Cachecatch
The first prompt-cache audit and optimization tool for AI agents.
- Run audit β 2. Run share β 3. Post on π
npx cachecatch@latest audit local --window 7d β npx --yes cachecatch@latest share

Cachecatch audits prompt-cache efficiency across two worlds: local IDE agent sessions (Claude Code, Codex, OpenCode on your machine) and production platform traces (LangSmith, Langfuse, Braintrust). Same engine, same report schema, same π banner. It finds the exact tokens breaking your cache prefix, estimates recoverable spend, and gives you the prompt-layout fix to ship.
Local IDE Agent Report
Audit your local coding agent sessions β no API key, no network, no config:
npx cachecatch@latest audit local --window 7d
Then generate your π banner and share it:
npx --yes cachecatch@latest share --handle @yourname
share fetches your X profile picture, renders a 1024x732 banner with your audit data, and saves it as cachecatch-x-share.png. It automatically picks up the most recent reports/ JSON β so the flow is just: run audit local, then share. Chrome is pre-warmed in the background on the first run, so the second command runs instantly with no install prompts. To share a specific report instead, pass its path:
npx cachecatch@latest share audit.json -o ./my-card.png
npx cachecatch@latest share local-report.json -o ./my-local-card.png
What it shows
- Agentic sessions parsed, token activity, tool calls, subagent runs
- Cache read profile when exact cache-token telemetry is present
- Context hygiene score (volatile data near prompt prefix)
- Recoverable cost estimate
- IDE agents used (Claude Code, Codex, OpenCode)
Expand the window
npx cachecatch@latest audit local --window 30d
npx cachecatch@latest audit local --window 30d --json ./local-report.json
Scope to one repo
npx cachecatch@latest audit local --project /path/to/repo --window 7d
Why some agents show cache telemetry and others do not
Local IDE agents expose different levels of local data:
- OpenCode exposes local token/cache telemetry directly, so Cachecatch can report observed cache-read percentage from local fields.
- Codex can expose token/cache fields through local JSONL token events and future OTel logs depending on version/config.
- Claude Code can expose usage, cost, cache, and tool telemetry through OTel when enabled.
- Cursor and some other IDEs may only expose transcript/context history.
Cachecatch separates visibility into exact cache telemetry, token telemetry only, transcript context only, and unavailable. It never treats missing cache telemetry as zero, never invents cache-read percentage, and never invents cost upside when the model/pricing/token basis is missing.
Enable Codex telemetry
npx cachecatch@latest init codex
npx cachecatch@latest daemon
codex
npx cachecatch@latest audit local --window 7d
Enable Claude Code telemetry
npx cachecatch@latest init claude
source ~/.cachecatch/claude-code-otel.env
npx cachecatch@latest daemon
claude
npx cachecatch@latest audit local --window 7d
Debug local telemetry
npx cachecatch@latest debug codex-telemetry
npx cachecatch@latest debug claude-telemetry
npx cachecatch@latest telemetry status
Platform Trace Report
Audit production agent traces from LangSmith, Langfuse, or Braintrust:
npx cachecatch@latest audit "your-project" --provider langsmith --window 7d
What it shows
- Traces analyzed, routes detected, cache-read rate
- Estimated recoverable cache loss ($)
- Top leaking routes with exact fix instructions
- Prompt layout issues (what's breaking prefix stability)
See which projects your key can access
npx cachecatch@latest projects --provider langsmith
Set the key once in your shell
export LANGSMITH_API_KEY="lsv2_..."
npx cachecatch@latest audit "your-project" --provider langsmith --window 7d
Langfuse
export LANGFUSE_PUBLIC_KEY="pk-lf-..."
export LANGFUSE_SECRET_KEY="sk-lf-..."
npx cachecatch@latest audit "your-project" --provider langfuse --window 7d
Or pass the key directly
npx cachecatch@latest audit "your-project" --provider langsmith --window 7d --key "$LANGSMITH_API_KEY"
Why Token Spend Matters
Cache-read tokens cost a fraction of input tokens. When your prompt assembly is unstable β timestamps, request IDs, or user data appearing before stable instructions β the provider sees every request as unique. You lose the cache discount entirely.
