Java Inspector

Java Inspector is an MCP server that gives AI agents fast, read-only access to code and metadata inside Maven dependencies, local JARs, executable JARs, WARs, and the JDK.

The server starts accepting MCP calls immediately. Maven dependency scanning continues in the background, while an interactive query can promote the JAR it needs to a foreground queue and return without waiting for the complete project index.

Requirements

• Node.js 16 or newer

• Java 17 or newer (java and javap); required by bundled Vineflower 1.11.2

• Maven for Maven project discovery and dependency resolution

Vineflower 1.11.2 is bundled in lib/.

Related MCP server: MCP Java Decompiler Server

Install and configure

npm install -g @mustafagoksever/java-inspector

{
  "mcpServers": {
    "java-inspector": {
      "command": "java-inspector",
      "args": ["start"]
    }
  }
}

For local development:

npm ci
npm run build
node dist/cli.js start

Startup discovery

After the MCP transport connects, Java Inspector performs non-blocking discovery:

1. If cwd/pom.xml exists, that Maven project is used.

2. Otherwise directories are searched breadth-first.

3. Every pom.xml at the first matching depth is selected.

4. Deeper POMs are not independently scanned.

5. Maven classpaths are resolved and their JARs enter the background light-scan queue.

Directories such as .git, node_modules, .gradle, target, build, and out are excluded from POM discovery.

Local, non-Maven JARs are not scanned at startup. They are found and indexed only when a tool call supplies a path, directory, or filename prefix.

Tools

Every tool supports text, json, and toon output where applicable.

Artifact selectors

Class, JAR, and resource tools share the same selection model:

• jarPath: exact absolute JAR path; fastest and unambiguous.

• jarDirectory + jarNamePrefix: useful for Maven-free projects with lib/ directories.

• workspacePath + jarNamePrefix: search the Maven classpath and workspace.

• workspacePath: use the live partial index and bounded lazy fallback.

• coordinates: groupId:artifactId[:version[:classifier]].

When no selector is supplied, tools use the server's current working directory. If Maven classpath resolution is still running, searches return complete: false instead of reporting a definitive miss. Per-JAR scan failures are returned in errors; an unreadable exact jarPath fails the tool call.

If multiple JARs or classes match, the server returns candidates instead of silently choosing one.

Inspect a class from a local JAR

{
  "className": "com.vendor.Client",
  "jarPath": "C:\\project\\lib\\vendor-client-2.1.jar",
  "view": "source"
}

Find an exception message

{
  "query": "Connection refused",
  "kind": "string",
  "workspacePath": "C:\\project",
  "mode": "balanced"
}

Find implementations

{
  "className": "com.vendor.Transport",
  "workspacePath": "C:\\project",
  "transitive": true
}

Inspect a JDK class

{
  "className": "java.util.ArrayList",
  "view": "all"
}

JDK source is read from src.zip when available. API and bytecode views use javap.

Foreground and background scheduling

All JAR reads pass through a shared coordinator:

• Up to eight JAR reads run concurrently.

• Background scanning may occupy at most six slots.

• Two slots remain available for interactive foreground work.

• A foreground request promotes a queued background JAR.

• Concurrent requests for the same JAR share one Promise and one cache write.

• Light scans time out after 30 seconds; deep scans time out after 60 seconds.

The background scan builds a light index from ZIP entries. More expensive class-file metadata is generated only when search_code or find_implementations needs it.

Context-wide deep searches are intentionally bounded. A partial response includes complete, scannedJarCount, and remainingJarCount. Calling the tool again continues using the already cached deep indexes.

Class inspection

inspect_class supports:

• source: local *-sources.jar, embedded .java, or Vineflower decompilation.

• api: fields, methods, constructors, modifiers, superclass, and interfaces via javap.

• hierarchy: resolved superclass chain.

• bytecode: javap -c -l -p -s output.

• all: all available views.

Source responses support methodName, paramTypes, offset, and limit.

The decompiler executable cannot be overridden by a tool argument. Java Inspector uses the bundled Vineflower JAR or the operator-controlled DECOMPILER_PATH environment variable.

JAR layouts

Light indexing understands:

• Ordinary and shaded JARs

• Multi-release entries under META-INF/versions/

• Spring Boot classes under BOOT-INF/classes

• WAR classes under WEB-INF/classes

• JMOD class paths under classes/

• Nested artifact metadata under BOOT-INF/lib, WEB-INF/lib, and lib

• Maven pom.properties, manifests, source JARs, and text resources

Nested JAR contents are extracted and scanned only when explicitly needed.

Cache

Cache data is stored under:

~/.cache/java-inspector/<context>_<hash>/

Important entries include:

classpath.json
class-index.jsonl
scan-state.json
jar-indexes-v3/
results-v3/
server-<pid>.log

Per-JAR caches use resolved path, size, and modification time as their fingerprint. Nested artifacts additionally use their parent fingerprint and ZIP entry identity. JSONL project indexes remain append-only and crash-safe, with cross-process locks protecting writes.

Environment variables

Maven command resolution order is MAVEN_CMD, mvnd, MAVEN_HOME/bin/mvn, then mvn from PATH.

Development

npm run build
npm test
npm run dev

The project is ES modules and emits JavaScript, declarations, and source maps to dist/.

License

Apache-2.0