dify-mcp: Let any AI agent build, test and ship Dify workflows programmatically (138 tools, MCP + CLI)

[alexjiaguo]

I built an open-source tool that exposes the entire Dify console API as 138 tools, so AI agents (Claude Code, Codex, Gemini CLI, Cursor, etc.) can create apps, author workflow graphs, validate, test-run, and publish - all without a browser.

What it does:

• 138 tools across 16 namespaces: apps, workflows, providers, plugins, triggers, RAG pipelines, snippets, agent configs, stats, audio, comments, annotations, etc.
• Two surfaces: a CLI (difywf) any shell-capable agent can drive, and an MCP server (stdio) any MCP-compatible host can attach. Same tools, same JSON contract.
• Handles Dify's cookie + CSRF console auth automatically, including refresh-token rotation. (This was the hard part - Dify's console doesn't use Bearer tokens.)
• Safe by default: confirm gates on destructive ops, offline graph validation, audit log, dry-run support.

Quick example - an agent builds and ships a workflow:

difywf app create --mode workflow --name "my-workflow"
difywf wf draft sync <app-id> --graph graph.json
difywf wf test <app-id> --input query="hello"
difywf wf publish <app-id> --yes

Connect to Claude Code:

claude mcp add dify -- difywf mcp serve

Repo: https://github.com/alexjiaguo/dify-mcp
License: Apache-2.0

Live-verified against cloud.dify.ai. Would love feedback from the community - especially on which workflow patterns you'd want agents to automate. Happy to add more tools or node-type schemas based on demand.