At production scale, that's real money. Cachecatch finds the exact tokens that are breaking your cache prefix and tells you where to move them.
Sample (No API Key)
Run a realistic demo report with no network access:
npx cachecatch@latest sample
npx cachecatch@latest sample --compact
npx cachecatch@latest sample --full
npx cachecatch@latest sample --explain-math
npx cachecatch@latest sample --out ./cachecatch-report.html
Export to HTML
npx cachecatch@latest sample --json > audit.json
npx cachecatch@latest export audit.json --format html --out ./cachecatch-report.html
Or directly from a report:
npx cachecatch@latest sample --out ./cachecatch-report.html
Supported Providers
| LangSmith | Primary | LANGSMITH_API_KEY |
| Langfuse | Covered | LANGFUSE_PUBLIC_KEY + LANGFUSE_SECRET_KEY |
| Braintrust | Covered | BRAINTRUST_API_KEY |
CLI Commands
cachecatch | Show quick start |
cachecatch sample | Render a deterministic sample report |
cachecatch sample --compact | Short executive summary |
cachecatch sample --full | Full route diagnostics |
cachecatch sample --json | Raw CachecatchReport JSON |
cachecatch sample --out ./report.html | Export sample as HTML |
cachecatch audit local --window 7d | Scan local Claude Code, Codex, OpenCode sessions |
cachecatch audit local --project /path/to/repo --window 7d | Restrict local audit to one repo |
cachecatch debug codex-telemetry | Inspect Codex local telemetry fields without raw prompts |
cachecatch debug claude-telemetry | Inspect Claude Code local telemetry fields without raw prompts |
cachecatch init codex | Configure Codex OTel to the local Cachecatch daemon |
cachecatch init claude | Write a safe Claude Code OTel env file |
cachecatch daemon | Receive local OTLP logs/metrics on localhost |
cachecatch telemetry status | Show daemon/config/event visibility |
cachecatch run claude | Launch Claude Code with the generated telemetry env |
cachecatch audit "project" --provider langsmith --window 7d | Run a live platform audit |
cachecatch audit "project" --json | JSON output for automation |
cachecatch projects --provider langsmith | List projects visible to a provider key |
cachecatch config set-key langsmith <key> | Save provider key to local .env |
cachecatch config set-key langfuse publicKey:secretKey | Save Langfuse keys |
cachecatch config get | Show redacted local config |
cachecatch export audit.json --format html --out ./report.html | Convert saved report JSON to HTML |
cachecatch share | Generate a shareable X card PNG |
cachecatch --help | Show CLI help |
All report commands support --no-color for plain terminal output.
Requirements
- Node.js 18+
- An observability provider API key for platform audits
- Rendered prompts and token usage in traces for high-confidence platform reports
- Chrome / Chromium is auto-downloaded by
share (and pre-warmed by every other command), so no separate browser install is needed
Privacy
- Runs locally from your terminal.
- Does not store prompts, traces, or reports unless you explicitly write an output file.
- Reads API keys from flags, environment variables, or a local
.env.
- Does not log API keys.
- The OTel daemon binds to localhost by default and does not send data anywhere external.
- Codex setup uses
log_user_prompt = false.
- Claude setup does not enable user prompts, assistant responses, tool content, or raw API bodies by default.
- Claude tool details are opt-in with
npx cachecatch init claude --include-tool-details.
- Raw OTLP bodies are not stored unless you explicitly run
npx cachecatch daemon --debug-raw.
- The web app path audits server-side; the browser only receives the generated report.
Where data lives
Cachecatch is local-first. Your API keys and trace data never leave your machine except to the provider you explicitly point it at.
Network calls Cachecatch makes:
| Your provider (LangSmith / Langfuse / Braintrust) | only during audit | your provider API key (read from env or .env) |
https://unavatar.io/x/<handle> | only during share | the public X handle you pass (no auth) |
There is no Cachecatch server, no analytics, no phone-home, no cookies, no telemetry from the CLI itself.
