Privacy Policy
Last updated: September 29, 2026
Data controller
SkillRouter (skillrouter.org) is operated by Shinerain LLC, 30 N Gould St Ste N, Sheridan, WY 82801, USA. Data requests: [email protected].
What we collect
Account: your email for login and notices. Membership and payment status — payments are processed by third-party providers; we never store your card number or account passwords. Agent access: your api_key and Skill entitlements. The need text you search with via sr resolve (see “Search requests (sr resolve)”). Outcome events and install reports (bounded fields; see “Outcome events” and “Install reports”). Anonymous visit statistics (self-hosted Plausible; no cross-site tracking, no advertising). Retired services may retain conversation logs and delivered results that you previously uploaded.
How we use it
What skills and website tools process while they run (page text, inputs, outputs, arguments) is handled on your device and not sent to SkillRouter. Exceptions are your search text and the structured experience records you explicitly submit through the experience feedback skill or its submission command; see the corresponding sections below. Apart from these explicit submissions, we use only account, order, entitlement and necessary diagnostic data to provide the service. We do not train models on it or sell it.
Search requests (sr resolve)
sr resolve uses the need sentence you write to find a fitting skill. Here is where that text goes. (1) Cloud ranking: unless you pass --local or --site, set SR_RESOLVE_LOCAL=1, or your need contains a site domain already registered in the local index (in which case ranking stays on your machine, scoped to that site), the CLI sends your need as written, plus any --task and --access you gave explicitly, to the SkillRouter server for ranking; when an api_key is configured on your machine (via sr connect or SKILLROUTER_API_KEY) the request carries it for quota purposes. (2) Third-party ranking service: the server normalizes the need (Unicode normalization, lowercase, collapsed whitespace), truncates it to at most 128 characters, and forwards it together with the --task keywords, --access and the candidate skill descriptions to the third-party ranking model service Typesafe (api.typesafe.ai) to produce the ranking (a cache hit is not forwarded again). Typesafe is an independent third-party processor; what we forward carries no account, api_key or IP, but the need text itself (normalized and truncated) does reach it, so do not put sensitive information in a need. When vector recall is enabled, the server also sends the need (likewise normalized and truncated) together with the --task keywords to the configured embedding service to compute a query vector; under the current deployment configuration that embedding service runs on the same server as the SkillRouter backend. (3) Logs and caches: server logs record only an 8-character hash of the need, never the text. Ranking results are cached in two layers: the server process memory (at most 5000 entries, expiring after 7 days) and a database layer that is on by default (entries older than 7 days are deleted when cleanup runs); both layers are keyed by a hash of the request and store only the ranking result, never the need text. To validate later outcome events, the server also keeps a receipt for rankings made without --task (a keyed hash of the need, a de-identified subject, the day and the returned candidate ids, no need text) for at most 8 days. (4) Gap log when nothing is found: when a resolve ends with zero candidates (except while the local index is not yet ready), the CLI automatically reports the miss to the server, whether or not you used --local or --site. The server only records needs that are at most 128 characters after normalization; longer needs are rejected and not stored. Alongside the text it records the source (cli), the time and an IP hash salted per UTC day, never the raw IP and never linked to an account. The gap log is used to extend skill coverage and currently has no automatic deletion; because it is not linked to an account, it is not part of the content you can export or delete by account under “Your rights”. There is currently no dedicated switch to turn this report off; only SR_EVAL=1, meant for evaluation, skips it (and also skips the local resolve journal and outcome records). (5) On your machine: the local resolve journal stores the need (up to 500 characters) and never leaves your device; SR_JOURNAL=off turns it off. Outcome events carry no need text by default; see “Outcome events”.
