You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
When using MCP servers in Jan Desktop with local GGUF models (llama.cpp / Cortex), large MCP servers like GitHub or Playwright can easily consume 10,000 to 30,000+ context tokens just for tools/list definitions, and local 7B to 14B models frequently struggle with unresolved JSON Schema $ref / $defs pointers or emit stringified integers ("page": "1").
We built mcp-slim, a zero-dependency Python MCP stdio proxy that wraps any MCP command in Jan:
Key improvements for local models in Jan
50% to 98.7% Token Compaction: Minifies tools/list schemas, compacts JSON array outputs into TOON tables (--toon), and supports lazy BM25 tool discovery (--dynamic-toolsets) so 80+ tools collapse into search_tools and call_tool.
Inlines $ref / $defs (--flatten-schemas): Flattens Pydantic/FastMCP JSON schemas into flat inline properties that local GGUF grammars handle cleanly.
Automatic Argument Type Coercion (--coerce-args): Converts "10" -> 10, "true" -> true, and JSON-encoded array strings into native arrays before calling the upstream server.
EmbeddedResource Unwrapping & SecretGuard (--unwrap-resources, --redact-secrets): Unwraps type: "resource" file blocks into plain text and redacts leaked credentials in real time.
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Hi Jan community,
When using MCP servers in Jan Desktop with local GGUF models (
llama.cpp/ Cortex), large MCP servers like GitHub or Playwright can easily consume 10,000 to 30,000+ context tokens just fortools/listdefinitions, and local 7B to 14B models frequently struggle with unresolved JSON Schema$ref/$defspointers or emit stringified integers ("page": "1").We built mcp-slim, a zero-dependency Python MCP stdio proxy that wraps any MCP command in Jan:
Key improvements for local models in Jan
tools/listschemas, compacts JSON array outputs into TOON tables (--toon), and supports lazy BM25 tool discovery (--dynamic-toolsets) so 80+ tools collapse intosearch_toolsandcall_tool.$ref/$defs(--flatten-schemas): Flattens Pydantic/FastMCP JSON schemas into flat inline properties that local GGUF grammars handle cleanly.--coerce-args): Converts"10"->10,"true"->true, and JSON-encoded array strings into native arrays before calling the upstream server.EmbeddedResourceUnwrapping &SecretGuard(--unwrap-resources,--redact-secrets): Unwrapstype: "resource"file blocks into plain text and redacts leaked credentials in real time.Repository: https://github.com/SalyyS1/mcp-slim
All reactions