mcpField
B
blinkbox

blinkbox

Remote
blinkbox.netProductivity3Updated 5d ago

About blinkbox

blinkbox is an open-source workflow automation platform for teams using AI agents to create, run, and debug automations. Its MCP interface exposes workflow management alongside a visual canvas and a catalog of integrations.

Key features

  1. 01Create and manage workflows through an agent
  2. 02Visual drag-and-drop workflow canvas
  3. 03250+ integration nodes
  4. 04AI workflow builder
  5. 05Browser-based scraping and JavaScript sandbox
  6. 06Encrypted credential vault

Use cases

  1. 01Draft an automation from a natural-language task
  2. 02Run workflows and inspect failed execution logs
  3. 03Connect services such as Gmail, Slack, or GitHub
  4. 04Edit agent-created workflows in the visual canvas

MCP support

The MCP server exposes tools to list, create, run, activate, and debug workflows, along with node and credential operations. It uses Streamable HTTP and can connect to clients such as Claude Code, Cursor, and ChatGPT.

Configuration

Remote servers need no local install. Pick your client, then copy the config or install in one click.

Endpointhttps://mcp.blinkbox.net/mcp
Copy install command
Claude CodeTerminal
Cursor.cursor/mcp.json
VS Code.vscode/mcp.json
Claude Desktopclaude_desktop_config.json
Windsurf~/.codeium/windsurf/mcp_config.json

Run commands in a terminal; paste JSON into the client’s MCP config and restart. Servers that need sign-in will prompt on first connect.

All endpoints
Streamable HTTPhttps://mcp.blinkbox.net/mcpAuth required

Tools

Details

From the project README

BlinkBox

Open-source automation platform (Zapier / Make / n8n alternative) with a native MCP server — connect Claude Code, Cursor or ChatGPT and let the agent build, run and debug your workflows.

Free hosted tier at blinkbox.net (no card) · self-host with Node 22 + MongoDB + Redis · MIT.


What it does

  • MCP server — the whole platform is exposed over the Model Context Protocol (Streamable HTTP). Your coding agent lists, creates, runs, activates and debugs workflows with tool calls — see Use it from Claude Code
  • Agent skill — a SKILL.md that teaches the agent node fields, expression syntax and execution rules, so a working flow takes ~4 tool calls instead of ~12 of guessing
  • Drag-and-drop canvas — everything the agent builds is a normal workflow you can open and edit visually
  • AI Agents — LLM nodes that reason over data and output structured results
  • Headless scraping — full Chromium pool, defeats anti-bot, renders JavaScript
  • Code sandbox — write JavaScript in an isolated V8 sandbox with memory limits
  • 250+ integrations — Gmail, Slack, Stripe, GitHub, Notion, Airtable, Shopify, and more
  • Encrypted credential vault — AES-256-GCM; secrets never leave the server decrypted

Use it from Claude Code, Cursor or ChatGPT (60 seconds)

  1. Sign up at blinkbox.net → Dashboard → MCP → create an API key.

  2. Add the server to your client:

    # Claude Code
    claude mcp add --transport http blinkbox https://mcp.blinkbox.net/mcp \
      --header "Authorization: Bearer <your key>"
    
    // Cursor / Claude Desktop / any client that takes a JSON config
    {
      "mcpServers": {
        "blinkbox": {
          "url": "https://mcp.blinkbox.net/mcp",
          "headers": { "Authorization": "Bearer <your key>" }
        }
      }
    }
    

    Clients that can't send headers can pass the key in the URL instead: https://mcp.blinkbox.net/mcp?key=<your key>.

  3. (Recommended) install the skill so the agent knows the node catalog and expression syntax:

    mkdir -p ~/.claude/skills/blinkbox && curl -fsSL \
      https://raw.githubusercontent.com/blinkboxhq/Blinkbox/main/.claude/skills/blinkbox/SKILL.md \
      -o ~/.claude/skills/blinkbox/SKILL.md
    
  4. Ask for what you want:

    "Every morning at 8, find dentists in Berlin on OpenStreetMap that list a website, email and phone, dedupe against my Google Sheet and append the new ones."

    The agent calls create_automation, runs it with run_automation, reads get_execution_logs when something fails, and activate_automation when it's green. OAuth apps (Google, Slack, Notion…) are connected once by you in the dashboard — the agent never sees your tokens.

Tools: list_automations · get_automation · create_automation · run_automation · activate_automation · deactivate_automation · rename_automation · delete_automation · list_executions · get_execution · get_execution_logs · list_nodes · get_node · list_node_actions · list_credentials · create_credential · blinkbox_api_get · blinkbox_api

Self-hosting? The same server is at <your backend>/api/mcp — create keys in your own dashboard.


