AgentSight

AgentSight is a local AI agent monitoring tool for developers investigating failed runs, slow steps or unexpected changes. It connects prompts and model calls to what agents actually do on the machine, including activity their own logs may miss.

• System activity: Track subprocesses, file access, network requests and resource use alongside model and tool calls, with kernel events independent of agent logging.
• Live monitoring: A ranked terminal view and local browser dashboard show active sessions, token use, process trees and metrics.
• Run analysis: Reports help identify retry loops, repeated calls and slow steps; offline profiles break down saved Claude and Codex token usage by project and model.
• Repository replays: Animated views show which files coding agents read, write, create, rename or delete, using local Claude, Codex and Gemini session records.

AgentSight works with Claude Code, Codex, Gemini CLI, OpenCode and OpenClaw. On Linux, eBPF and TLS tracing can observe existing command-line agents, including closed-source programs, without code changes or a proxy. Live tracing requires an eBPF-capable kernel and root privileges; Docker capture also needs privileged host access.

Windows, macOS and Linux can inspect native agent session files without eBPF. Cursor support reads its local sessions. AgentSight is open source under the MIT license and saves recorded sessions in local SQLite files, which can include model request and response payloads.