Argus - See what Claude Code actually did, and what it cost

Hi Product Hunt, I'm Krish. I built Argus because there was no local observability tool for the coding agents we actually use all day. We have good tracing for LLM API calls. Claude Code, the tool most of us spend the most hours in, is a black box. It writes a full transcript of every session to ~/.claude, every turn, tool call and sub-agent, and then tells you almost nothing about it. When a session goes sideways you can't see which tool call did it, what a sub-agent was told, or what happened inside a workflow. Argus reads those transcripts into a SQLite archive on your machine and serves a dashboard on localhost: - Every session turn by turn, with the tool calls inside each turn and their errors. - Every sub-agent, including ones inside workflows: the exact task it was given, its tools, its own timeline. - Tool error rates, with a detector that flags a tool the day it starts failing. - Tokens and estimated cost per turn. Tokens exact, costs estimated. - Opt-in full-text search across prompts, replies, thinking and tool output. - A permanent archive. Claude Code rotates its logs; Argus keeps what it has ingested. Nothing is uploaded. It binds to 127.0.0.1, has no telemetry, and the only network call is an optional price refresh. pip install argus-code argus start MIT, code on GitHub. Claude Code first; ingest is adapter-based so other agents can follow. What can't you see about your agents today that you wish you could? That's the question shaping the next version.