MCP Compression: Preventing tool bloat in AI agents
Mar 30, 2026Tim Esler

Summary
Introduced mcp-compressor, an open-source MCP proxy that wraps existing MCP servers to reduce tool-description token usage by 70-97%. The proxy replaces a server's full tool inventory with generic wrapper tools (get_tool_schema, invoke_tool, optionally list_tools) to enable on-demand schema fetching. This reduces prompt costs, improves cache behavior, and scales to multiple servers. The post details configurable compression levels and their impact on token counts, and discusses design principles like preserving structure, namespacing, cache-friendliness, and explicit discovery paths. It also contrasts this approach with code-mode execution.
More from the AI-Native Software Development Capabilities in Jira trail
- A simpler way to run AI agents in Bitbucket Pipelines
- We’re bringing governed agent loops to the AI-Native SDLC
- How we made vulnerability fixes review-ready with Agentic Pipelines
- Agentic automation in practice: putting standard engineering work on autopilot
- Opening the door to agent autonomy: The architecture behind Rovo’s agent harness