MCP Context Manager

Instructions

Implementation Reference

• src/retriever.ts:126-165 (handler)

  Implementation of getRelevantContext method which performs the search and prepares the response.

  async getRelevantContext(query: string, maxTokens: number): Promise<string> {
    // Use advanced search engine
    const allSymbols = this.getAllSymbols();
    const files = this.indexer.getAllFiles();
  
    const searchOptions: SearchOptions = {
      maxResults: 50, // Get more candidates, then filter by tokens
      fuzzyThreshold: 0.7,
      includeCodeBodies: true,
      includeComments: true,
      boostExactMatches: true,
    };
  
    const matches = CodeSearchEngine.search(query, allSymbols, files, searchOptions);
  
    if (matches.length === 0) {
      // Provide helpful suggestions
      const queryTokens = CodeSearchEngine.tokenize(query);
      const suggestions = this.getSuggestions(queryTokens);
  
      return `No relevant context found for query: "${query}"\n\n${
        suggestions.length > 0
          ? `Did you mean one of these?\n${suggestions.map(s => `  - ${s}`).join('\n')}\n\n`
          : ''
      }Try:\n- Using more specific terms\n- Checking for typos (fuzzy search is active but needs some similarity)\n- Using search_code for regex pattern matching\n- Using find_symbol for exact symbol lookup`;
    }
  
    // Build response within token limit
    let result = `🔍 Found ${matches.length} relevant results for: "${query}"\n\n`;
    let estimatedTokens = this.estimateTokens(result);
    let included = 0;
  
    for (const match of matches) {
      const formattedMatch = this.formatSearchMatch(match, true);
      const matchTokens = this.estimateTokens(formattedMatch);
  
      if (estimatedTokens + matchTokens > maxTokens) {
        result += `\n... (${matches.length - included} more results available, but truncated to stay within ${maxTokens} token limit)`;
        break;
      }

• src/index.ts:144-152 (registration)

  Registration of the get_relevant_context MCP tool.

  name: 'get_relevant_context',
  description: '⭐ PREFERRED SEARCH: Use this INSTEAD OF Grep for finding code by description. Natural language search with BM25 ranking. Saves 70-90% tokens vs reading files directly. Examples: "authentication middleware", "payment processing", "gacha spin mechanics". Returns only relevant code snippets with relevance scores.',
  inputSchema: {
    type: 'object',
    properties: {
      query: {
        type: 'string',
        description: 'Natural language description of what you need (e.g., "user authentication flow", "error handling logic")',
      },