HN 표시: Claude Code 사용 제한을 리버스 엔지니어링하려고 했습니다.

claude-meter is a local research proxy for understanding how Claude Code usage maps to Anthropic's hidden quota system. The core problem is simple: - Claude Max / Pro users can see rough usage state, but not the real hidden budget - Anthropic appears to enforce multiple windows such as 5h and7d - Claude Code traffic includes useful usage and rate-limit signals, but they are not surfaced in a way that is easy to study over time claude-meter sits between Claude Code and Anthropic, captures those signals locally, normalizes them, and helps estimate rough budget bands from real traffic. This project is currently an alpha research tool , not a polished end-user product. What that means: - it is already useful for reverse-engineering and observing hidden quota behavior - it is not yet an exact “Claude Max limit meter” - current estimators produce rough ranges, not precise truth The right way to think about the project today is: working local proxy working local data capture working local normalization early-stage estimator The immediate goal is to answer questions like: - what 5h and7d quota windows does Claude Code actually see? - how do different models affect those windows? - do cache reads behave more like cheap API cost than raw token count? - can we estimate a rough 5h budget band with enough confidence to track changes over time? The longer-term goal is bigger: - help users understand their own hidden Claude usage limits - detect changes in Anthropic behavior over time - eventually compare behavior across plans, models, and accounts - transparent pass-through proxy for api.anthropic.com - asynchronous logging so proxying stays ahead of disk writes - all capture is local - color-coded CLI logging for every request (model, tokens, utilization) - periodic status summaries via --status-interval - live web dashboard at http://localhost:7735 with auto-refresh setup subcommand to auto-configure your shell - full raw request / response exchange capture - raw JSONL written under ~/.claude-meter/raw/ - request / response bodies persisted locally - sensitive headers redacted before persistence - raw directories and files created with private permissions - normalized JSONL written under ~/.claude-meter/normalized/ - background derivation from raw exchanges /v1/messages parsing/v1/messages/count_tokens parsing- SSE parsing for real streamed Claude responses - best-effort handling for partial gzip event streams - header-driven fallback records for unknown endpoints From live Claude Code traffic, claude-meter can already capture: anthropic-ratelimit-unified-* headers5h and7d utilization windows- model-specific windows such as 7d_sonnet when present input_tokens output_tokens cache_creation_input_tokens cache_read_input_tokens - model names - session ids - declared plan tier The current analysis layer can already: - summarize observed windows - compare raw-token and price-weighted usage formulas - build interval-based estimates - compute filtered 5h estimate bands - estimate dollar budgets per 5h and 7d window using API pricing - show token usage breakdown (input, output, cache read, cache create) - report current and peak utilization per window - output a human-readable summary via --summary flag These are important gaps, not footnotes: - no exact hidden-limit estimator - no high-confidence scoring yet - no automatic reset-aware experiment mode - no packaged installer or background service - no built-in anonymized sharing flow - no cross-account comparison yet - no strong claim that the current estimate band is “the true limit” In particular, the estimator still has outliers. That is expected at this stage. The data capture is real; the hard part is interpreting a coarse hidden utilization meter. The current working thesis is: - raw token counting is too naive - price-weighted usage appears to explain Anthropic's hidden 5h meter better than raw token totals - cache reads should likely be weighted much more cheaply than fresh input or output - the useful product output is probably a rough 5h budget band , not a fake exact cap So the project is moving toward: - “you probably have roughly this much 5h budget” - not “you have exactly N tokens remaining” - Claude Code is pointed at the local proxy with ANTHROPIC_BASE_URL - The proxy forwards all traffic upstream unchanged - Raw exchanges are stored locally - A background normalizer derives structured records - Offline analysis scripts estimate behavior from those records This separation is deliberate: - live traffic path stays simple - raw capture remains the local source of truth - normalization can evolve - estimator logic can be rewritten without touching the proxy curl -sSL https://raw.githubusercontent.com/abhishekray07/claude-meter/main/install.sh | bash git clone https://github.com/abhishekray07/claude-meter.git cd claude-meter go build -o claude-meter ./cmd/claude-meter ./claude-meter start --plan-tier max_20x claude-meter setup This detects your shell (bash, zsh, fis