Observability for AI applications and AI coding agents, built on OpenTelemetry.
LoongSuite is a family of open-source instrumentation and telemetry projects: compile-time and runtime instrumentation for several languages, a shared set of GenAI semantic conventions, and a local-first collector for AI coding agents. Everything speaks OpenTelemetry, so the data lands in whatever backend you already run.
The core repositories live under github.com/alibaba. This organization hosts companion projects — ecosystem integrations and plugins that are published and versioned on their own cadence.
| Project | What it does |
|---|---|
| loongsuite-go | OpenTelemetry compile-time instrumentation for Go — no code changes, no runtime agent |
| loongsuite-python | OpenTelemetry instrumentation for Python applications and AI frameworks |
| loongsuite-java | Java GenAI telemetry utility library built on the OpenTelemetry GenAI semantic conventions |
| loongsuite-js | OpenTelemetry instrumentation plugins for JavaScript-based AI coding agents |
| loongsuite-pilot | Local-first telemetry collector for AI coding agents: token usage, cost, traces, security audit |
| loongsuite-semantic-conventions-genai | The GenAI semantic conventions the projects above emit |
| Project | What it does |
|---|---|
| dsh-plugin | OpenTelemetry tracing for DeepSeek Harness: one GenAI span tree per turn — steps, LLM calls with TTFT, tool executions, token usage — over standard OTLP to any compatible backend |
Pilot discovers the AI coding agents already installed on a machine, deploys the hook or plugin each one needs, normalizes their activity into one GenAI event schema, and exports it where you want it — local JSONL by default, or OTLP traces, HTTP, and Alibaba Cloud SLS.
curl -fsSL https://loongcollector-community-edition.oss-cn-shanghai.aliyuncs.com/loongsuite-pilot/installer.sh \
-o /tmp/loongsuite-pilot-installer.sh && bash /tmp/loongsuite-pilot-installer.sh install
loongsuite-pilot statusSupported agents, configuration, privacy controls, and the output event schema are documented in the Pilot repository.
- OpenTelemetry-native. GenAI semantic conventions rather than a proprietary schema, so traces and metrics work with the backends and dashboards teams already have.
- Local-first. Collection happens on the developer's machine and writes locally by default; every export destination is something the user configures explicitly.
- Privacy as a setting, not a promise. Per-agent content-capture policy and secret masking are applied before anything is exported.
- Zero-touch instrumentation. Compile-time rewriting, hooks, and plugin injection instead of asking application authors to change their code.
Each repository carries its own contributing guide, issue tracker, and license. Bug reports and pull requests are welcome in the repository the change belongs to.
基于 OpenTelemetry 的 AI 应用与 AI 编码 Agent 可观测性。
LoongSuite 是一组开源的埋点与遥测项目：覆盖多语言的编译期/运行期埋点、一套共享的 GenAI 语义约定， 以及一个面向 AI 编码 Agent 的本地优先采集器。全部以 OpenTelemetry 为数据格式，可以直接落到你现有的 后端。
核心仓库位于 github.com/alibaba；本组织托管配套项目—— 按自己节奏发布和版本管理的生态集成与插件。
| 项目 | 简介 |
|---|---|
| loongsuite-go | Go 的 OpenTelemetry 编译期埋点，无需改代码、无需运行时 agent |
| loongsuite-python | Python 应用与 AI 框架的 OpenTelemetry 埋点 |
| loongsuite-java | 基于 OpenTelemetry GenAI 语义约定的 Java GenAI 遥测工具库 |
| loongsuite-js | 面向 JavaScript 系 AI 编码 Agent 的 OpenTelemetry 埋点插件 |
| loongsuite-pilot | AI 编码 Agent 的本地优先遥测采集器：token 用量、成本、调用链、安全审计 |
| loongsuite-semantic-conventions-genai | 上述项目共同遵循的 GenAI 语义约定 |
| 项目 | 简介 |
|---|---|
| dsh-plugin | DeepSeek Harness 的 OpenTelemetry 调用链插件：每轮生成一棵 GenAI span 树——步骤、带 TTFT 的 LLM 调用、工具执行、token 用量——通过标准 OTLP 上报到任意兼容后端 |
Pilot 会发现机器上已安装的 AI 编码 Agent，为每个 Agent 部署所需的 hook 或插件，把它们的活动归一成 同一套 GenAI 事件模型，再导出到你指定的位置——默认写本地 JSONL，也支持 OTLP 调用链、HTTP 与阿里云 SLS。
curl -fsSL https://loongcollector-community-edition.oss-cn-shanghai.aliyuncs.com/loongsuite-pilot/installer.sh \
-o /tmp/loongsuite-pilot-installer.sh && bash /tmp/loongsuite-pilot-installer.sh install
loongsuite-pilot status支持的 Agent 列表、配置项、隐私控制与输出事件模型见 Pilot 仓库文档。
- OpenTelemetry 原生。 采用 GenAI 语义约定而非私有模型，调用链和指标能直接用在团队现有的后端与大盘上。
- 本地优先。 采集发生在开发者机器上、默认只写本地；每一个导出目标都由用户显式配置。
- 隐私是可配置项，不是口头承诺。 分 Agent 的内容采集策略与密钥脱敏在导出之前生效。
- 无侵入埋点。 用编译期改写、hook 与插件注入，而不是要求应用作者改自己的代码。
每个仓库都有自己的贡献指南、issue 区与许可证。欢迎在对应仓库提 issue 和 pull request。