelf (OTS) clients-across 20 servers and 169 tools from MCP-Bench and MCP-Universe. Our analysis reveals a distinct inversion in performance bottlenecks: topologies with customized clients devote 56-72% of total tokens and 60-67% of latency to planning and schema injection, whereas OTS clients concentrate over 75-85% of latency in final answer synthesis. Crucially, actual tool execution constitutes a negligible fraction of the total cost across all configurations. These findings establish a quantitative baseline for protocol overhead and demonstrate that future optimization must target schema orchestration and transport efficiency rather than tool execution speed. The code is available at: https: //github.com/ResponsibleAILab/ProMCP.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
- MCP-Bench: Benchmarking Tool-Using LLM Agents with Complex Real-World Tasks via MCP Servers cites ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs
- MCIP: Protecting MCP Safety via Model Contextual Integrity Protocol cites Toolformer: Language Models Can Teach Themselves to Use Tools
- ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs cites Toolformer: Language Models Can Teach Themselves to Use Tools
Builds on7
- Toolformer: Language Models Can Teach Themselves to Use ToolsTimo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu et al.NeurIPS 2023 · 5,989 citations
- ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIsYujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu et al.ICLR 2024 · 1,469 citations
- MCP-Bench: Benchmarking Tool-Using LLM Agents with Complex Real-World Tasks via MCP ServersZhenting Wang, Qi Chang, Hemani Patel, Shashank Biju et al.ICLR 2026 · 109 citations
- Validation of Modern JSON Schema: Formalization and ComplexityLyes Attouche, Mohamed-Amine Baazizi, Dario Colazzo, Giorgio Ghelli et al.POPL 2024 · 19 citations
- MCIP: Protecting MCP Safety via Model Contextual Integrity ProtocolHuihao Jing, Haoran Li, Wenbin Hu, Qi Hu et al.EMNLP 2025 · 3 citations
Related papers
- MCP Security Bench (MSB): Benchmarking Attacks Against Model Context Protocol in LLM AgentsDongsen Zhang, Zekun Li, Xu Luo, Xuannan Liu et al.ICLR 2026 · 47 citations
- Parasites in the Toolchain: A Large-Scale Analysis of Attacks on the MCP EcosystemShuli Zhao, Qinsheng Hou, Zihan Zhan, Yanhao Wang et al.S&P 2026 · 20 citations
- MCP-Flow: Facilitating LLM Agents to Master Real-World, Diverse and Scaling MCP ToolsWenhao Wang, Peizhi Niu, Zhao Xu, Zhaoyu Chen et al.ACL 2026 · 8 citations
- MCP-Guard: A Multi-Stage Defense-in-Depth Framework for Securing Model Context Protocol in Agentic AIWenpeng Xing, Zhonghao Qi, Yupeng Qin, Yilin Li et al.ACL Findings 2026 · 2 citations
- Tool Preferences in Agentic LLMs are UnreliableKazem Faghih, Wenxiao Wang, Yize Cheng, Siddhant Bharti et al.EMNLP 2025
