marmutapp/superbased-observer

SuperBased Observer is a single Go program that watches how AI coding assistants like Claude Code, Codex, Cursor, Cline, GitHub Copilot, and several others are actually used on your machine, and turns that raw activity into a clear picture of what is happening and what it costs. Rather than working with just one tool, it organizes everything it sees by the git repository being worked on, so knowledge picked up from one assistant is shared with all the others touching that same project. The problem it solves is that billing dashboards from AI providers only show a total dollar figure, not which project, model, or session actually generated the spend. Observer answers more specific questions: which sessions are burning through an expensive model when a cheaper one would have done the job, whether cached prompts are actually getting reused, and how much money gets wasted when an assistant re-reads a file that has not changed since the last time it looked at it. It works passively, reading each tool's own session logs and optionally sitting between the tool and the AI provider's API to get exact token counts. Everything runs locally and nothing is sent anywhere except the normal upstream calls your tools already make. Results show up three ways: a local web dashboard with tabs for cost breakdown, per-session detail, and a waste detector called Discovery, a command line interface, and an MCP server exposing more than a dozen tools so the AI assistants themselves can query their own history, check whether a file has already been read, or recall what a past session did. To keep costs and context size down, it also compresses and indexes large tool outputs before they reach the model, with each compression layer able to be turned on or off independently. Installation is a single Go binary, available through npm or go install, backed by a pure Go SQLite database with no external dependencies.