agentailor/slimcontext-mcp-server

Summary

A thin MCP server wrapper around the SlimContext npm library that exposes two chat history compression tools: token-based trimming (drop oldest messages) and AI-powered summarization via OpenAI. It's essentially a bridge so MCP-compatible clients like Claude Desktop can call SlimContext's compression algorithms as tools. The project is a single-author, single-commit repo at v0.1.2 with ~150 lines of meaningful source code.

Great for

people interested in context window management for LLM applications, specifically implementing chat history compression strategies within the MCP ecosystem

Easy wins

• +Add tests — package.json literally has 'echo No tests configured' as the test script; even basic Jest/Vitest unit tests for the trim and summarize handlers would be a meaningful PR
• +Fix the model name mismatch bug: types/index.ts defaults openaiModel to 'gpt-5-mini' but src/index.ts defaults it to 'gpt-4o-mini' — these are inconsistent and 'gpt-5-mini' doesn't exist as a public model
• +Remove the dead import in src/index.ts: `import ru from 'zod/v4/locales/ru.js'` is imported but never used — this is a clear leftover artifact
• +Add support for non-OpenAI models (Anthropic, local Ollama) in the summarize tool — the OpenAISlimContextModel adapter pattern in summarize.ts is already cleanly separated and would make the tool more useful

Red flags

• !Model name bug: types/index.ts defaults openaiModel to 'gpt-5-mini' (non-existent public model) while index.ts uses 'gpt-4o-mini' — whichever schema default wins will silently fail or use the wrong model
• !Dead import: `import ru from 'zod/v4/locales/ru.js'` in src/index.ts suggests AI-assisted code generation left in a stray import that has no purpose
• !No license file despite README and package.json claiming MIT — the repo metadata shows license: null
• !Zero tests with a test script that just echoes a message — the project has no validation that the compression logic actually works end-to-end
• !Single commit history means there's no way to understand design decisions or evolution of the codebase
• !@langchain/core is listed as a dependency in package.json but is never imported in any source file — unnecessary ~50MB dependency

Code quality

The architecture is clean — tools are properly separated in src/tools/, types are centralized with Zod validation, and error handling returns structured JSON consistently rather than throwing. However there are real bugs: the Russian locale import (`import ru from 'zod/v4/locales/ru.js'`) in index.ts is dead code that will cause a module resolution error depending on zod version, the tool callback in index.ts double-parses by calling handleTrimMessages then JSON.parse(result.content[0].text) which is a clunky round-trip, and the default model inconsistency between types/index.ts ('gpt-5-mini') and index.ts ('gpt-4o-mini') is a real functional bug.

What makes it unique

This is a straightforward glue-layer project — it doesn't implement any novel compression algorithms, it just wraps the separate 'slimcontext' npm package as MCP tools. The value proposition is convenience for MCP users who don't want to integrate slimcontext directly. There are a handful of similar context-management MCP servers, but this specific wrapping of slimcontext is unique. Worth noting: the actual compression logic lives entirely in the slimcontext dependency, so contributing here means contributing to the adapter/interface layer, not the algorithms.