Kong Context Management Is GA: Fix MCP Tool Bloat Now

Kong shipped Context Management as generally available on October 14, adding a Code Mode that replaces upfront MCP tool definitions with a single SDK-like interface where agents retrieve schemas on demand at runtime. Kong cites a measured production case in which three MCP servers consumed 143,000 of a 200,000-token context window before a user typed anything, and notes each MCP tool definition averages 300 to 600 tokens. The feature builds on Kong AI Gateway 2.2, released September 30, whose MCP Server Bundling consolidates multiple MCP servers behind one route, and adds identity-aware policies tying rate limits, cost attribution and spending controls to authenticated users or agents via an existing OpenID Connect provider.

Three MCP servers. Before a user types a single character, 143,000 of your 200,000-token context window is already gone — burned on tool definitions your agent may never use. That’s not a worst-case scenario. That’s a measured production result, and it’s exactly why Kong just shipped Context Management as generally available https://sdtimes.com/context/kong-announces-context-management-for-efficient-ai-agent-use/ on October 14. The math is punishing at enterprise scale. Each MCP tool definition averages 300 to 600 tokens. Connect a realistic enterprise stack — five to ten MCP servers covering authentication, databases, CRMs, internal APIs — and you’ve routinely handed 60 to 80 percent of your agent’s entire reasoning budget to schema declarations before the conversation begins. The industry consensus in 2026 is that seven MCP servers is roughly the ceiling before agent reasoning starts degrading. Most enterprise environments blew past that limit months ago. What Kong’s Code Mode Actually Changes Standard MCP tool calling exposes every operation as a separate tool definition, all loaded at startup. Kong’s Code Mode flips this: instead of dumping the entire API surface into context, it presents a single SDK-like interface. The agent searches for what it needs at runtime, retrieves only the relevant schema on demand, and executes it. Nothing unnecessary enters the context window. There’s a compounding benefit. Because agents can compose multiple operations — including loops, conditionals, and error handling — without pre-loading every definition, the number of model round-trips drops for complex workflows. You get both a lighter context footprint and fewer back-and-forth calls to the model. Context Mesh https://developer.konghq.com/context-mesh/ , Kong’s earlier GA product, turns existing REST APIs and MCP servers into these Code Mode interfaces automatically, presenting them as a single governed entry point. The Governance Layer Matters More Than the Token Count Performance is the headline, but governance is the reason enterprises will actually adopt this. Kong AI Gateway 2.2, released September 30, laid the foundation: MCP Server Bundling consolidates multiple MCP servers behind a single Kong route and exposes only the tools a caller is authorized to invoke. Context Management extends this with identity-aware policies that tie rate limits, cost attribution, and spending controls to authenticated users or agents — not just gateway-level consumers. In practice, this means you can give your sales team’s agents access to a different tool surface than your engineering team’s agents, with separate spending budgets, all governed through your existing OpenID Connect provider. That kind of control didn’t exist cleanly before without custom middleware on every MCP server. This Is a Workaround, and That’s Fine Kong is solving the symptom. MCP’s design still requires tool definitions to be sent upfront in many implementations, and Code Mode is a gateway-layer patch that works around this rather than fixing it at the protocol level. That’s worth being clear about. It’s also not the only patch. Anthropic’s Claude Agent SDK has ENABLE TOOL SEARCH=auto:5 , which activates BM25-based tool search automatically when definitions exceed five percent of context — Speakeasy measured a 96.7 percent token reduction with this approach. Research-grade solutions https://arxiv.org/html/2604.21816v1 like the Tool Attention paper achieve 95 percent reduction using semantic gating and lazy schema loading. These approaches work client-side without Kong infrastructure. The open question is whether MCP’s specification will eventually add native lazy-loading or tool discovery at the protocol level. If it does, gateway solutions like Kong’s become less necessary for the context efficiency problem specifically — though the governance and identity layer remains valuable regardless. For now, Kong has the most complete enterprise story: context efficiency, tool authorization, identity policies, and cost attribution in a single platform. Who Should Move on This If you’re running enterprise APIs exposed through MCP to agents and hitting performance or cost walls, Context Management is the most production-ready option today. If you’re already on Kong, this is already in Kong AI Management https://konghq.com/blog/product-releases/kong-ai-gateway-2-2 — the barrier to turning it on is low. If you’re not on Kong, evaluate whether the governance layer justifies the infrastructure addition against client-side alternatives that require no new platform. For teams already at the five-to-seven MCP server ceiling, the cost-benefit case is straightforward: 60 to 80 percent of your context budget back, per-user tool governance, and spending controls that integrate with your IdP. That’s a real return.