Pheebs

Pheebs hooks into Claude Code, Cursor, and Codex and fires on the events that mark real work: session start and end, prompt submission, tool calls and failures, skill and slash-command expansion, sub-agent spawns, compaction, and background tasks. Seventeen event types cover that range. Each prompt becomes a character count, plus an intent label when the optional prompt classifier is switched on. Claude Code analytics and Cursor usage tracking come through the same hook mechanism, so one install covers both.

Every event is appended to a local JSONL file, which stays the durable copy whether or not a backend is configured. A fresh install ships with no endpoint at all. Point it at your own server through a base URL and token, and the same events also post to your infrastructure. The reference backend and the endpoint contract live in the open repository, so the receiving side is something you can read before you trust it.

This is the feature most teams buy the tool for. Pheebs assigns every task prompt a scope, from a single-file change up to open-ended design, and judges the session by its hardest prompt. It then compares that size with the model class that ran. An over-provisioned session burns budget quietly, and an under-powered one shows up later as repair prompts. The gap between the two gets priced at API list rates, with an explicit caveat that the figure is an estimated upper bound.

File contents, file paths, directory structures, prompt text, and command strings never reach the log. A command like npm test is read in process and recorded as tool_intent: test_run. Identity is either the id behind your token or a truncated hash of your git email. Pheebs doesn't look up GitHub handles. For teams where AI measurement died in legal review, these defaults are the part that gets it approved.

Raw events are hard to argue about in a leadership meeting, so Pheebs ships a model that turns them into two layers. The first covers repertoire: 24 practices across six competencies, covering model choice, reusable artifacts, MCP connections, verification wired into the agent loop, context management, and multi-agent orchestration. The second covers judgement: verification coverage, pushback rate, refinement-to-repair ratio, wholesale-accept rate, and model fit. Each practice is detectable from telemetry, which means the reporting doesn't depend on self-reported surveys.

Claude Code and Codex emit native OpenTelemetry metrics and logs, and Pheebs proxies those signals to your collector instead of asking you to ship an observability credential to every laptop. That keeps existing tracing and logging stacks in play. Cursor runs on hooks alone, since that agent doesn't expose the same native export path.