Arthas Workbench
Connects to Kubernetes Java workloads using configured kubeconfig, namespace, Pod, and optional container/remote port, then uses an owned kubectl port-forward to reach an existing Arthas MCP endpoint for diagnostics.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
Arthas Workbench
A shared Web workbench and one MCP endpoint for local and Kubernetes Java diagnostics.
中文说明 · UI preview · Screenshots · Development · Security
Prototype — review before deployment. This repository proves the local/Kubernetes connection paths with an actual Rust connector and simulated upstreams. It is not a production-ready release. There is no one-click installer yet. The public UI preview is entirely simulated and cannot inspect your computer or connect to a cluster.
Interface preview
Open the interactive prototype →
The new v0.2 design prototype organizes diagnostics by environment → application → exact Pod/JVM. Explore production, staging and development deployments of the same application; group replicas by worker node; launch parallel tasks; inspect structured results; and simulate an independent reconnect or same-name Pod replacement. A single light theme keeps target identity, connection state and results in focus.
This is an offline design prototype, separate from the runnable gateway UI. Per-Pod supervision, environment discovery, persistent tasks, OGNL and Profiler are simulated design interactions, not newly implemented backend features. Every metric is fabricated. Refresh resets the scenario; closing the page ends the simulation. No process, cluster or real MCP endpoint is contacted. Open docs/preview/index.html directly after cloning, or use the hosted preview. See design behavior and capability boundaries.

Send UI feedback. Please include the environment, page, viewport and steps; remove credentials and private target names from screenshots.
Related MCP server: mcp-grafana
What works today
Capability | Prototype status |
Design preview | Single-theme environment/application workspace, multi-Pod task and lifecycle simulations; separate from runtime |
Runtime Web UI | Existing target search, explicit MCP verification, diagnostic forms, cancellation and in-memory session history |
Local Java | Outbound connector, loopback Arthas forwarding, |
Kubernetes | Configured kubeconfig/namespace/Pod → owned |
MCP endpoint | Streamable HTTP at |
Tools |
|
Isolation | Separate user/node credentials, node access rules, user-scoped upstream sessions, shared per-target concurrency |
Resilience | Bounded streaming, backpressure, cancellation, reconnect and owned child-process cleanup |
Kubernetes targets currently require an already-running Arthas MCP endpoint. Long trace/watch calls require the target to advertise task support. A successful TCP/HTTP probe alone does not prove that a JVM was correctly attached.
Not implemented: Kubernetes resource upload/Attach, Perl file transfer, hot swap, MySQL, SSO, automatic device pairing, a bundled launcher, full legacy tool compatibility or production audit storage. No OBS integration is included.
Why a local connector exists
A web page cannot enumerate OS processes or attach to a JVM on its own. The small native connector runs on the target computer, performs these operations within that user's permissions, and initiates a connection to the workbench. Your browser stays the UI.
flowchart LR
Browser["Browser UI"] --> Gateway["Web API + MCP gateway"]
IDE["MCP client"] -->|"Streamable HTTP"| Gateway
Local["Connector on developer computer"] <-->|"Outbound-established WS/WSS channel"| Gateway
Local --> JVM["Local JVM / Arthas"]
Cluster["Connector with cluster access"] <-->|"Outbound-established WS/WSS channel"| Gateway
Cluster -->|"kubectl port-forward"| Pod["Pod / Arthas"]Proposed end-user experience: open the team's shared workbench → select an environment and application → diagnose a specific JVM. Kubernetes users need only a browser and authorization. For local Java, launch the local helper once and pair it with the team server; the helper discovers processes and obtains centrally managed Arthas resources. Pairing, packaging and resource delivery are design directions, not implemented installers or services. See the onboarding design.
Developer quick start
The current runtime is validated on Windows with Python 3.13. Python 3.11–3.13 is supported by the package constraints. Rust CI includes Linux unit checks; full connector/JVM/platform acceptance is still limited and tracked explicitly.
Requirements: Git, uv, Python 3.13, Rust 1.94.1 or newer; a JDK for local scanning/Attach and kubectl for real Kubernetes targets.
git clone https://github.com/yy568821795-max/arthas-workbench.git
Set-Location arthas-workbench
$project = (Get-Location).Path
$env:UV_CACHE_DIR = "$project/.cache/uv"
$env:UV_LINK_MODE = 'copy'
$env:CARGO_HOME = "$project/.cache/cargo"
$env:CARGO_TARGET_DIR = "$project/.build/connector"
$env:PYTHONDONTWRITEBYTECODE = '1'
uv sync --locked --python 3.13
cargo build --locked --manifest-path "$project/connector/Cargo.toml"Try the isolated full-stack demo
& "$project/.venv/Scripts/python.exe" -B "$project/tests/smoke_connector.py" --serveThis starts the real Web/MCP server and Rust connector against fake Arthas endpoints and a strict fake kubectl relay. It prints an ephemeral URL and test-only login token. Use Ctrl+C, or create the printed stop_file, to stop all owned demo processes. No real JVM or Kubernetes API is contacted. This integration demo currently uses Windows fixtures; the static preview works without them.
