Code Intelligence MCP Server — Joern CPG + ArangoDB powered code analysis for AI agents. 14 tools: symbol search, call graphs, data flow, impact analysis, React components, and more.
15 tools with mostly complete schemas and descriptions, but significant gaps in parameter documentation and output schema clarity. Tool names follow verb_noun convention well (get_, find_, list_, etc.). Descriptions are present but often lack actionable context for LLM selection. Many parameters lack type constraints (enums, min/max). Output schemas are not documented, LLMs cannot predict response structure. Error handling is minimal. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels. Composition is sound (single responsibility per tool), but parameter relationships are underdocumented.
Clear the query cache for a project — forces fresh queries from ArangoDB.
Find all usages of a symbol — where it's imported, called, or referenced. Includes file:line for each usage.
Trace the call path from function A to function B. Like running a debugger trace without running the code. Shows the exact execution route.
grep -A 50 'function login(' src/ → USE THIS INSTEAD. Shows everything a function calls across the whole dependency tree. For React components, use get_react_components instead.
grep -rn 'functionName(' src/ → USE THIS INSTEAD. Shows every caller of a function up to N levels deep — follows the full call chain, not just string matches.
Get the full code context around a symbol — function body, class definition, variable initialization. Returns the actual source code.
Output schemas not documented. LLMs cannot predict response structure for any tool. Responses from symbol_search, get_callers, get_code_context, etc. lack field definitions, types, and pagination info. This forces LLMs to guess at response shape and breaks downstream tool chaining.
Missing tool annotations. All tools lack readOnlyHint, destructiveHint, or idempotentHint. index_project, watch_project, and clear_project_cache are WRITE operations but carry no annotation. Agents cannot distinguish safe reads from destructive writes.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Render tree of a React component — shows parent→child relationships, props passed down, and re-render triggers.
grep -rn 'userId' src/ → USE THIS INSTEAD for tracking data. Traces how a variable flows through the code — forward to sinks, backward to sources. Answers: where does this value come from and where does it end up?
MUST RUN before changing shared code. grep -rn 'import.*foo' src/ → USE THIS INSTEAD. Shows blast radius — every file that breaks if this symbol changes, following imports and calls transitively.
Get detailed status of an indexed project — file count, node count, last indexed time, and any errors.
Find React components by name. Returns component definition, props interface, and all child components it renders.
REQUIRED FIRST — parse project code into the graph DB. Run once per project, then all query tools work. Incremental: only re-parses files with changed SHA256 hashes. Like 'git init' for code search.
List all indexed projects and their status — shows which projects are ready for querying.
grep -r 'functionName' src/ → USE THIS INSTEAD. Find any function, class, method, or variable by name (partial match, exact file:line location).
Enable live file watching for a project — automatically re-indexes when files change. Incremental updates only.
Parameter constraints incomplete. nodeType enum in symbol_search is good, but depth/maxDepth in get_callers, get_callees, get_call_chain lack min/max bounds in descriptions. contextLines in get_code_context has min/max in schema but no explanation in description. maxResults in get_impact_analysis has bounds but no rationale.
Descriptions lack actionable context. Many descriptions (get_callers, get_callees, find_usages) state WHAT the tool does but not WHEN to use it vs similar tools or what prerequisites exist. 'Shows every caller of a function' doesn't explain when to call get_callers vs get_impact_analysis. LLMs cannot disambiguate.
No error handling guidance. Tools return no documented error cases, recovery steps, or actionable messages. If index_project fails (Joern parse error, ArangoDB connection lost), agents have no guidance on retry vs user intervention. If symbol_search finds nothing, no suggestion to broaden query.