swatchdog

How it works

one standard · one connection · checked as it builds

swatchdog keeps an AI build on its design standard. The standard can be yours or ours; the check is the same — deterministic, on-demand, and called by your agent on its own output. The whole loop is three steps.

  1. 01

    Set the standard

    Your agent needs a reference to check against. Two ways:

    Bring your own — point your agent at your design tokens (CSS variables, a tokens file, a Tailwind theme). It extracts them and checks against your system. Preview how swatchdog reads them →

    Pick a pack — a complete, curated visual system: tokens, CSS, components, rules, agent prompt. Your agent builds from it and checks against it. From $19 →

  2. 02

    Connect swatchdog

    One connection, in the tool you build in:

    In Claude — Settings → Connectors → Add custom connector. Name it swatchdog, paste this URL, leave the auth fields empty:

    https://swatchdog-connector-970396648818.us-central1.run.app/mcp

    In Claude Code — run this once in your project (Claude Desktop, Cursor & other MCP clients: add the same URL in your MCP config):

    claude mcp add --transport http swatchdog https://swatchdog-connector-970396648818.us-central1.run.app/mcp

    Connecting is free (shared, rate-capped lane). A $12 key adds your own uncapped lane — sent as an Authorization header; your purchase page shows it pre-filled.

  3. 03

    Build — it checks as it goes

    Ask for what you want. As your agent builds, it checks its own output with swatchdog: every off-token color, radius, spacing, and type value comes back with the exact token to use instead. The agent fixes and re-checks; you get the finished build, on standard — and one line telling you what was caught.

What the check is (and isn't). Deterministic and on-demand. swatchdog reads the declared styles and tokens your agent produced — not pixels, not screenshots — and reports the drift with suggested fixes. It never edits your code and never stores it: tokens in, code in, report out. The building and the fixing are your agent's; swatchdog is the instrument.

No design system yet? The packs are complete, coded visual systems — fifteen looks across three families, $19–$49 one-time — that your agent builds from and checks against. They also work with no connection at all: paste a pack's single text file into any chat builder (Base44, Lovable, v0, Replit) and the look comes through. See the packs →

Get the check — $12 See the packs →