◉ vibescope

Local-first observability for the vibe-coding era — for every coding agent.

You code with AI now. How much? Through which agents? On what repos? At what cost? What did you even do last week? Nobody can see their own AI-assisted development — vibescope makes it visible in one command:

npx vibescope

No signup, no API keys, no tracking, no cloud. It reads what's already on your machine — your coding agents' local session histories and your git repos — and opens a dashboard:

• 💸 Estimated AI spend — per model, per project, per day (standard API rates, cache-aware)
• 🤖 Per-agent breakdown — Claude Code, Cursor, Codex, Gemini, Aider… who did what
• 🔨 What the agents did — ranked tool invocations (Bash vs Edit vs Read vs MCP)
• 🗓 Daily rhythm — calendar heatmaps of prompts and commits, side by side
• ⏱ Pair-coding time — actual active hours with agents, idle gaps excluded
• 📁 Where the effort went — sessions, prompts, commits and cost per project
• 📝 Your week, drafted — an auto-generated standup with week-over-week trends
• 🖼 Wrapped — a shareable stats card, rendered and downloaded entirely client-side

Agent-agnostic by construction

Every agent is a provider behind one contract — detect() plus scan() returning normalized stats. The dashboard, per-agent breakdown and recap work off that shape, so supporting a new agent is one file:

| Agent | Reads | Status | |---|---|---| | Claude Code | ~/.claude/projects/**/*.jsonl | ✅ full (tokens, cost, tools, pair time) | | Cursor | SQLite state via bundled sqlite3 CLI | ✅ sessions + prompts (no token data persisted) | | Codex CLI | ~/.codex/sessions/**/*.jsonl | 🧪 experimental | | Gemini CLI | ~/.gemini/tmp/*/logs.json | 🧪 experimental | | Aider | <repo>/.aider.chat.history.md | 🧪 experimental | | Your agent | docs/providers.md | PRs welcome |

Fields a format doesn't expose stay at zero — vibescope never guesses.

Privacy

Everything stays on your machine. vibescope reads local files, aggregates in-process, and serves the dashboard on localhost. It makes zero network requests. The dashboard shows aggregates (counts, tokens, costs, titles) — never your prompt or code content. The Wrapped card is rendered in your browser and saved to your Downloads; sharing it is your call.

Usage

npx vibescope                          # scan cwd + ~/work, serve dashboard on :4177
npx vibescope --roots ~/code,~/oss     # where your git repos live
npx vibescope --months 3               # look-back window (default 6)
npx vibescope --authors [email protected]    # git identity match (default: your git config)
npx vibescope --providers claude-code,cursor   # scope to specific agents
npx vibescope --json > stats.json      # raw aggregated data, no server

Tip: Claude Code prunes old sessions (default ~30 days). To build a longer history, raise cleanupPeriodDays in ~/.claude/settings.json — vibescope can only see what still exists.

Cost estimation notes: token usage is deduped per API request (streamed chunks share a requestId), cache reads bill at 0.1× input rate, cache writes at 1.25×/2× (5m/1h TTL), priced from each message's own model id. Subscription users: read cost as "API-equivalent value", not a bill.

Team mode

Your whole team's picture — adoption, spend, rhythm — without a server and without anyone's data leaving their machine unwillingly:

# each member, locally (--redact strips session titles):
npx vibescope --json --name alice --redact > alice.json

# anyone, with the collected files:
npx vibescope team ./snapshots/

The team dashboard shows combined spend/prompts/pair-time, per-member bars, agent adoption (who uses Claude Code vs Cursor vs …), a team rhythm heatmap with per-member tooltips, and a full member table. Sharing a snapshot is an explicit, inspectable act — it's a JSON file you can read before sending.

Roadmap

• v0.4 — --compare period dashboards · more providers (Windsurf, opencode) · team trends over time
• Ideas welcome — open an issue

Development

git clone https://github.com/ajitjha393/vibescope && cd vibescope
node bin/vibescope.js        # no dependencies, no build step — Node ≥ 18
npm test                     # fixture-driven node:test suite

Zero runtime dependencies by design: the scanners are plain Node, the dashboard is one self-contained HTML file with hand-rolled SVG charts (colorblind-validated in light and dark). Want to add an agent? Start at docs/providers.md.

License

MIT