# AgentSight — Zero-Instrumentation Agent Observability

> AgentSight is an MIT eBPF tool that records LLM/agent traffic without SDK changes, helping you observe Claude Code or Gemini CLI interactions locally.

## Install

Save the content below to `.claude/skills/` or append to your `CLAUDE.md`:

## Quick Use

```bash
wget -O agentsight https://github.com/eunomia-bpf/agentsight/releases/latest/download/agentsight
chmod +x agentsight
sudo ./agentsight record -c claude
open http://127.0.0.1:7395
```

## Intro

AgentSight is an MIT eBPF tool that records LLM/agent traffic without SDK changes, helping you observe Claude Code or Gemini CLI interactions locally.

**Best for:** developers who want local, low-overhead observability for agent tools without changing app code

**Works with:** Linux hosts with eBPF support, sudo privileges, Claude Code / Node / Python agent processes

**Setup time:** 10-15 minutes

### Key facts (verified)

- GitHub: 318 stars · 49 forks · pushed 2026-04-22.
- License: MIT · owner avatar + repo URL verified via GitHub API.
- README-verified entrypoint: `docker run --privileged --pid=host --network=host \`.

## Main

- Use it when instrumentation is too expensive: eBPF captures behavior at the system boundary without modifying your agent code.

- Start by recording one process name (claude/node/python) and validate you can correlate events with failures or regressions.

- Treat the trace as evidence: use it to debug latency spikes, unexpected network calls, or policy violations.

### Source-backed notes

- README states “Zero Instrumentation Required” and focuses on observing at the system boundary using eBPF.
- README includes command examples to record Claude Code and gemini-cli processes via `agentsight record -c ...`.
- README suggests a local web UI at http://127.0.0.1:7395 for viewing recordings.

### FAQ

- **Does it require code changes?**: No—README emphasizes zero instrumentation.
- **Where does it run?**: README examples target local processes on a host with sudo permissions.
- **Is it cross-platform?**: It’s eBPF-based, so Linux support is the primary path (check README).

## Source & Thanks

> Source: https://github.com/eunomia-bpf/agentsight
> License: MIT
> GitHub stars: 318 · forks: 49

---

<!-- ZH -->

## Quick Use

```bash
wget -O agentsight https://github.com/eunomia-bpf/agentsight/releases/latest/download/agentsight
chmod +x agentsight
sudo ./agentsight record -c claude
open http://127.0.0.1:7395
```

## Intro

AgentSight 是 MIT 许可的 eBPF 可观测性工具，无需接入 SDK 即可记录 LLM/agent 的网络与进程行为，适合监控 Claude Code、Gemini CLI 等本地工具的交互轨迹并做审计分析。

**Best for:** 想在不改代码的前提下，对本地 agent 工具做低开销观测的开发者

**Works with:** 支持 eBPF 的 Linux 主机、sudo 权限、Claude Code/Node/Python 等 agent 进程

**Setup time:** 10-15 minutes

### Key facts (verified)

- GitHub：318 stars · 49 forks；最近更新 2026-04-22。
- 许可证：MIT；作者头像与仓库链接均已通过 GitHub API 复核。
- README 中核对过的入口命令：`docker run --privileged --pid=host --network=host \`。

## Main

- 当接入 SDK 代价太高时用它：eBPF 能在不改 agent 代码的情况下捕获系统边界行为。

- 先从一个进程名（claude/node/python）录制起，确认能把事件与失败/回归对应上。

- 把 trace 当证据：用于定位延迟尖峰、异常网络调用或策略违规行为。

### Source-backed notes

- README 强调“Zero Instrumentation Required”，通过 eBPF 在系统边界做观测。
- README 给出用 `agentsight record -c ...` 记录 Claude Code / gemini-cli 等进程的示例。
- README 提示可在本地 http://127.0.0.1:7395 查看记录结果。

### FAQ

- **需要改代码吗？**：不需要。README 强调零侵入。
- **它运行在哪里？**：按 README 示例，它在本机以 sudo 权限观察本地进程。
- **跨平台吗？**：基于 eBPF，主要面向 Linux（以 README 为准）。

## Source & Thanks

> Source: https://github.com/eunomia-bpf/agentsight
> License: MIT
> GitHub stars: 318 · forks: 49


---
Source: https://tokrepo.com/en/workflows/agentsight-zero-instrumentation-agent-observability
Author: Script Depot