# mcp-compressor wrapper proxy in front of a local-model agent > The model first sees a small compressed surface, asks for the full schema of one tool, then invokes it. Parent: [Mac local LLMs: Agent clients, context and compaction](https://llms-explorer.com/tree/mac-local-llms-agent-clients-context-and-compaction/) · 1 facets · 6 facts · page: https://llms-explorer.com/tree/mcp-compressor-wrapper-proxy-in-front-of-a-local/ ## Facts - The model first sees a small compressed surface, asks for the full schema of one tool, then invokes it. — source: `asserted` - Pattern 1 is a wrapper in front of any MCP client; pattern 2 is a local session proxy for SDKs and generated clients, which call it with a session token and need no stdio subprocess. — source: `asserted` - No source found measures a local small model choosing tools through the proxy; the existing open question stands. — source: `asserted` - Task success of a 9B to 35B local model through get_tool_schema then invoke_tool round trips. — source: `asserted` - The mcp-compressor docs describe a wrapper pattern and a local session proxy pattern with session tokens — [source](https://atlassian-labs.github.io/mcp-compressor/) - No source read gives a measurement of mcp-compressor with a local model; the claim stays open — source: `asserted`