Who it is for
Design system owners keeping agent-generated interfaces on brand.
Overview
swatchdog compares UI code against a set of design tokens and reports where the code departs from them. Tokens are supplied with each call, either pasted, described, or accepted from a set the tool proposes, and the response is a structured drift report: which axis is at fault, colour, radius, spacing, or type, what value the code actually used, which token it should have matched, and where in the source that violation sits.
It reads declared styles rather than rendered pixels. The publisher states it is stateless, keeping neither tokens nor code between calls, and read-only, reporting drift without editing anything. It works with no key on a rate-capped shared lane, with a one-time personal key removing the cap in clients that can send headers.
Design system owners keeping agent-generated interfaces on brand.
One tool, check_design_drift, and the whole product is that call: candidate code and a reference token set in, a structured list of violations out. Nothing writes, and nothing persists between calls.
Availability
swatchdog is listed in the Claude connectors directory only. swatchdog has no app under the same domain in the ChatGPT directory; 392 of the 3,150 products here are on both marketplaces, and this is not one of them.
The MCP server behind a Claude connector is usually the same server a ChatGPT app would need, so the absence is a distribution decision rather than a technical one.
Claude listing
swatchdog publishes 1 tool to Claude, a deliberately narrow surface; only 0.7% of the 1,253 Claude-listed connectors expose fewer.
The median Claude-listed connector filed under Developer Tools publishes 13, so this one sits 12 tools below its bucket's median, 91st of 95.
It falls in the 1-3 tools band, which holds 16.4% of the Claude directory.
A direct call to the endpoint returned 1 tools, matching the 1 the listing publishes. Probed 2026-08-07.
check_design_driftswatchdog publishes a remote MCP endpoint. Claude connects to it over HTTP; there is nothing to install locally.
Every connector in the Claude directory is remote, all 1,253 of them, so transport is not a differentiator there.
https://swatchdog-connector-970396648818.us-central1.run.app/mcp swatchdog is community-listed in the Claude directory, the tier 770 of 1,253 connectors sit in (61.5%). It is the default for a self-published server, and says nothing either way about quality.
In the publisher's words
swatchdog keeps an AI build on-brand. Give your agent your design system. Paste your tokens, describe them, or approve a set it proposes and swatchdog deterministically checks generated UI code against them, returning a compact, structured drift report: for each violation, the token axis (color / radius / spacing / type), the value found, the expected token and its value, and the location. It reads declared styles and tokens, not pixels. It is stateless. Your tokens and code are passed with each call and never stored - and read-only: it reports drift; it does not modify code. Enforcement is your agent's job; the source of truth is yours; swatchdog is the instrument. Works with no key (shared free lane, rate-capped). A one-time $12 personal key (swatchdog.dev) removes the cap in header-capable MCP clients (Claude Code, Claude Desktop, Cursor). The full pack library and hosted suite are at swatchdog.dev.
Quoted from the connector's own Claude directory listing as captured in August 2026. Node8 did not write it and does not vouch for it.
Nearby
Trust ratings and identity for software agents
Claude · ChatGPT
Load test scenarios, runs, and results on the OctoPerf platform
Claude
Turn a plain-language brief into a Pega application design
Claude
Policy rules for AI coding agent access
ChatGPT
Edit and manage Mintlify documentation sites
Claude · ChatGPT
Check domain availability and pricing at Namecheap
ChatGPT
This is an independent catalogue entry compiled by Node8 from swatchdog's public marketplace listing as it stood in August 2026. Listing facts come from the marketplaces; the comparisons against the rest of the directory and any live-endpoint measurements are Node8's. Node8 is not affiliated with swatchdog, and inclusion here is not an endorsement.
Node8 builds MCP servers end to end: deciding which tools are worth exposing, the authentication and permission model, the review submissions, and the onboarding that gets them used. One server serves ChatGPT, Claude, and Microsoft Copilot.