javalens-mcp

# JavaLens: AI-First Code Analysis for Java

[![GitHub release](https://img.shields.io/github/v/release/pzalutski-pixel/javalens-mcp)](https://github.com/pzalutski-pixel/javalens-mcp/releases)
[![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)
[![Java 21](https://img.shields.io/badge/Java-21-orange.svg)](https://openjdk.org/projects/jdk/21/)

An MCP server providing 75 semantic analysis tools for Java, built directly on Eclipse JDT for compiler-accurate code understanding.

## Built for AI Agents

JavaLens exists because **AI systems need compiler-accurate insights that reading source files cannot provide**. When an AI uses `grep` or `Read` to find usages of a method, it cannot distinguish:

- A method call from a method with the same name in an unrelated class
- A field read from a field write
- An interface implementation from an unrelated class
- A cast to a type from other references to that type

This leads to incorrect refactorings, missed usages, and incomplete understanding of code behavior.

## Compiler-Accurate Analysis

JavaLens provides **compiler-accurate code analysis** through Eclipse JDT—the same engine that powers Eclipse IDE. Unlike text search, JDT understands:

- Type resolution across inheritance hierarchies
- Method overloading and overriding
- Generic type arguments
- Import resolution and classpath dependencies
- Java source from version 1.1 through **Java 25** (markdown Javadoc, module imports, compact source files, flexible constructor bodies)
- **Lombok**-generated members — a bundled agent makes `@Data` accessors and the like resolve, so code using them is not flagged as undefined

**Example:** Finding all places where `UserService.save()` is called:

| Approach | Result |
|----------|--------|
| `grep "save("` | Returns 47 matches including `orderService.save()`, `saveButton`, comments |
| `find_references` | Returns exactly 12 calls to `UserService.save()` |

## AI Training Bias Warning

> ⚠️ **Important for AI developers and users**

AI models may exhibit **trained bias toward native tools** (Grep, Read, LSP) over MCP server tools, even when semantic analysis provides better results. This happens because:

1. Training data contains extensive grep/text-search patterns
2. Native tools are "always available" in the model's experience
3. The model may not recognize when semantic analysis is superior

**To get the best results:**

Add guidance to your project instructions or system prompt (e.g., `CLAUDE.md` for Claude Code):

```markdown
## Code Analysis Preferences

For Java code analysis, prefer JavaLens MCP tools over text search:
- Use `find_references` instead of grep for finding usages
- Use `find_implementations` instead of text search for implementations
- Use `analyze_type` to understand a class before modifying it
- Use refactoring tools (rename_symbol, extract_method) for safe changes

Semantic analysis from JDT is more accurate than text-based search,
especially for overloaded methods, inheritance, and generic types.
```

## What is JavaLens?

JavaLens is an MCP server that gives AI assistants deep understanding of Java codebases. It provides semantic analysis, navigation, refactoring, and code intelligence tools that go beyond simple text search.

## Why Not LSP?

Language Server Protocol was designed for IDE autocomplete and basic navigation—not for AI agent workflows that require deep semantic analysis.

| Capability | Native LSP | JavaLens |
|------------|------------|----------|
| Find all `@Annotation` usages | ❌ | ✅ |
| Find all `new Type()` instantiations | ❌ | ✅ |
| Find all casts to a type | ❌ | ✅ |
| Distinguish field reads from writes | ❌ | ✅ |
| Detect circular package dependencies | ❌ | ✅ |
| Calculate cyclomatic complexity | ❌ | ✅ |
| Find unused private methods | ❌ | ✅ |
| Detect possible null pointer bugs | ❌ | ✅ |
| Project-wide dead-code reachability from entry points | ❌ | ✅ |
| Find the tests that exercise a symbol, transitively | ❌ | ✅ |

JavaLens wraps **Eclipse JDT Core** directly via OSGi, providing:

- **Fine-grained reference types**: Find specifically casts, annotations, throws clauses, catch blocks, instanceof checks, method references, type arguments
- **Read vs write access distinction**: Track where fields are mutated vs just read
- **Indexed search**: JDT pre-builds an index at load time, so symbol/reference queries do not walk source files
- **Full AST access**: Direct manipulation for complex refactorings

## Installation

### Prerequisites

- **Java 21** or later (must be on PATH or set `JAVA_HOME`) — required for both install paths.
- **Node.js 18+** — required *only* if you use the npm/`npx` install path below. Skip if you use the direct-download path.

JavaLens is an analytical server, not a compiler. It uses Eclipse JDT 2025-12 to parse and understand Java source code from **version 1.1 through 25**. Java 21 is required only as the server runtime.

