Show HN: Relay – The open-source Claude Cowork for OpenClaw

SeventeenLabs built Relay as a desktop client for Anthropic’s Claude Cowork, with OpenClaw as the routing backbone — and the pitch lands squarely on the compliance angle that OpenClaw alone can’t fix.

The Austin studio released Relay as an MIT-licensed Electron app (TypeScript, React, Vite) that routes Cowork tasks to any OpenClaw-compatible backend. Where Cowork is limited to Claude and processes files on Anthropic infrastructure, Relay promises model flexibility and data sovereignty. The third promise is the one that draws the regulated-industry audience: a full audit trail with exportable execution history, approval records, and action rationale.

Here’s what the shipped code actually contains, as of version 2026.3.24 on GitHub:

The activity page renders from in-memory chatMessages and coworkMessages — no backend persistence, no database, no audit log. There is no export button. The only export function, handleExportChat, outputs Markdown chat transcripts, not execution history. Timestamps are fabricated client-side: Date.now() minus a fixed offset, not actual execution times. The approval gates are real — safety-page.tsx implements per-action risk scopes (low/medium/high/critical) with a requiresApproval flag persisted to localStorage. CoworkProjectTask tracks six status states — queued, running, needs_approval, approved, rejected, completed, failed — with needs_approval acting as a human-in-the-loop gate before high-risk actions execute. That part works as described.

Anthropic’s own documentation — cited directly in Relay’s README — is unambiguous: Cowork activity is not captured in standard audit logs, the Compliance API, or data exports. The platform was not designed for regulated workflows. That’s not a Cowork bug; it’s a product boundary. Relay attempts to close that gap with a local-first client layer, which genuinely helps the solo practitioner or small team that wants approval gates and local execution history. For enterprise compliance — SOC 2, HIPAA, anything requiring immutable centralized audit records — in-memory approval records on a developer’s laptop are not audit-ready.

Our read: the README’s audit trail language is accurate for individual use, overstated for enterprise. The gap between those two audiences is where this story lives. SeventeenLabs identified real Cowork limitations and built a real client. The audit trail claim in the marketing is doing work the shipped architecture can’t back up.

Repository and primary sources linked below.