The draw is precision: call hierarchy and inspections come from the IDE's index, so an agent asks one question where it would otherwise grep four times. The setting to look at before you enable it is brave mode, which lets connected clients run terminal commands and run configurations without asking you each time.
The MCP server built into JetBrains IDEs from version 2025.2. External clients such as Claude Code, Codex or VS Code connect to the running IDE and use what it knows about your project: the call hierarchy, the inspections, the debugger, the VCS state.
- Call hierarchy through analyze_calls — who calls a symbol and what it calls, from IDE data rather than a text search
- Inspections and compile errors via get_file_problems, lint_files and build_project
- Refactoring the IDE performs rather than text replacement: rename_refactoring
- A debugger an agent can drive — xdebug_set_breakpoint, xdebug_evaluate_expression, xdebug_get_stack
- Database work over the IDE's own connections: introspect_schema, execute_sql_query, preview_table_data
- Symbol, text, regex and file search, plus execute_terminal_command and git_status
- Per-tool switches under Settings | Tools | MCP Server | Exposed Tools, and a router-only mode that keeps unused tool descriptions out of the client's context
A JetBrains IDE at 2025.2 or newer with the bundled MCP Server plugin left enabled, and the server switched on under Settings | Tools | MCP Server. Detected clients configure themselves from that screen; anything else takes a copied SSE, stdio or HTTP stream config.
One command — npx -y @jetbrains/mcp-proxy