Security and privacy
What leaves a developer's computer, where it goes, who can read it, and what we have not built yet. Last checked against the product on 15 October 2026 (change this date only after re-checking the facts).
What the agent reads
The StackSignal Agent reads the session logs that AI coding tools already keep on the computer: Claude Code, OpenAI Codex CLI, GitHub Copilot (VS Code, CLI and app), Gemini CLI, Antigravity and OpenCode. It reads them; it never changes them. It does not capture keystrokes or the screen, and it never uploads source files.
What leaves the computer, every 15 minutes, over HTTPS
| Always | Work email and display name; counts and timings of sessions and requests; the AI tool and model used; token counts; practice findings and scores; which repository a working folder belongs to (host/owner/name, with any credentials stripped); committed lines per repository per day; up to six topic keywords for a request someone repeats often. |
|---|---|
| If your organisation keeps excerpts on | Short excerpts of requests: at most 300 characters per request over the last 14 active days, at most 120 characters for a session's first request, at most 1,000 characters of prompt and 1,000 of reply for requests that included a screenshot, and at most 2,000 characters for up to five selected prompts a month. Every limit is a maximum, so a short request can arrive whole. |
| For AI subscriptions | Token counts per message, usage-limit readings (Claude 5-hour and 7-day windows, Codex, a weekly Google reading), which account is signed in and its plan, and the names of the most-used AI tool add-ons. No prompts, replies or code. |
| Never | Whole conversations, files or attachments, images (only a count), commit messages, a cost per person. |
Redaction
Before any text is shortened, the agent replaces credential-shaped strings (API keys, passwords written as name=value, connection strings) and email addresses with labels such as [REDACTED:email]. It is a defence against leaked credentials, not a privacy guarantee: it does not recognise personal or customer information written in ordinary sentences, and code pasted into a request can appear in an excerpt.
Who can read what
| A developer | Their own dashboard only, including their own excerpts. |
|---|---|
| A manager | Figures and excerpts for the people in the team they lead; the company's subscription screens, read-only. |
| Owners and administrators | Everyone. Only Owners change people, roles, teams, machine owners and the excerpt switch. |
| Systems you give an API token | What that token's grants allow; every call by a token that reads people is logged with the address asked about. |
Access is enforced on the server for every request; the browser never receives data its user is not entitled to. Automated tests call every protected route as each role. Five practice rules about working hours and tone never reach a manager's screen or the score, though managers can see when AI was used.
Where your data lives
Each customer gets a dedicated instance with its own database; no data is shared between companies. Enterprise customers can run it in their own cloud or on their own servers. Practice summaries are kept until you ask us to delete them; subscription usage events are kept for 90 days and hourly totals for 13 months.
How it is protected
- All traffic from agents and browsers uses HTTPS.
- Agent updates are signed (Ed25519) and their checksum is verified again just before they run.
- Passwords are stored with argon2id; sessions use secure cookies.
- Microsoft Entra sign-in is available for one Microsoft tenant per instance.
Not in place yet
- No SOC 2 or ISO 27001 report yet; a penetration test is planned before general availability.
- The Windows installer is not yet code-signed, so browsers and SmartScreen show a warning.
- No self-service export or erasure of a person's data: email [email protected] and we delete it.
- Single sign-on other than Microsoft, and multi-factor authentication on password logins, are on the roadmap.
Subprocessors
Our hosting provider for your dedicated instance (named in your order form). Microsoft, only if you switch on Microsoft sign-in or emailed reports. An AI provider only if you switch on one of two optional features: AI-written report summaries (counts, scores and folder and branch names, never prompt text) or AI drafting of a reusable skill (the request's keywords and up to five of that person's 300-character excerpts).
Report a vulnerability
Email [email protected]. See also security.txt.