JVMind Community Edition

An AI-powered JVM diagnostics agent. Single-user · local-first · open source.

中文文档 (Chinese) · Report a bug · Request a feature

JVMind CE is a JVM performance diagnostics assistant built on OpenAI-compatible LLMs. It runs the agent as a local web service, accepts uploaded logs/dumps, and streams diagnostic conclusions back over SSE.

• 🧠 LangGraph agent — Tool orchestration via LangGraph state machines (native OpenAI function-calling).
• 🪵 GC log analysis — JDK 8 / 11 / 17 / 21 / 25 collectors (G1, Parallel, ZGC, Shenandoah, Serial, CMS). Pure regex parsers, no external runtime deps.
• 🧵 jstack thread analysis — Deadlock detection, lock-contention hotspots, thread-pool distribution, flame graph, per-thread drill-down.
• 💾 Heapdump analysis (optional) — Parses GB-sized hprof files via Eclipse MAT (requires JDK 21). Ships with the bundled query-service plugin.
• 🔌 OpenAI-compatible LLM — DeepSeek, OpenAI, Qwen, Kimi, Ollama (local), and any compatible endpoint. Hot-reloadable UI config.
• 🏠 Ollama local inference — Run entirely offline with no data leaving your machine. Zero-cost, no API key required.

The community edition ships without billing, team management, or rate limits. For those, use a commercial offering.

Quick Start

Install from PyPI

pip install jvmind-ce
jvmind
# open http://127.0.0.1:8000

Install from source (developer mode)

git clone https://github.com/jvmind/jvmind-ce.git
cd jvmind-ce
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
jvmind

Configure the LLM

The UI prompts for an API key on first launch. Two options:

Option A — Remote LLM provider (DeepSeek, OpenAI, etc.):

cp .env.example .env
# Edit .env:
OPENAI_API_KEY=sk-xxxxxxxxxxxxxxxx
OPENAI_BASE_URL=https://api.deepseek.com/v1
OPENAI_MODEL=deepseek-chat

Option B — Local Ollama (no API key, zero-cost, fully offline):

# 1. Install & start Ollama: https://ollama.com
# 2. Pull a model:
ollama pull qwen2.5
# 3. In the JVMind UI, click "Ollama · Local" preset in Settings (⚙️)
#    or set in .env:
OPENAI_BASE_URL=http://localhost:11434/v1
OPENAI_MODEL=qwen2.5
#    Leave OPENAI_API_KEY empty — Ollama doesn't need one.

The agent uses native OpenAI function-calling, so any model supporting tools works. DeepSeek's deepseek-chat is a sensible default for remote inference; qwen2.5 or llama3.1 are good choices for local Ollama.

Privacy: When using Ollama with a local model, no data ever leaves your machine — all GC logs, thread dumps, heap dumps, and conversation content stay local.

Features

GC log analysis

Open the 📊 GC Logs tab, upload a log file. The pipeline:

1. Stream-upload → store raw text in DB
2. Parser (pure regex) extracts events, computes statistics, builds overview
3. Render: 8 status cards + per-collector table + heap chart + pause histogram + top-10 slowest
4. Trigger a streamed LLM diagnosis with the structured template (overall health, key issues, parameter tuning, follow-up metrics)

Supported collectors: G1 · Parallel · Serial · ZGC · Shenandoah · CMS. JDK 9+ unified log format and JDK 8 PrintGCDetails are both supported.

jstack thread analysis

Open the 🧵 Thread Dumps tab, upload a jstack -l dump. Features:

• Thread state histogram (RUNNABLE / BLOCKED / WAITING / TIMED_WAITING)
• Deadlock detection with lock chain visualization
• Lock contention hotspots (holder + waiters list, click-to-drill)
• Thread pool distribution, flame graph, single-thread drill-down
• Streamed LLM diagnosis (whole dump or single thread)

Heapdump analysis (optional — requires Eclipse MAT + JDK 21)

The 🔍 Heap Dumps tab uploads GB-sized .hprof files. The architecture:

  Browser ─▶ FastAPI (upload routes) ─▶ local disk: <dump_dir>/<report_id>/
                                              │
                                              ▼
                            worker_loop ─▶ ParseHeapDump.sh
                                              │
                                              ▼
                          ┌───── query-service (HTTP) ◀─────┐
                          │   bundled in vendor/mat/         │
                          └──────────────────────────────────┘
                                              │
                                              ▼
  Browser ◀────── SSE progress / JSON query results ◀───── FastAPI proxy

Install MAT and the query-service in one step:

./scripts/install_mat.sh /opt/mat
# or: MAT_HOME=/opt/mat ./scripts/install_mat.sh

This downloads Eclipse MAT, extracts it, and copies the bundled com.jvmind.mat.query-0.1.0.jar into the MAT plugins directory. Note: MAT requires JDK 21. Configure and run:

# .env
MAT_HOME=/opt/mat
MAT_QUERY_SERVICE_URL=http://127.0.0.1:8090

# Terminal 1: query-service
/opt/mat/MemoryAnalyzer -consoleLog -nosplash \
    -application com.jvmind.mat.query.QueryServiceApp

# Terminal 2: heapdump worker (separate process)
jvmind-worker

The bundled query-service jar ships inside vendor/mat/, so users don't need to compile Java themselves.

Agent internals

The agent is a LangGraph state machine in react_agent/graph/:

• facade.LangGraphAgent — public API for the agent
• graph_builder.build_graph — wires nodes (Agent → Tools → Finalize)
• nodes — tool execution + structured reasoning (cross-domain diagnosis for OOM)
• sse_adapter — yields user / token / step / fact_added / final / error / done events
• llm_compat — non-streaming chat helper used by the summarizer and skills route

Native OpenAI function-calling is the default. If a provider rejects the tools parameter, the agent surfaces a clear error rather than silently degrading.

Configuration

See .env.example for the full list. Key knobs:

The web UI also has a Settings (⚙️) dialog for LLM config — changes hot-reload the agent and re-encrypt stored keys.

Development

# Run dev server with auto-reload
uvicorn server:app --reload --port 8000

# Run all tests
python -m pytest _tests --no-cov

# Frontend dev (Vite hot-reload)
cd frontend && npm install && npm run dev

# Frontend production build
cd frontend && npm run build

# Heapdump worker (separate process)
jvmind-worker

Project layout:

jvmind-ce/
├── server.py              # FastAPI entry point
├── app/                   # routes, middleware, helpers
├── react_agent/
│   ├── graph/             # LangGraph agent (default)
│   ├── gc_analyzer/       # GC log parsers (JDK 8 + JDK 9+)
│   ├── jstack_analyzer.py
│   ├── mat_tools.py       # Heapdump tool wrappers
│   ├── heapdump_worker/   # Background MAT parser
│   ├── memory_db.py
│   ├── user_manager_db.py # single-user
│   └── db.py / models.py
├── frontend/              # vanilla JS + Vite
├── vendor/mat/            # bundled query-service jar
├── scripts/install_mat.sh # one-step MAT install
└── _tests/                # pytest suite (146 tests)

License

MIT — see LICENSE.

Contributing

See CONTRIBUTING.md and CONVENTIONS.md.

Credits

JVMind CE is extracted from the commercial JVMind product. The commercial edition includes multi-user auth, team workspaces, Paddle billing, PostHog analytics, and other features that this repository deliberately omits.