The paired tools are the practical hook: every research generator has a version that writes its Markdown to a path, so 'draft the project guidelines' or 'document this API' ends as a file rather than a chat message you copy out. Take the README's own security warning seriously before wiring it up — the filesystem-write and execute_terminal_command tools mean this is not a read-only research server, and it applies minimal model safety filters.
A server exposing 29 tools over Google Cloud's Vertex AI Gemini models, aimed at coding assistance: query answering with or without web search, documentation-grounded research, code and architecture analysis, plus filesystem and terminal tools. Every model tool also has a save_ variant that writes its output to a path.
- answer_query_websearch answers with the model grounded in Google Search results, and answer_query_direct answers from the model's own knowledge alone
- Documentation-grounded research: explain_topic_with_docs synthesises official docs found by web search, and get_doc_snippets returns precise code snippets without the surrounding explanation
- Code and design review — code_analysis_with_docs checks a snippet against documented best practices for bugs and vulnerabilities, technical_comparison builds pros/cons tables, architecture_pattern_recommendation and microservice_design_assistant suggest structure, and database_schema_analyzer reviews a schema for normalization and indexing
- Focused advisors: dependency_vulnerability_scan, security_best_practices_advisor, testing_strategy_generator, regulatory_compliance_advisor and documentation_generator, each grounded via web search
- generate_project_guidelines turns a tech stack into a structured guidelines document from the latest official docs
- A save_ twin for every generator — save_answer_query_websearch, save_doc_snippet, save_topic_explanation, save_generate_project_guidelines and the rest — writing the result straight to an output_path instead of returning it inline
- Direct filesystem access: read_file_content and write_file_content (single or batch), edit_file_content with exact-then-whitespace-tolerant matching, list_directory_contents, get_directory_tree, move_file_or_directory, search_filesystem and get_filesystem_info
- execute_terminal_command runs a shell command with an optional working directory and timeout, returning combined stdout and stderr
A Gemini backend chosen with AI_PROVIDER: set it to vertex and provide GOOGLE_CLOUD_PROJECT with the Vertex AI API enabled (authenticating through ADC or credentials), or set it to gemini and provide GEMINI_API_KEY. Node, run as npx vertex-ai-mcp-server. Note the security boundary the README flags: this server bundles filesystem-write and terminal-execution tools, so run it under a dedicated low-privilege identity, and its safety filters are set to BLOCK_NONE.
One command plus a key — bunx vertex-ai-mcp-server, then supply credentials