Files Cachecatch writes to your disk:
| Auto-saved JSON reports | ./reports/cachecatch-*.json (in the directory you ran the command from) | Gitignored. Contains the report only β no API key. |
| Exported HTML | ./cachecatch-report.html (or path from --out) | Gitignored. |
| X card PNG | ./cachecatch-x-share.png (or path from --out) | Visible to you by design. |
API keys from config set-key | ./.env (in CWD) | Gitignored. Read on next run. |
| Local OTel telemetry | ~/.cachecatch/telemetry/<agent>/*.jsonl | Only if you opt in via init + daemon. Local token/cost events from your IDE agent sessions. |
| Daemon PID file | ~/.cachecatch/telemetry/daemon.pid | Only if you run daemon. Auto-cleared on shutdown. |
| Claude Code OTel env | ~/.cachecatch/claude-code-otel.env | Only if you run init claude. |
| Codex OTel config | ~/.codex/config.toml (edited in place) | Only if you run init codex. |
To delete everything Cachecatch left on your machine:
rm -rf ./reports ./.env ./cachecatch-x-share*.png ~/.cachecatch
Troubleshooting
Missing project
npx cachecatch@latest audit "your-project-name" --provider langsmith --window 7d
Use projects if you are unsure which names your key can see.
Missing API key
export LANGSMITH_API_KEY="lsv2_..."
Langfuse:
export LANGFUSE_PUBLIC_KEY="pk-lf-..."
export LANGFUSE_SECRET_KEY="sk-lf-..."
No trace data found
Try a wider window:
npx cachecatch@latest audit "your-project-name" --provider langsmith --window 30d
Confirm that your traces include rendered prompts and LLM token usage.
Export says no report JSON was provided
npx cachecatch@latest sample --json > audit.json
npx cachecatch@latest export audit.json --format html --out ./cachecatch-report.html
JSON output for CI
npx cachecatch@latest sample --json > audit.json
No spinner or status text is printed in JSON mode.
Support CacheCatch
CacheCatch is a small tool for finding wasted AI agent token spend, prompt cache misses, and recoverable cost gaps.
If it helped you save money, debug your agent workflow, or understand your token usage better, you can support the project here.
Crypto wallets
Solana
SOL or USDT on Solana / SPL:
r4yxhFNtnSrZi9YSTpH326WavHengxbFTmfo7Ud1M81
Ethereum
ETH or USDT on Ethereum / ERC-20:
0xC3B39db02BF5D3fEFaAdd0FD1AE687B3f3F48713
TON
TON or USDT on TON:
UQAeQz7NomoaAAXlG4-XDifSQ7QC5DSqVr19QvfnqctvTJsq
Development
npm install
npm run build
npm run lint
npm test
npm run test:live
npm run cachecatch -- sample
npm run dev | Start the Next.js web app |
npm run build:cli | Compile the CLI to dist/index.js |
npm run build | Build CLI and web app |
npm run typecheck | Run TypeScript checks |
npm run lint | Run ESLint |
npm test | Run engine, adapter, HTTP plumbing, and CLI tests |
npm run test:live | Run live provider smoke tests when real keys are set |
npm run cachecatch -- sample | Run the local CLI |
Architecture
src/
bin/ CLI entry point and commands
adapters/ LangSmith, Langfuse, Braintrust, and mock provider I/O
engine/ Provider-agnostic trace analysis plus local IDE session audit
reporting/ Terminal, HTML, and X card renderers
types/ Shared CachecatchReport and NormalizedTrace types
util/ HTTP and environment helpers
The CLI and web app share the same engine and CachecatchReport schema. Provider-specific HTTP code stays in src/adapters/*; cache analysis stays provider-agnostic in src/engine/*. Local IDE agent scanning is implemented in src/engine/local-agent-audit.ts and produces a LocalAgentReport. X card banners are generated from src/reporting/x-card.ts or src/reporting/x-card-local.ts (HTML templates) and src/reporting/html-to-png.ts (Puppeteer screenshot).
Uninstall / Re-test from 0
To fully wipe Cachecatch and re-test a freshly published version:
rm -rf ./reports ./.env ./cachecatch-x-share*.png ~/.cachecatch
npx clear-npx-cache
rm -rf ~/.npm/_npx
npx --yes cachecatch@latest --version
License
MIT