626Labs · Claude Code plugin

The /insights you wish you had.

vibe-insights reads your Claude Code transcripts across every machine and your whole history — walls your employer's sessions local-only — and hands back coverage, recall, the cache reveal, trends, and a synthesized read of how you actually work. The verbose /insights, over everything.

v0.2.2·cross-machine·work-walled·MIT

Includes1 skill

ValidatedProven on a live 195-session personal index across two machines.

vibe-insights ~
> /vibe-insights
[vibe-insights] Reading sessions across your machines...
280 sessions · 2 machines · work walled · back to Jan
 
Token burn 144M · output share 18%
Cache reveal 22.8B cache-read (96% of input-side)
Top friction env / tooling — 44% of judged sessions
 
[vibe-insights] Narrative read written → insights.html
01 · What it does

Everything /insights leaves on the table.

A deterministic engine reads your session transcripts and produces one branded report — split by account and machine, with the parts your employer owns kept local.

Coverage & recall

Where was I, across every machine.

Sessions, repos, and burn split by account (work vs personal) × machine. Plus the most-recent sessions with titles, branches, and machines — and the open feature branches you walked away from.

The cache reveal

The number the flat report buries.

Burn, output share, and the cache-read tokens carrying most of your input-side context. The headline burn is the small number; vibe-insights shows you the rest.

The wall

Employer sessions stay local.

Configure your work repos — those sessions are labeled work and kept local-only, never synced cross-machine, never pushed — but still visible in your own report. The wall is by repo, not by home.

How it went

A read, not just a count.

An LLM tags each session for friction, satisfaction, and outcome; deterministic charts roll it up. Then a synthesized 'how you actually work' interpretation, grounded in the numbers.

02 · How it's built

Deterministic engine, LLM only where it earns it.

The engine is deterministic and offline. It never calls an LLM or MCP — it reads local JSONL transcripts and a few cache files. The narrative read, the decisions overlay, and the per-session tags come from the /vibe-insights skill, where the LLM lives, and are handed to the engine as files. That keeps the analytics reproducible and the tool marketplace-portable.

Cross-machine works through tiny per-machine indexes you sync (Syncthing or anything that moves a folder); the report merges every machine. vibe-insights runs fully standalone — the optional cc-logs package keeps token numbers aligned with sibling tools like Sanduhr when it's installed. Decisions are MCP-agnostic: a file, or your dashboard.

No telemetry. Everything stays on your machine — and the work shard never leaves its origin box.
03 · Get it

Install.

Stable marketplace

Tagged releases via the Vibe Plugins marketplace. The Python engine installs from the repo; the optional cc-logs extra keeps token numbers aligned with sibling tools.

/plugin marketplace add estevanhernandez-stack-ed/vibe-plugins
/plugin install vibe-insights@vibe-plugins

Standalone. Reads ~/.claude and ~/.claude-personal; set your work repos to wall employer sessions local-only.