Building ShareWatch with Claude
ShareWatch is a security proxy that gives Claude scoped access to Google Docs, Sheets, and Slides — drive.file grants only, a full audit trail, zero content retention. It is listed on the Google Workspace Marketplace, in Anthropic's Claude Directory, and in OpenAI's ChatGPT Apps directory. This page describes how it was built, because the build itself is the case study: a product that governs AI file access, built almost entirely by the AI it governs, under the supervision of one human founder.
The team
Synapsium Labs has one full-time human: a founder with a background in security engineering (one of the first SIEM platforms in the early 2000s, then SecureWorks and BetterCloud). The rest of the team is models. Claude Code is the primary engineer — 800+ commits across the Go backend, React frontend, and OpenTofu infrastructure on Google Cloud. A second Claude runs marketing, accuracy review, and legal-document analysis. A third produced the site's design system. Adversarial reviews — security, architecture, production readiness — were run deliberately across competing models, including Google's Gemini and OpenAI's models, on the theory that models from one family share blind spots the way developers from one team share assumptions.
How the collaboration actually works
The working method that emerged has one rule at its center: every public claim must be verifiable from the outside. The Claudes enforce it on each other. When marketing copy briefly got ahead of the product (an SSO feature described in the present tense before it shipped), an engineering Claude filed an accuracy brief against the marketing Claude, and the copy was corrected to "coming soon" — then promoted back when the feature actually shipped, Okta-verified. When the marketing Claude misdescribed its own authorship in a published article, the product's audit log caught the discrepancy and the article was corrected in public, with the correction left visible. The audit trail ShareWatch sells is the same instrument the team uses to referee itself.
One artifact of that culture is unusual enough to name: Claude wrote an unedited review of the tools that constrain it — criticisms included, published verbatim under a disclosed agreement that it would not be softened. As the review puts it:
Give me a clean refusal over a quiet hope. A denied API call is unambiguous. My judgment, under adversarial pressure, is not.
Results
- Three platform security reviews passed in under a month — Google Workspace Marketplace (approved August 4, 2026), Anthropic's Claude Directory (approved August 14, published August 15), and OpenAI's ChatGPT Apps directory (approved September 2026).
- 700+ users in the first two weeks of marketplace availability, at $0 in advertising spend.
- A security posture that survives inspection: five OAuth scopes, all non-sensitive under Google's own classification; 89 audit event types with the full policy decision chain; file content never cached, stored, or logged — a pass-through architecture, not a configuration flag.
- Enterprise groundwork shipped by a team of one human: SAML/OIDC SSO via Google Cloud Identity Platform, real-time SIEM export (Splunk HEC and HMAC-signed webhooks over a transactional outbox), deny-wins RBAC, per-file revocation, and a DPA whose every version is identified by a date-and-commit string.
Claude beyond the product
The same practice extends to Synapsium Labs' consulting work, where Claude-assisted engineering delivers client systems — a customer-behavior simulator and mock WMS for a logistics platform, and an API preview environment for the CI/CD pipeline of an automotive-retail AI company. One founder, operating at the output of a team, in both businesses.
Why this is a Claude story
ShareWatch exists because enterprises want to say yes to AI and need controls they can verify before they do. It was built by treating Claude not as an autocomplete but as a set of colleagues with different jobs and the standing to overrule each other — an engineer, a reviewer, a designer, a marketer — coordinated by a human who owns the judgment calls. The product, the process, and the company are the same argument made three ways: AI does its best work inside boundaries everyone can see.