Stack

LayerTech
FrontendReact 18 + Vite + Tailwind CSS + ReactFlow
BackendNode.js + Express + MongoDB Atlas + Redis
AIAnthropic Claude (primary) → Groq → Gemini (fallback)
ScrapingPuppeteer + Chromium
ExecutionCursor-based distributed engine, 4 worker cells
DeploymentRailway (nixpacks)

Local setup

Prerequisites
  • Node.js ≥ 22
  • MongoDB Atlas cluster (or local MongoDB)
  • Redis (local or Upstash)
1. Clone
git clone https://github.com/blinkboxhq/Blinkbox.git
cd Blinkbox
npm install
2. Configure environment
cp apps/backend/.env.example apps/backend/.env

Edit apps/backend/.env — at minimum you need:

MONGODB_URI=mongodb+srv://...
REDIS_URL=redis://localhost:6379
JWT_SECRET=your-32-char-secret
ENCRYPTION_KEY=your-exactly-32-char-key
ANTHROPIC_API_KEY=sk-ant-...          # enables the AI workflow builder behind create_automation
3. Run
# Terminal 1 — backend (port 3000)
cd apps/backend && npm run dev

# Terminal 2 — frontend (port 5174)
cd apps/frontend && npm run dev

Open http://localhost:5174.


Environment variables

Required
VariableDescription
MONGODB_URIMongoDB connection string
REDIS_URLRedis connection string
JWT_SECRET≥ 32 chars, any random string
ENCRYPTION_KEYExactly 32 chars — used for credential vault AES encryption
AI (at least one recommended)
VariableDescription
ANTHROPIC_API_KEYEnables the AI workflow builder (Claude) used by the create_automation MCP tool
GROQ_API_KEYFallback LLM (Llama 3.3 70B)
GOOGLE_AI_KEYFallback LLM (Gemini 2.0 Flash)
OAuth integrations (optional)
VariableDescription
GOOGLE_CLIENT_ID / GOOGLE_CLIENT_SECRETGmail, Drive, Calendar OAuth
SLACK_CLIENT_ID / SLACK_CLIENT_SECRETSlack OAuth
GITHUB_CLIENT_ID / GITHUB_CLIENT_SECRETGitHub OAuth
MICROSOFT_CLIENT_ID / MICROSOFT_CLIENT_SECRETMicrosoft 365 OAuth
NOTION_CLIENT_ID / NOTION_CLIENT_SECRETNotion OAuth
AIRTABLE_CLIENT_ID / AIRTABLE_CLIENT_SECRETAirtable OAuth
Deployment
VariableDescription
BACKEND_PUBLIC_URLPublic-facing backend URL (e.g. https://api.blinkbox.net) — required for OAuth callbacks
FRONTEND_URLFrontend URL for CORS allowlist
PORTBackend port (default: 3000)

Architecture

┌─────────────────────────────────────────────────────┐
│                    Frontend (Vite)                  │
│  Dashboard · Workspace Canvas · MCP keys            │
└───────────────────────┬─────────────────────────────┘
                        │ REST API
┌───────────────────────▼─────────────────────────────┐
│                Backend (Express)                    │
│  Auth · Automations · Executions · Credentials      │
│  MCP server · AI builder · OAuth · Webhooks         │
└──────────┬──────────────────┬───────────────────────┘
           │                  │
    ┌──────▼──────┐    ┌──────▼──────┐
    │  MongoDB    │    │    Redis    │
    │  Atlas      │    │   Queues   │
    └─────────────┘    └──────┬──────┘
                              │
                    ┌─────────▼───────────┐
                    │  Execution Workers  │
                    │  (4 cursor cells)   │
                    │  Chromium pool      │
                    └─────────────────────┘
Execution engine

Workflows run as cursor-based state machines. Each execution is a MongoDB document with a cursor array. Workers claim cursors atomically via arrayFilters, run the node handler, advance the cursor, and release. Redis queues decouple trigger fan-out from execution. A resumer process recovers crashed executions every 5 seconds.

Node system

Every integration is a node with two halves:

  • Frontend (nodeRegistry.js) — config panel, icon, label, category
  • Backend (nodes/) — stateless handler async (config, context) => result

The same backendType key connects both sides. 250+ nodes are registered.


AI workflow builder

create_automation (the MCP tool) hands your description to a server-side builder:

"When a new Stripe payment comes in, look up the customer in HubSpot, and send a Slack alert to #revenue"

The builder calls Claude with the node knowledge base and forced tool use, and returns a fully configured, canvas-correct workflow — real field values, variable chaining ({{trigger.data.from}}), node positions — as a draft you can run, inspect and activate from the agent or the UI.

Provider fallback chain: Anthropic → Groq → Gemini


Security

  • JWT auth on all API routes, workspace-scoped queries
  • AES-256-GCM credential encryption (key never stored in DB)
  • SSRF guard on all outbound HTTP requests
  • Code sandbox: isolated V8 with memory/time limits (isolated-vm)
  • Shell tool nodes gated behind ENABLE_SHELL_TOOLS=true (off by default)
  • OAuth state tokens (CSRF protection), postMessage to explicit origins only
  • Rate limiting on webhook triggers (Redis-backed, survives restarts)

Deployment on Railway

  1. Create a Railway project, add the repo
  2. Set all required env vars in Railway dashboard
  3. Set BACKEND_PUBLIC_URL to your Railway backend URL
  4. Register that URL as an OAuth redirect URI in Google Cloud Console (if using Gmail)

The nixpacks.toml in apps/backend/ installs Chromium and all native deps automatically.


License

MIT

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.