Privacy policy

Short version: trimwire collects nothing by default. It runs entirely on your machine, and it never sees, stores, or transmits your code or conversations. The only data that ever leaves your machine is an opt-in, anonymized, content-free statistics upload that you turn on yourself — described in full below.

Last updated: 2026-06-11.

The trimwire tool

trimwire is a local proxy. Your Claude Code (or other harness) traffic passes through it on your own machine, on its way to your model provider on your own API key.

• It reads each request only to prune it in memory before forwarding it, then discards it. It does not store your prompts, code, file contents, file names, or model responses.
• It keeps a small local ledger (a SQLite file in your data directory) of content-free counters — bytes saved, cache-hit rate, which strategies fired — the same numbers trimwire stats shows you. This never leaves your machine unless you explicitly opt in to sharing (below).
• It makes no network calls of its own on the default path. (The optional summarizer, if you enable it, calls a model you choose on your own key — see Summarizer.)

Opt-in telemetry (trimwire share)

Nothing is uploaded unless you run trimwire share enable (or pass --yes). You can opt out any time with trimwire share disable. Until you opt in — and until the project’s community collector is deployed — trimwire share stats only prints the payload and sends nothing.

When you do opt in, each upload is a single small JSON of coarse, bucketed, aggregate numbers derived from your local ledger. By design it contains no prompts, code, file paths or names, message text, session IDs, machine or install IDs, raw IP, timestamps finer than a calendar day, or any raw byte/token counts.

How your privacy is protected:

• Content-free. Only ledger-derived metadata is ever sent — never message content or paths.
• Coarsened on your machine. Every percentage and size is bucketed before it leaves, so even the row that reaches the collector is already anonymized.
• No cross-day identity. A random install ID stays on your machine and is never transmitted. Each upload carries only a daily-rotating HMAC token, so uploads on different days cannot be linked to you.
• No IP stored. The collector uses your IP only for transient rate-limiting and never writes it to its database.
• Aggregate-only public dashboard. The public dashboard shows only aggregates across many contributors, with small groups suppressed (k-anonymity), so no individual’s data is surfaced.

The complete, field-by-field list of exactly what is and isn’t sent — the single source of truth — is in Telemetry. A test in the codebase fails the build if the payload ever contains a field not on that list.

Benchmark sharing (trimwire share benchmark) is a separate opt-in upload that scores a model against a bundled synthetic corpus — never your session — and is likewise content-free. See Telemetry.

The website

This site is static and adds no analytics, tracking pixels, or tracking cookies. It’s served by a CDN host (Cloudflare), which — like any web host — processes requests and may keep standard, short-lived access logs for security and abuse-prevention. The community dashboard fetches only the pre-aggregated, anonymized statistics described above.

Changes

trimwire is open source: this policy, the exact telemetry fields, and the code that produces them are all auditable in the repository. Material changes will be noted in the changelog and reflected here.