
JVM Source Lens
io.github.Sintexer由 GitHub 账号 Sintexer 发布v1.16.0更新于 Oct 24, 2026
Gradle-accurate JVM classpath; fetch Java source, signatures, and class structure for agents.
发布方自己的服务:Source
概览
本地 MCP 服务器,为编码助手提供与 Gradle 一致的 JVM 类路径查询:按项目实际依赖版本搜索类、查看结构、签名和源码
- 可以做什么
- 按简单名或通配符搜索已解析依赖中的类,返回全限定名和所属库,并分析项目实际使用的依赖图与各模块版本。获取类的概览或已声明签名,并取得方法的真实重载、参数名和泛型信息。在已解析的类内做正则或子串搜索,或在单个依赖 JAR 的所有类中检索文本。作为最后手段读取方法体或行范围源码,并标明是真实源码还是 CFR 反编译结果。
- 适用场景
- 适用于使用 Gradle 构建的 Java 和 Spring Boot 项目,尤其是第三方或内部库没有公开 Javadoc、助手容易猜错版本或翻查 Gradle 缓存的情况。它返回签名和结构而非整份文件,有助于控制上下文长度。
- 运行要求
- 通过 npx 以 stdio 在本地运行。需要 Node.js 20 或更高版本,以及 PATH 中的 Java(用于 CFR 反编译器和 javap)。项目必须是 Gradle 构建(不支持 Maven 和 Bazel);对项目的首次调用会运行 Gradle,耗时数秒。可选环境变量用于限定允许的项目根目录、缓存与日志目录、自定义 CFR JAR、Gradle 超时和源码输出上限。
安装
在 SourceWeft 中
- 打开 控制台中的 JVM Source Lens,将其添加到工作区。
- 为需要使用其工具的对话启用该服务。
Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。
其他 MCP 客户端
参照 仓库 中的启动说明。
README
jvmsrc — Give your coding agent a Java IDE
[CI] [npm version] [License: MIT] [Node]
An MCP server and CLI that gives your coding agent the one thing it's missing on JVM codebases: the actual classpath.
The Problem
You use an IDE to write Java. Your coding agent doesn't have one.
When your agent hits an unfamiliar library type — say, a superclass from a proprietary internal library — it spends 25+ turns walking ~/.gradle/caches, opening JARs by hand with jar tf, picking one by guesswork, and trying to answer a question your IDE would answer in one keystroke: does this superclass have a public utility method called X?
The Solution
jvmsrc queries your build tool (Gradle) for this project's resolved classpath, then hands your agent real source, real signatures, and real structure — for the exact version your build actually uses.
- As an MCP server – Connect to Claude Code, Cursor, Windsurf, or any other MCP host and equip your agent with six purpose-built classpath tools.
- As the
jvmsrcCLI – Same engine, scriptable, pipe-friendly, and useful on its own when you just want to read a class.
[jvmsrc resolving a Spring class in Claude Code]
Claude Code resolving HandlerInterceptor signatures via jvmsrc — 3 calls, correct answer, no cache walking.
What it saves, daily
On JVM projects with private libraries and no public Javadocs, jvmsrc is a force multiplier:
- ~50K tokens saved per "what's on this external class" investigation — roughly 70% reduction in cost.
- 15+ agent panic loops avoided daily — no grinding through Gradle caches, opening wrong JARs, or giving up.
- Unlock delegation for complex tasks touching internal SDKs that you previously had to do yourself.
The Difference in Action
Without jvmsrc (The 25-turn blind grind)
User: Override the audit hook from
AbstractTradingService— is there a public utility method on it we should reuse?Agent:
- Searches workspace for AbstractTradingService.java ➔
0 hits- Runs:
find ~/.gradle -name "trading-core*"- Finds 4 versions:
2.1.0,2.3.0,2.4.1,3.0.0-SNAPSHOT- Guesses: Picks
trading-core-2.4.1.jar(the project actually uses3.0.0-SNAPSHOT!)- Runs:
jar tfandjavap -pon the wrong JAR- [22 turns later] "I don't see a utility method, you'll have to implement it yourself."
Reality:
3.0.0-SNAPSHOTaddedmaskSensitiveFields()as a protected helper exactly for this case. The agent didn't hallucinate — it was just blind.
