❯privacy

❯the short version

• model traffic never routes through our cloud. requests to Anthropic leave from your machine, with your own credentials.
• your credentials stay on your machine. we never see, store, or relay your Anthropic API key or subscription token.
• the free tier sends us nothing. no account, no calls to our servers.
• paid-tier scoring sends chunk text to us. the local proxy sends chunk text plus structural features to our scoring API and gets scores back. details below.
• telemetry is opt-in and off by default. the product is fully functional with it off.

❯what runs where

data plane — your machine. the proxy runs locally. every model request goes straight from your machine to Anthropic, authenticated with your own key or subscription. subscription OAuth tokens are never routed through our cloud, on any tier.

control plane — our cloud. accounts, entitlements, the savings ledger, and the scoring API. the learned model that does the scoring lives only here; its weights never leave our servers, and your model traffic never enters them.

❯what crosses the wire on the paid tier

each turn, the proxy chunks your context locally and sends the chunk text, plus structural features it computed, to our scoring API. the API embeds the text server-side, scores each chunk, and returns scores. the keep/cut decision is applied on your machine — the scoring API is never told which chunks were dropped.

to be explicit: chunk text does leave your machine on the paid tier. an earlier design sent only embedding vectors; as of 2026-07-20 the embedding step runs server-side, so text crosses the wire. we say that plainly rather than keep a claim the architecture no longer supports.

❯the savings ledger

every request writes one row: token counts (billed input, output, cache reads and writes, and the measured counterfactual), the model id, and opaque conversation and session ids. no message text, ever — the row schema has no field that could carry it.

rows are written locally first, to ~/.parsec/ledger.jsonl — that file is what the savings report and status line read. rows ship to your account only after you set an API key; clearing the key stops shipping.

❯training telemetry (opt-in)

the curator improves on traces users choose to share. sharing is tiered, granular, and off by default:

• tier 0 — off. the default. the product is fully functional here; consent by degradation is not consent.
• tier 1 — metrics only. tokens saved, cut percentage, cache ratios, outcome signal, harness version, latency. no content, no file paths, no prompts.
• tier 2 — featurized traces. chunk embeddings plus trace structure, tool names, and token counts. no source code, no prompts — and, honestly: a vector is a mitigation, not anonymity. embedding inversion exists.
• tier 3 — full traces. design partners only, governed by contract.

how consent works:

• we ask after your first savings report — never at install.
• a preview command dumps the exact bytes that would upload, locally, before anything ships. before featurization, a local scrub masks secrets and hashes file paths.
• a persistent status-line indicator shows whenever sharing is on; turning it off stops uploads instantly.
• a tier or schema change re-prompts. consent never silently expands.
• purge deletes your data from the corpus and excludes it from all future training runs. stated plainly: erasing its influence from an already-trained checkpoint requires a retrain.

❯accounts & payments

we build on managed services rather than holding more than we must. the subprocessors:

• Supabase — authentication and the database: accounts, entitlements, ledger rows, and the telemetry consent registry.
• Stripe — billing. card details go to Stripe and never touch our servers.
• Google Cloud — hosting for the control plane.

❯contact

questions about any of this: hello@dasein.rocks.