Privacy
Last updated: 2026-10-06 (draft).
What runs where
TokenPak is a local proxy. Every compression, routing, caching, and telemetry decision happens on your machine. Your prompts, completions, code, and business data are not sent to any TokenPak-operated service: none sits in your request path. By default, the one thing that leaves your machine is the LLM request you ask TokenPak to proxy: it travels to the upstream provider you configure, using your own credentials (see What leaves your machine below).
What we collect by default
By default, nothing: zero telemetry, no usage pings, no install counters. The opt-in debug and logging controls described below can expand what is stored locally on your machine when you enable them — they write only to local destinations (file, stdout, or your system's local syslog socket) and do not initiate any outbound network connection for logging purposes. The few optional network requests TokenPak can make are listed under What leaves your machine.
What TokenPak stores locally on your machine
- A SQLite ledger at
~/.tokenpak/monitor.db(or~/.tpk/monitor.dbon fresh installs) with per-request metadata (model, token counts, latency, cost, cache-origin). Default retention: 90 days of rolling history. - Configuration at
~/.tokenpak/config.yaml(or~/.tpk/config.yamlon fresh installs). Contains profile selection, port, compression knobs. Never contains API keys or prompts.
Optional debug and logging controls
Several opt-in controls will expand what is stored locally if you turn them on. They are off by default. We disclose them here so you know exactly what each knob does before enabling it.
TOKENPAK_DEBUG=1(env var)- Enables verbose debug output on stdout/stderr and debug-level entries in local logs. Debug output may include request headers (never credential values — those are redacted at the logging boundary). Off by default.
TOKENPAK_LOG_ENABLED=1+TOKENPAK_LOG_DESTINATION(env vars)- Enables the structured request logger. Destinations:
file(default path),stdout,syslog. Logger records request metadata — not prompt/response bodies unlessstore_promptsis also enabled. SeeTOKENPAK_LOG_LEVELandTOKENPAK_LOG_RETENTION_DAYSfor tuning. telemetry.store_prompts = true(config flag)- This flag stores prompt and response bodies to local disk. Off by default. Exists for debugging and benchmarking. When enabled, prompt/response content is written to the same retention store as metadata. Clients with regulated data should leave this off.
What leaves your machine
By default, only what you ask TokenPak to proxy:
- LLM requests — forwarded to the provider you configured (Anthropic, OpenAI, etc.), using your credentials on the way out. TokenPak forwards the request to that provider; it does not send a copy to any TokenPak-operated service.
The requests below happen only if you opt in or run the command yourself. None of them sends your prompts, responses, file contents or credentials.
- Update check (asks first)
- TokenPak asks, in an interactive terminal, whether it may check for new releases. The default answer is no, and nothing is requested until you say yes. If you agree, it reads
https://pypi.org/pypi/tokenpak/jsonat most once a day. The request carries no TokenPak project, prompt, completion, usage, credential or file data, but PyPI sees ordinary connection metadata such as your IP address. Turn it off withtokenpak update --disable-checksorTOKENPAK_NO_UPDATE_CHECK=1. Runningtokenpak update --checkyourself makes a one-time request to the same address. - Usage metrics (
TOKENPAK_METRICS_ENABLED=1) - Off by default. If you turn it on, TokenPak can send content-free usage metrics to TokenPak-operated endpoints (
metrics.tokenpak.aiandapi.tokenpak.ai). The data is limited to the date and time of the report, token counts, model names, compression ratio, latency, the profile and invocation mode in use, the TokenPak version, your operating system and Python version, a random install ID and a 24-hour request count. It never includes prompts, responses, file names or API keys. tokenpak fingerprint sync(a command you run)- Sends a structural fingerprint of the text you give it (segment types, token estimates and counts, not the text itself) to
intelligence.tokenpak.aiand receives suggested directives back. Add--dry-runto see exactly what would be sent.
Third parties
TokenPak OSS has no sub-processors, and no TokenPak-operated service sits in your request path. Your prompts, completions, and credentials never reach TokenPak. The only network destination for the requests you proxy is the upstream LLM provider you configure (Anthropic, OpenAI, Google, etc.), reached using your own credentials. The optional requests listed under What leaves your machine are separate from your proxied requests.
Your controls
tokenpak uninstall --softrestores your Claude Code settings, reverses the Codex companion install and stops a running proxy. It keeps your configuration and local data.tokenpak uninstall --hardalso deletes TokenPak's local files, such as configuration, request databases and caches. It keeps your session journal, budget history and saved capsules; delete those yourself if you want them gone. Add--dry-runto either command to see exactly what it would change.- Clients you configured with
tokenpak integrate <client> --applycan be restored withtokenpak integrate <client> --revert. - Unset any optional escape-hatch env var to disable that source of data collection.
- Uninstalling the package (
pip uninstall tokenpak) stops all TokenPak execution on your machine.
Contact
Questions: hello@tokenpak.ai.