With jvmsrc (The 3-turn instant resolution)
User: Override the audit hook from
AbstractTradingService— is there a public utility method on it we should reuse?Agent:
search_classes("AbstractTradingService")➔ Finds FQN & exact resolved library.get_class_structure(scope: "overview")➔ DiscoversmaskSensitiveFields()in3.0.0-SNAPSHOT.get_method_signature("maskSensitiveFields")➔ Gets accurate signature and generics.Result: Writes the override correctly on the first try. No cache walking, no guessing, no wrong version.
How It Works
- Build Tool Inquiry:
jvmsrcqueries your active build tool (e.g., Gradle) for the exact resolved classpath configuration. - Intelligent Caching: It caches the resolved classpath, tracking changes to build files to stay current.
- Precision AI Tools: Instead of full-code dumping, it exposes precise, high-granularity tools (signatures, structure, search) to keep context windows small and token usage ultra-low.
Installation & Quick Start
1. Install CLI
Or use it directly via npx:
[!IMPORTANT]
Requires Node ≥ 20 and Java onPATH(for CFR decompiler +javap).
2. Add the MCP server
Paste this into your AI assistant config (Cursor, Claude Code, Windsurf, etc.), then restart the host:
Optional: jvmsrc config (or jvmsrc config --project /path/to/gradle-project) prints a paste-ready block plus environment hints. Most users can skip it and copy the snippet above.
MCP Server Reference
The MCP server runs over stdio via jvmsrc mcp. The default config needs no environment variables:
Private repository credentials (optional)
Only needed when your Gradle build requires credential env vars for a private Maven/Artifactory/Nexus-style repo. MCP hosts often do not inherit your interactive shell, so those vars must be set on the jvmsrc MCP process (then restart the server).
REPO_USER / REPO_PASS below are sample names only — they are not required by jvmsrc. Use whatever variable names your project’s Gradle scripts document:
Omit the env block entirely when the project does not need them.
Tools your agent gets
[!TIP]
Every source response includessourceAvailable:truefor real sources (Javadoc, parameter names, generics),falsefor CFR decompilation (structure reliable, names may be synthetic).
[!NOTE]
Multimodule: omitmodulePathand jvmsrc auto-picks the unique owning module; on a miss it lists candidatemodulePaths. Methods:search_classesmatches declared method names when the index has source enrichment; for body text in a known JAR usesearch_in_artifact.get_class_sourcemethodNamesalso walks superclasses for unmatched names.
Make your agent use it
Tool descriptions help, but a built-in Bash tool the agent already trusts will often win. Two small additions make jvmsrc the default for dependency questions.
1. Add a rule to your agent's instructions (CLAUDE.md, AGENTS.md, Cursor rules) — many clients do not show MCP server instructions at all:
2. (Claude Code) Block the shell fallbacks with a hook. docs/hooks/block-jvm-fallbacks.mjs denies javap, unzip/jar on JARs, and ~/.gradle/caches / ~/.m2 access, and tells the agent to use jvmsrc instead, so it redirects instead of silently reading the wrong version. Save the script somewhere stable and add to .claude/settings.json (project) or ~/.claude/settings.json (all projects):
Set JVMSRC_HOOK_DISABLE=1 to switch it off temporarily. Other clients have their own hook mechanisms; check their current docs before porting.
Check that your client shows the server instructions: ask the agent "What instructions does the jvmsrc MCP server give?". If it cannot answer, rely on the rule in step 1.
[!NOTE] Inputs are forgiving.
projectRootcan be omitted (the workspace is used, or setJVMSRC_PROJECT_ROOT);modulePathacceptsapp,/app,:app, or a file/directory inside the module (omit it, or pass"*", to search all modules);classNameaccepts a simple name,Outer.Inner, or a trailing.class/.java. The first call on a project runs Gradle (typically 5–10s); later calls use the cache. PassforceRefresh: trueonly after something changed that the build files do not show, such as a SNAPSHOT republish.
How It Compares
Target Audience
Primarily Java + Spring Boot projects on Gradle. Other JVM languages (Kotlin, Scala) and Android work today on a best-effort basis and are on the roadmap as first-class targets — see ROADMAP.md.
If you're on Maven or Bazel, it's planned but not shipping yet. Star the repo or open an issue and I'll prioritize accordingly.
Detailed Reference
Requirements & Compatibility
Runtime: Node.js ≥ 20, Java on PATH.
