// the find
chiphuyen/sniffly
Claude Code dashboard with usage stats, error analysis, and sharable feature
Sniffly parses Claude Code's local JSONL session logs and serves a local web dashboard showing token usage, cost breakdowns, error patterns, and a searchable command history. It's for developers who use Claude Code heavily and want to understand where their API budget is going or what kinds of mistakes the agent makes repeatedly. The sharing feature uploads a snapshot to sniffly.dev for team review.
1. Zero-dependency quickstart via `uvx sniffly@latest init` — no virtual env setup, no config file before first run, just works. 2. The error breakdown by type is genuinely useful: most AI coding tools give you raw logs; this actually categorizes failures so you can spot patterns across sessions. 3. Local-first with opt-in sharing is the right privacy model — all processing happens on disk, shared dashboards require an explicit click. 4. The test suite is real: baseline regression JSON, processor correctness tests, performance tests, and mock JSONL data — not just a few smoke tests.
1. Tightly coupled to Claude Code's undocumented JSONL log format. Anthropic can change that format in any release and break the whole tool with no warning; there's no schema version guard in the code. 2. The sharing backend is a Cloudflare Workers + static site stack (`sniffly-site/`) bundled into the same repo — when the hosted service goes away or changes, the share feature silently breaks for everyone who's already shared links. 3. Error 'analysis' is pattern-matched categorization, not root-cause diagnosis; you see that 30% of your errors are tool failures but not why, which limits how actionable the dashboard actually is. 4. No support for team/org aggregation — if you want to compare usage across a team, each person runs their own instance and you share screenshots, which is awkward given the sharing feature already exists.