### Install from GitHub Releases (recommended — Java only)

This is the simplest path if you already have Java 21 and don't have Node.js. Download from [Releases](https://github.com/pzalutski-pixel/javalens-mcp/releases):

| Platform | File |
|----------|------|
| All platforms | `javalens.zip` or `javalens.tar.gz` |

Extract to a location of your choice (e.g., `/opt/javalens` or `C:\javalens`). Then point your MCP client at the bundled jar — see [Configure MCP Client](#configure-mcp-client) below.

### Install via npm (requires Node.js 18+)

If you already have Node.js, `npx` will download and cache the JavaLens distribution (~23 MB) on first run:

```json
{
  "mcpServers": {
    "javalens": {
      "command": "npx",
      "args": ["-y", "javalens-mcp"],
      "env": {
        "JAVA_PROJECT_PATH": "/path/to/your/java/project"
      }
    }
  }
}
```

### Configure MCP Client

Add to your MCP configuration (e.g., `.mcp.json` for Claude Code):

```json
{
  "mcpServers": {
    "javalens": {
      "command": "java",
      "args": ["-jar", "/path/to/javalens/javalens.jar", "-data", "/path/to/javalens-workspaces"]
    }
  }
}
```

The `-data` argument specifies where JavaLens stores its workspace metadata. See [How Workspaces Work](#how-workspaces-work) below.

### Auto-Load a Project

Set `JAVA_PROJECT_PATH` to auto-load a project when the server starts:

```json
{
  "mcpServers": {
    "javalens": {
      "command": "java",
      "args": ["-jar", "/path/to/javalens/javalens.jar", "-data", "/path/to/javalens-workspaces"],
      "env": {
        "JAVA_PROJECT_PATH": "/path/to/your/java/project"
      }
    }
  }
}
```

> **Note:** Project loading happens asynchronously in the background. The MCP server responds immediately while the project loads. Use `health_check` to monitor loading status—it will show `"project.status": "loading"` until complete, then `"loaded"` when ready.

## How Workspaces Work

Unlike in-memory code models, Eclipse JDT requires a **workspace directory** to store:

- Search indexes for fast symbol lookup
- Compilation state and caches
- Project metadata

### Workspaces Are Outside Your Source

JavaLens creates its workspace **outside your source project** to keep your codebase clean:

```
Your Java Project (unchanged)
├── src/main/java/
├── pom.xml
└── (no Eclipse files added)

JavaLens Workspace (specified by -data)
└── {session-uuid}/
    ├── .metadata/          <- JDT indexes and state
    └── javalens-project/   <- Links to your source (not copies)
```

**Why this matters:**

1. **No pollution**: Your source tree stays clean—no `.project` or `.classpath` files
2. **No conflicts**: Works alongside any build system without interference
3. **Session isolation**: Each MCP session gets its own workspace, enabling concurrent analysis

### Session Lifecycle

1. JavaLens starts and creates a unique workspace: `{base}/{uuid}/`
2. `load_project` creates linked folders pointing to your source
3. JDT builds indexes in the workspace (not in your project)
4. When the session ends, the workspace is cleaned up

## Tools

### Navigation (10 tools)

| Tool | Description |
|------|-------------|
| `search_symbols` | Search types, methods, fields by glob pattern |
| `go_to_definition` | Navigate to symbol definition |
| `find_references` | Find all usages of a symbol |
| `find_implementations` | Find interface/class implementations |
| `get_type_hierarchy` | Get inheritance chain |
| `get_document_symbols` | Get all symbols in a file |
| `get_symbol_info` | Get detailed symbol information at position |
| `get_type_at_position` | Get type details at cursor |
| `get_method_at_position` | Get method details at cursor |
| `get_field_at_position` | Get field details at cursor |

### Fine-Grained Reference Search (9 tools)

These use JDT's unique reference type constants—not available through LSP:

| Tool | Description |
|------|-------------|
| `find_annotation_usages` | Find all `@Annotation` usages |
| `find_type_instantiations` | Find all `new Type()` calls |
| `find_casts` | Find all `(Type) expr` casts |
| `find_instanceof_checks` | Find all `x instanceof Type` |
| `find_throws_declarations` | Find all `throws Exception` in signatures |
| `find_catch_blocks` | Find all `catch(Exception e)` blocks |
| `find_method_references` | Find all `Type::method` expressions |
| `find_type_arguments` | Find all `List<Type>` usages |
| `find_reflection_usage` | Find `Class.forName()`, `Method.invoke()`, and other reflection calls |

### Analysis (20 tools)

| Tool | Description |
|------|-------------|
| `get_diagnostics` | Get compilation errors and warnings |
| `validate_syntax` | Fast syntax-only validation |
| `get_call_hierarchy_incoming` | Find all callers of a method |
| `get_call_hierarchy_outgoing` | Find all methods called by a method |
| `find_field_writes` | Find where fields are mutated |
| `find_tests` | Discover JUnit/TestNG test methods |
| `find_unused_code` | Find unused private members |
| `find_unreachable_code` | Project-wide dead code — members unreach