Llm Lens

llm-lens-web

 

A local, offline web UI for auditing, pruning, and cleaning the conversation history your LLM CLI has written to disk. Currently supports [Claude Code](https://claude.ai/code); the architecture accommodates other providers (Codex, Gemini) but only Claude is implemented today.

**Local only.** No API key. No auth. No outbound network. Never invokes `claude`. Reads and rewrites `~/.claude/projects/*.jsonl` on your machine, nothing else.

**Status: alpha.** Active development, fast-moving surface. APIs, JSONL marker formats, sidecar layouts, and word-list semantics all change without notice between commits. Pin a version if you depend on any of it. Bug reports and pull requests welcome; expect churn.

Why you'd use it

1. Know what you're spending

Token counts and USD costs come from the actual `message.usage` fields Anthropic returned for each turn — not estimates. Per-model breakdowns, per-project rollups, per-day/week/month buckets. The overview chart shows where your money went and which sessions were expensive. Pricing table in `utils.js` (captured from `claude.com/pricing` on 2026-04-14; update when rates change).

2. Make future `/resume` cheaper

This is the lever that's easy to miss. Anything you remove from a conversation shrinks what Claude Code sends as context the next time you `/resume` it. Less context sent → fewer input tokens billed per turn going forward. The editing tools aren't just cleanup; they're direct downstream cost reduction:

• Edit — rewrite a prose message's text in place; usage, UUIDs, and the resume chain stay intact. For fine-tuned customization. Prose-only: tool/thinking blocks are locked.
• Redact — redact a message's text to .. Original usage is preserved (historical accuracy), but on resume the redacted content is what gets sent.
• **Nor