# zapier-dev-mcp An [MCP](https://modelcontextprotocol.io) server for people who **build Zapier integrations**. [Zapier MCP](https://docs.zapier.com/mcp/home) lets AI *use* Zapier apps — this server helps you *ship* one. Point Claude at your [zapier-platform-cli](https://github.com/zapier/zapier-platform) project: - *"What's in this integration?"* — triggers, creates, searches, auth, platform version, request hooks - *"Is it ready for app review?"* — a lint pass encoding the things Zapier's review commonly bounces integrations for - *"Run the validator"* — wraps your project's own `zapier validate` ## Quick start **Claude Code** ```bash claude mcp add zapier-dev -- npx -y zapier-dev-mcp ``` **Claude Desktop** — add to `claude_desktop_config.json`: ```json { "mcpServers": { "zapier-dev": { "command": "npx", "args": ["-y", "zapier-dev-mcp"] } } } ``` Then: *"Lint C:\\code\\my-zapier-app and walk me through fixing the errors."* ## Tools | Tool | What it does | |------|--------------| | `inspect_integration` | Structured summary of the app definition | | `lint_integration` | Review-readiness report: errors / warnings / info, each anchored to a component path | | `generate_test_fixtures` | Ready-to-run bundle mocks (typed `inputData`, `authData` from your auth fields) plus an appTester test skeleton per component | | `validate_integration` | Runs the project's own `zapier validate` (schema + style), with install guidance if the CLI is missing | ## What the linter checks - Missing `operation.sample` on visible actions — **required for public apps**, and the single most common review bounce - Dynamic dropdowns (`"trigger_key.id.label"`) referencing triggers that don't exist - Missing `display.label` / `display.description` / `noun`; unlabeled input fields - Definition `version` vs `package.json` version mismatch; `platformVersion` vs the `zapier-platform-core` dependency - Auth without a `connectionLabel` (users can't tell accounts apart) - Key/map-key mismatches and duplicate keys `hidden` components skip the visibility checks. The linter complements `zapier validate` — schema validation catches malformed definitions; this catches *reviewable-but-rough* ones. ## Notes - Inspecting a project **evaluates its entry module** in a child process (with a timeout and output cap) — exactly what the Zapier CLI itself does. Only point it at projects you trust, i.e. your own. - Computed input fields (functions) can't be linted statically; they're skipped, not flagged. - Not affiliated with or endorsed by Zapier. `zapier-platform-cli` is Zapier's open-source SDK. ## Development ```bash npm install npm test # offline tests — synthetic projects written in-suite npm run build # tsc → dist/ node scripts/smoke.mjs # end-to-end: generates a project, drives the server over stdio ``` Architecture: [`src/extract.ts`](src/extract.ts) (definition extraction via child process) and [`src/lint.ts`](src/lint.ts) (pure review checks) carry the logic; [`src/index.ts`](src/index.ts) is the MCP wiring. ## License MIT