AgentSight is an open-source, local-first observability and profiling harness for AI coding agents developed by the Eunomia project. Operating similarly to system tools like top or strace, it correlates high-level prompts, model calls, and tool decisions with kernel- and OS-level effects on the local host, including process spawns, file modifications, and network connections. In practical use, AgentSight operates directly from the terminal. Developers run commands such as `agentsight top` for a live interactive TUI ranking active agent sessions by model, token consumption, health, process family, and tool activity. It inspects running agents non-invasively by capturing plaintext LLM traffic at SSL/TLS boundaries and hooking system events, recording runs to local SQLite databases or exporting them via OpenTelemetry. It supports closed-source and open-source agents alike, including Claude Code, Codex, Gemini CLI, OpenCode, and OpenClaw. Unlike application-level observability platforms (such as Langfuse, LangSmith, or Phoenix) or API proxy gateways that require SDK instrumentation or traffic rerouting, AgentSight uses Linux eBPF to trace agents externally. This enables full observability into closed-source CLI binaries, uninstrumented subprocesses, and agent-spawned scripts that bypass standard framework loggers. AgentSight is written in Rust and licensed under the MIT license. It is available as a standalone CLI binary and as a Rust crate (`agentsight-capture`). Linux systems with kernel 4.1+ support full eBPF monitoring, while macOS, Windows, and Linux can parse agent-native session files directly for offline inspection and visualization.
Tags: AI, ai agent, cli, coding agent, developer tools, OpenSource
top) — Provides a live terminal UI displaying real-time agent sessions, token usage, tool invocations, and system resource consumption. - Repository Replay (agentsight vis) — Visualizes file creation, editing, renaming, and deletion sequences across Git worktrees into self-contained HTML or animated GIF replays. - Semantic Profiling (agentpprof) — Generates flamegraphs of agent call stacks to pinpoint slow steps, retry loops, and token-heavy model interactions. - Local SQLite & OpenTelemetry Output — Stores recorded sessions locally in structured SQLite files (agentsight-*.db) and supports exporting traces through standard OpenTelemetry collectors. - Broad CLI Agent Compatibility — Out-of-the-box monitoring and log ingestion for Claude Code, Codex, Gemini CLI, OpenCode, OpenClaw, and custom shell commands.agentpprof to detect redundant context re-sends, slow tool execution loops, and compute bottlenecks across multi-step agent runs.AgentSight is built for software engineers, security auditors, and AI infrastructure teams who execute autonomous coding agents locally or in CI environments and require low-level visibility into agent system behavior. It is not intended for teams looking for managed cloud web dashboards for simple chat completions without local process execution.
bash # Via Cargo cargo install agentsight # Via Homebrew (Linux x86_64) brew tap eunomia-bpf/tap brew install eunomia-bpf/tap/agentsight # Or download the prebuilt binary wget https://github.com/eunomia-bpf/agentsight/releases/latest/download/agentsight chmod +x agentsight sudo mv agentsight /usr/local/bin/sudo enables eBPF tracing on Linux): bash sudo agentsight topbash # Navigate to your project directory cd /path/to/your/project # Generate an HTML visual replay of Claude/Gemini/Codex file interactions agentsight vis -o output/agent-nebula.html # Query recorded SQLite sessions agentsight reportPart of CLI Coding Agents
Categories: AI Coding Agents, Developer Tools