llm-kb
← Back to social
Community

User reports tool-selection accuracy drops linearly with more MCP servers due to token bloat

A user connected 4 MCP servers to one agent and observed tool-selection accuracy declining linearly with server count. They traced the issue to every tool's name, description, and JSON Schema being serialized into every request, causing token bloat. With 4 servers and ~60 tools, the serialized definitions consumed significant context, degrading performance.

11 engagement·1 source·Sat, Jul 11, 2026, 08:49 PM
On July 11, 2026, a Reddit user reported that connecting multiple MCP servers to a single agent caused tool-selection accuracy to drop consistently and almost linearly with server count. Investigating, they found that every tool exposed by an MCP server gets serialized into the request on every turn, including the tool name, description, and the entire JSON Schema for its inputs with descriptions for every property. With 4 servers, they ended up with around 60 tools, and the serialized tool definitions consumed a significant portion of the context budget, degrading performance. This highlights a practical challenge in scaling MCP-based agent systems.

Entities

MCP(concept)tool-selection accuracy(concept)token bloat(concept)Reddit(company)

Related

CommunitySat, Jul 11, 2026, 06:01 AM

Developer reports Claude performance degrades beyond 500K tokens, uses handover files to reset context

A developer working on a large C# plugin reports that Claude becomes increasingly unreliable after 500K tokens of context, with performance degrading significantly by 800-900K tokens. To work around this, they manually create a handover.md file summarizing the project state and start a new session, asking Claude to re-read key files. The post questions whether this is common practice and whether tools exist to automate the process.

7 engagement·1 source·reddit
IncidentSat, Jul 11, 2026, 11:18 AM

Users report Claude Code consuming excessive usage due to large context bug

Multiple users report that Claude Code consumes 20-40% of their usage allowance per session due to a bug that sends 300k+ message contexts even on fresh sessions. One user lost significant paid usage before realizing the issue. Another user describes a layered configuration setup with CLAUDE.md and AGENTS.md files, and a cleanup that removed ~700MB of empty content but still faced model selection inefficiencies.

15 engagement·2 sources·reddit
CommunitySat, Jul 11, 2026, 09:47 PM

MCP's OAuth scoping undermined by static API key proxies

A developer reports that many agent CLIs still store static API keys in environment variables, creating a large blast radius if leaked. The Model Context Protocol (MCP) aims to fix this with scoped OAuth consent per tool, but half of the servers examined simply proxy a static key, defeating the purpose.

5 engagement·1 source·reddit
CommunitySun, Jul 12, 2026, 03:19 PM

Developer reports 534K-token single turn in 1,320-turn agent session on 1M-context model

A developer shared that one turn in a 1,320-turn agent session consumed 534,000 tokens, 2.7x the typical 200K context ceiling. The session remained coherent only because it ran on a 1M-token window; on a 200K model it would have fragmented. The setup uses a long-running agent coordinating subagents.

3 engagement·1 source·reddit
CommunitySun, Jul 12, 2026, 12:24 PM

Users report tool-call failures and looping in Qwen3.6-27B

A Reddit user reports persistent tool-call failures and looping behavior when using Qwen3.6-27B as a local model. The user describes needing constant monitoring to prevent the model from entering infinite loops or hallucinating tool calls, and has developed an extensible workaround to mitigate these issues.

18 engagement·2 sources·reddit