01Who we are
Claros is run by Timur Monasypov, operating as Claros, Massachusetts, USA. No company has been formed yet. Questions and requests: the contact on the home page.
02This website
No ad trackers. Page views are counted by Vercel Web Analytics, which sets no cookies. Fonts are served from this site.
“Book a demo” either opens a scheduling page run by its own provider, or writes an email in your own mail app. This site does not store what you type.
Cookies: claros_session after you sign in, and claros_preview after you enter the preview password. Both are httpOnly and signed or encrypted.
03Connected tools and documents
When a workspace admin connects Slack, GitHub, Google Drive, Notion or Linear, or imports an export or a policy, Claros reads decision-bearing items (approvals, reverts, review comments, corrections, policies) with their authors and who can see them, and drafts suggested rules with an owner. Tokens are encrypted at rest and never shown again.
To draft rules, each item's text (up to 1,500 characters) goes to the model with personal data redacted: emails, phone numbers and keys become placeholders, and the people involved appear only as “[person 1]”, “[person 2]”. Their handles and emails stay on the Claros server, which uses them to name each rule's owner. A suggested rule is visible only to people who could see one of its sources. Nothing acts until its owner confirms it.
04Claros for Mac (preview)
- Reads only at a commit moment (Send, Submit, Approve, Pay, Merge, Delete, ⌘↩): the app, window title, page address, the text of the field being sent, and the button's name.
- Redacts on the Mac before anything is sent: emails, phone numbers, card numbers, IBANs, SSNs and keys become placeholders.
- Sends the action only, to be checked. Nothing is recorded: no screenshots, audio, keystrokes, clipboard or history. Password fields and excluded apps are never read.
- When a check asks, the question reaches the rule's owner in their notch, a Slack DM or an email, as one line with a link to decide.
05Agent checks
Each agent call that is checked sends the app, the tool and the record's fields, with the agent and the person it acts for. Before any model sees a record, personal data in its text is replaced with placeholders. The check log stores the record PII-redacted unless your workspace chooses to keep it as sent.
Learning from your corrections is off unless you turn it on (getclaros --learn-corrections, or later in Settings). Then a hook in your coding agent reads each prompt on your machine and keeps only corrections (“no, use our Button, not a div”). It sends the correction and the agent action it corrects, redacted, and the example images it references unless you chose no images. Images are read by a vision model that is told not to write out personal data; an image is kept only if you chose to keep them. What Claros learns becomes a suggestion only you see until you keep it.
06Recording rules by voice
Only when an editor chooses to record rules by voice: they share one window and talk. Change detection runs in their browser and only key moments are sent, with personal data on screen blurred before any model sees a frame. Microphone audio goes to ElevenLabs for live speech; the transcript goes to the Claros server. Transcripts and screen text are not blurred before they reach the models below.
“Off the record” pauses watching and listening; “strike that” deletes the last 30 seconds. Voice clips are kept only if the editor turns them on (off by default).
07What we store
On the Claros server: rules and their versions, connector items and suggested rules, checks, asks and answers, and, from voice recordings, transcripts, blurred keyframes and any consented clips.
The evidence log (checks, approvals, overrides, outcomes, rule versions) is hash-chained and signed. It is kept until a workspace admin sets a retention period, at least 183 days. The preview has no fixed retention for other data yet; we delete it when you ask or when the preview ends.
08Who processes it
We don't sell data or use it for advertising. These providers process data so Claros can work (details in the DPA):
| Provider | Purpose | Data |
|---|---|---|
| WorkOS (AuthKit) | Web sign-in: email codes, Google and Microsoft sign-in, invitation emails | Your email address and name, sign-in events with IP address and browser |
| Isoquant (GLM-5.3-Flash) | Main model: drafting rules from connected tools and policies, judging and routing checks, reading the screen when an editor records rules by voice | Connector items with their authors (not PII-redacted), policy text, PII-redacted check records and past decisions, redacted screen frames and transcripts of voice recordings |
| OpenRouter, Google Gemini (if configured) | Fallback models, only when the main model fails | Same as the main model |
| ElevenLabs | Live voice, only while an editor records rules by voice | Their microphone audio, the transcript, Claros's spoken replies |
| Jina AI, Exa, Bright Data (if configured) | Search over rule sets, ranking, looking up public app docs | Rule text, search queries, public URLs |
| Fastino, Cloudflare Workers AI (if configured) | Small, fast decisions during live voice sessions | Short text snippets |
| Resend, or your own SMTP server (if configured) | Email to a person who is asked to decide | Their email address, a one-line summary of the action, a signed link to decide |
| Tailscale (Funnel) | Relays the public connection to the self-hosted Claros server | Encrypted traffic only: TLS ends on the Claros server, so it sees addresses, not content |
| Vercel | Hosts the web app and counts page views (Web Analytics, no cookies); passes sign-in, account and /mcp requests on to the Claros server | Web traffic, including the requests it passes on |
Slack DMs for asks go through your workspace's own Slack connection.
09Changes
We'll update this page and its date when anything above changes.