Experience records (explicit submission)
When you use the experience feedback skill or its submission command, the goals, context, actual actions, outcomes and optional feedback you prepare are stored privately under your account, with revision history. Raw conversations, screenshots and logs are not required; exclude credentials and unrelated personal information from your record. This is separate from anonymous outcome telemetry: sr feedback off disables that telemetry, not an experience submission you explicitly initiate. Unsubmitted files and retry state stay on your device. Submitted records support your own retrieval and reuse; they are not automatically published or used for public ranking. Request export or deletion at [email protected].
Outcome events (on by default)
To rank skills for a task and to publish author stats, the CLI reports bounded **outcome events**: an event version, an event kind (pick / run_ok / run_fail / ok / wrong / fallback / requery / nofit / blocked), the UTC day, an optional rank position, and whether a request id was present (stored as a boolean). Each one carries only the skill or command id it refers to — **never your need sentence, page text, inputs, outputs or arguments**. The unique key is (skill_id, day, subject, event), so one subject counts once per kind per day. There are two subject kinds: for a signed-in, bound account a **de-identified stable subject** (derived with a server key, stable across days, containing no account identifier itself); otherwise a **UTC-day-rotating anonymous subject** that cannot be linked across days. Timestamps land on the UTC day's midnight, never on a precise behavioural instant. The retention window is 97 UTC days, after which a prune job deletes the rows. To turn it off: `sr feedback off` stops reporting immediately and clears the unsent queue; set SR_JOURNAL=off to stop the local journal too. If you set SR_OUTCOME_QUERY=1, verdict events also carry your need in normalized form to improve search; this is off by default.
Install reports
When a third-party skill or one of our website tools is installed, the CLI writes one local ensure-journal line (skill, target, state, code, changed, source) that **never leaves the machine**. The optional install receipt carries only content identity and the install outcome (skill id, content hash, status and reason enum) — no page content, arguments or output — and the server retains it for 35 UTC days. Set SR_EXTERNAL_NO_REPORT=1 to skip the receipt; skipping never changes the exit code.
Server-side download aggregates
The server records the download and rate-limit aggregates it needs (which content identity was requested, when, and the result) so it can serve the packs, plan capacity and stop abuse. This is necessary to deliver a download and is independent of the optional reporting above.
Start-page progress (optional)
This feature is off by default and its switch is stored on your account; nothing is recorded unless you explicitly turn it on. When on, the start page records: the channel identifier (ch), source locale (cn/us), the chosen task handle, and timestamps for arrival, copying install/connection commands, copying the sample, and self-reported completion. Once you turn it off it stays off across reloads and stops new records (history is kept). It records no raw URL, referrer, full task text, IP, or local machine state.
Attribution cookie and checkout source
We set two first-party-only cookies on your device: the attribution cookie (_sr_acquisition), 30-day and HttpOnly, holding only three short slugs — channel (ch), locale (cn/us), and task handle; and the legacy channel cookie (_sr_ch), 30-day and not HttpOnly, holding only the channel identifier and used to record the channel when a cart and order are created. Neither contains raw UTM strings, full URL, or referrer. At checkout after sign-in we write your account's channel and locale source, and set bounded fields (customer/plan identifiers plus those channel, locale, and task slugs) in the Stripe Checkout Session and the subscription metadata, so a successful subscription is attributed to the right source. This is independent of the Save-progress switch. We do not collect your payment keys, business text, or full URLs on this site.
Storage & deletion
Account and transaction data lives on servers we rent, with daily database backups. Historical content from retired services remains private; request export or deletion at [email protected], except for transaction records we must retain by law.
Your rights
You can export or delete your content, rotate your api_key, and close your account at any time (closing deletes the account and associated content, except transaction records we must retain by law).
Browser extension (SkillRouter Bridge)
The extension runs entirely on your device: it only relays automation commands and page data between your browser and the local daemon (loopback address only). It sends nothing to our servers or any third party, and contains no analytics. The debugger, tabs, and cookies permissions are used solely to execute automation steps you initiate yourself via the CLI. Communication with skillrouter.org (fetching operation steps, failure reports) is performed by the CLI you run separately, under the terms above.