Connect your own authorized target
& "$project/.venv/Scripts/python.exe" -B "$project/scripts/init_config.py" `
--output-dir "$project/.runtime/local" --port 18880 --local-port 8563
# Terminal 1
& "$project/.venv/Scripts/python.exe" -B -m workbench.cli `
--config "$project/.runtime/local/server.json"
# Terminal 2 (set $project to the same checkout)
& "$project/.build/connector/debug/arthas-connector.exe" `
--config "$project/.runtime/local/local-connector.json"Open http://127.0.0.1:18880/. Use the operator user token from the generated server configuration. User tokens are distinct from connector tokens and are kept only in page memory. Configuration generation does not contact a target and refuses to overwrite existing credentials.
Existing local Arthas: point the local target at its HTTP loopback origin.
Optional local Attach: supply
--arthas-homepointing to a complete offline distribution when generating configuration. Attach is disabled by default, requires an explicit PID/identity and user scope, and stages isolated configuration under the connector data directory.Kubernetes: generate a separate configuration directory with explicit
--kubeconfig,--namespace,--pod, and optionally--container/--remote-port; then start that K8s connector. The target must already provide Arthas/mcp.Remote deployment: configure the public HTTPS origin, TLS termination and connector WSS URL. Loopback defaults are for local development only.
See runtime behavior and limitations before real target use.
MCP client configuration
Use the Streamable HTTP URL and a user Bearer token:
URL: http://127.0.0.1:18880/mcp
Authorization: Bearer <your-user-token>Call list_instances, then provide the exact instance_id on every diagnostic call. For example, jvm(instance_id="local-node:jvm"). Trace/watch accept a timeout of 1–300 seconds; they wait for the target method to execute.
Clients must support Streamable HTTP and the configured authentication method. A session ID is not a user credential. Tokens must not be placed in URLs or public issue reports.
Validation
The initial local validation completed 12 Python integration tests and 8 Rust unit/native fixture tests. Tests cover both routes, standard MCP clients, >35-second tasks, JSON/SSE chunks, authorization, cancellation, reconnect and owned process cleanup. A manual-probe recovery regression also passed. These results are mock-backed, not proof of real-cluster or real-JVM acceptance; the Actions badge shows the current published CI result.
$env:WORKBENCH_LONG_TESTS = '1'
& "$project/.venv/Scripts/python.exe" -B -m unittest discover -s "$project/tests" -v
cargo test --locked --features test-fixtures --manifest-path "$project/connector/Cargo.toml"中文说明
这是一个专注 Arthas 的独立开源原型:把 Web 工作台、MCP 服务和本机/K8s 执行能力统一起来。
先审查界面: 新版交互原型采用单一浅色主题,包含多环境同名应用、多 Pod 状态、并行诊断及断线/重建演示。全部数据与操作均为模拟,设计预览与现有运行界面分开维护。
本机能力的原理: 主入口是团队共享服务器。K8s 用户只需浏览器;诊断本机 Java 时,由用户运行一次本机助手完成配对、扫描、资源准备和连接。该简化接入流程仍待实现,浏览器自身不会扫描进程。
当前状态: 两条连接链路和一组 MCP 工具已实现;自动配对、一键启动、K8s 上传与 Attach、Perl 文件传输、MySQL 尚未实现。开发配置流程不代表最终产品体验。
反馈方式: 提交界面建议,说明页面、操作步骤、屏幕尺寸与期望效果即可。
Contributing and license
Read CONTRIBUTING.md, CODE_OF_CONDUCT.md, SECURITY.md and third-party attribution.
Project-authored code and documentation are MIT licensed. Dependencies retain their own licenses. Arthas and kubectl binaries are not bundled. This is an independent community project and is not an official Alibaba Arthas distribution.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Guarded MCP server for agent-readable business truth, provenance, readiness, and discovery.
Stateless MCP gateway and OTel span-streaming bridge for hosted MCP servers.
Agent-native registry to discover APIs, MCP servers and CLIs, with live health checks.
The Cortex MCP server provides read-only access to real-time engineering context from the Cortex developer portal, allowing AI coding assistants to answer natural language questions about your organization's catalog (microservices, libraries, domains, teams, infrastructure), scorecards (engineering standards and best practices), initiatives (goals and deadlines), and Engineering Intelligence metrics. It includes tools for querying documentation, tracking personal entities, and accessing AI-assisted insights across the entire Cortex ecosystem.
Related MCP Servers
- AlicenseAqualityCmaintenanceMCP server for profiling Java applications via JDK utilities (jcmd, jfr, jps). Enables AI assistants to diagnose performance, analyze threads, and inspect JFR recordings without manual CLI usage.2663 npm10MIT
- FlicenseNot gradedqualityDmaintenanceIn-cluster MCP server for read-only diagnostics of Grafana, Prometheus, Alertmanager, and Loki, enabling metric queries, alert listings, and log queries through Grafana datasource proxies with Kubernetes RBAC authentication.-
- AlicenseNot gradedqualityDmaintenanceEnables Kubernetes cluster introspection via MCP tools, such as listing pods, namespaces, nodes, and events.4ISC
- AlicenseAqualityBmaintenanceEnables MCP-compatible AI assistants to inspect projects, processes, ports, logs, HTTP endpoints, Git state, and development databases for local runtime diagnostics.7MIT