Project types: JVM codebases (Java, Kotlin, Scala, Groovy). jvmsrc calls the build tool, not your editor.
Point -p / projectRoot at the Gradle root (settings.gradle(.kts) or root build.gradle(.kts)). Uses ./gradlew when present, else gradle on PATH. Maven-only trees get an explicit unsupported error.
Known Limitations
Early software; the supported path is narrow:
Composite builds, Android-only layouts, and exotic configurations are not fully validated. See ROADMAP.md.
Security & Privacy
- No telemetry.
- Local only — caches and diagnostics stay on disk; never writes under your project root.
- Subprocesses via argv only (no shell interpolation) — see SECURITY.md.
- Optional
JVMSRC_ALLOWED_ROOTSto lock down which projects jvmsrc may resolve.
CLI Command Reference
Useful flags: -p / --project, --module (:core:api), --configuration, --include-test, --force-refresh, --verbose (Gradle stderr only), --method, --start-line / --end-line.
Repo fixture for testing:
test/fixtures/gradle-smoke —
jvmsrc get com.smoke.Core -p test/fixtures/gradle-smoke --module :core.
Troubleshooting
- Resolution failures: Run
jvmsrc diagnostics last(orjvmsrc diagnostics last 5) - Custom JDK install roots: Add once with
jvmsrc config jdk-roots add /path/to/jdks - JDK mismatch debugging: Run
jvmsrc doctor java -p /path/to/project - After upgrading jvmsrc: Restart your MCP host
- Stale classpath: Run
jvmsrc resolve --force-refresh
Environment Variables
Defaults follow env-paths conventions per OS. Full layout: SPEC.md §6.
When JVMSRC_JAVA_HOME is not set, jvmsrc auto-discovers local JDKs from common paths such as ~/.jdks (IntelliJ), ~/.gradle/jdks, SDKMan, jenv, asdf, and OS-specific system install directories, plus your global configured JDK roots from jvmsrc config jdk-roots ....
AI Agent Reviews
Finally, an MCP That Doesn't Make Me Decompile JARs
"This tool is a revelation for anyone tired of LLMs hallucinating non-existent Spring APIs. It actually reads bytecode, providing accurate class definitions and source lookups without the usual 'vibes-based' guesswork. The
search_classesfunctionality is incredibly precise, and the thoughtful implementation ofjavapfallback and granular scope controls (overview/declared/effective) makes navigating complex JARs painless. It’s fast, honest when it can't find a class, and handles cache management perfectly. A must-have for any dev struggling with dependency hell — it’s like having a senior engineer who actually enjoys reading documentation." — Claude (AI Reviewer)
Project Documentation
Building from Source
Full contributor workflow: CONTRIBUTING.md.
I built jvmsrc because I kept running into the same wall: agents that are great at writing Java but blind to the actual classpath. If it saves you the same 25-turn grind it saved me, that's exactly why this exists. Found a bug, have an idea, or just want to say it helped? Open an issue or a PR — I read everything.
License
MIT — see LICENSE.
来源:README.md,提交 56c4bf0
工具
0版本历史
1- v1.16.0最新Sep 16, 2026
评分
已满足适用检查项的 50%
该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。
- 近 30 天内可连通25 分
请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。
- 已声明认证方式10 分
请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。
- 已列出工具15 分
请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。
- 描述必需15 分
- 图标10 分
- 文档15 分
- 许可证10 分
其他 Source 集成
Source
app.source官方发布者
Source 提供的远程 MCP 端点,让助手访问已接入 Source 的网站的营销归因与分析数据。

ArxSource
由 GitHub 账号 nicolaiprodromov 发布
本地 MCP 服务器,让助手检索 arXiv 论文、按章节或 LaTeX 原文阅读、追踪引用,并找到论文对应的代码。
SourceMed
io.sourcemed
SourceMed 为助手提供经来源核实的临床证据库:可提出证据问题、浏览书架并阅读带引用的卡片。

Spree Commerce: open-source B2B and marketplace ecommerce
由 GitHub 账号 spree 发布
Spree 的 MCP 服务器让 AI 助手用自然语言管理 Spree 商店的商品目录、回答报表问题并完成店铺设置。

PlushSource
由 GitHub 账号 quickman924 发布
本地 MCP 服务器,内置定制毛绒玩具制造商名录,让助手搜索、筛选、比较并引用厂商条目。