@kintsugi-ai/mcp

The kintsugi MCP server — one stdio server that gives AI coding agents (claude / codex / agy) everything they need to set up and verify Kintsugi visual regression flows.

Tool groups (single kintsugi alias)

Browser tools — reused from the official @playwright/mcp: browser_navigate, browser_click, browser_type, browser_snapshot, browser_take_screenshot, ... (the full upstream set is passed through unchanged, over an in-memory MCP bridge).

Flow tools — implemented here on top of @kintsugi-ai/core:

| Tool | What it does | | --- | --- | | save_flow | Saves a user flow as ordered steps (navigate/click/type/scroll/wait/screenshot) with stable selectors | | capture_baseline | Replays a flow headlessly and stores the reference screenshots + aria snapshots | | list_flows | Lists recorded flows with step counts | | check_flows | Replays flow(s) and pixel-diffs every step against the baseline |

Running

kintsugi mcp                # via the CLI (what agents are configured with)

or programmatically:

import { startKintsugiMcpServer } from '@kintsugi-ai/mcp';
await startKintsugiMcpServer({ headless: false });

• The agent-driven browser is headed by default so users can watch the agent explore; set KINTSUGI_MCP_HEADLESS=1 for headless.
• The browser viewport matches .kintsugi/config.json, so what the agent validates visually is pixel-identical to the captured baselines.
• Project root resolves from KINTSUGI_PROJECT_DIR or cwd.

Registration

kintsugi init registers this server with each detected agent:

• Claude Code → .mcp.json (project scope, tools appear as mcp__kintsugi__*)
• Codex CLI → .codex/config.toml ([mcp_servers.kintsugi])
• Antigravity → .agents/mcp.json

The launch command defaults to npx -y @kintsugi-ai/cli@latest mcp; override with KINTSUGI_MCP_COMMAND="node /path/to/dist/index.js mcp" for local development.

E2E verification

pnpm -r build
node scripts/e2e-stdio.mjs

Drives the server over stdio exactly like an MCP client: initialize → tools/list (asserts both tool groups) → real browser_navigate + browser_snapshot → full save_flow → capture_baseline → list_flows → check_flows setup loop.