AgentStack MCP

One MCP server. A whole backend for your AI agent.

Data, auth & RBAC, payments & wallets, CRM, hosting, RAG, bots, scheduling and automation: 700+ actions behind a single tool, agentstack.execute, at https://agentstack.tech/mcp.

Website · MCP docs · Pricing · Action catalog · Status

Why AgentStack

• One tool, the whole stack. agentstack.execute runs a batch of catalog actions (auth.*, projects.*, crm.*, hosting.*, payments.*, rag.* …) in a single call. No tool-menu bloat in your context window.

• Remote first. Nothing to host: point your client at https://agentstack.tech/mcp (MCP Streamable HTTP).

• OAuth or API key. Sign in in the browser (OAuth 2.1 + PKCE, dynamic client registration), or send X-API-Key.

• Works everywhere. Cursor, Claude (Desktop & Code), VS Code, Windsurf, Gemini CLI, ChatGPT / OpenAI and any MCP client. stdio-only clients use the tiny bridge in this repo: npx -y @agentstack/mcp.

Related MCP server: NiPlex MCP Server

Quick start

1. Add the server (pick your client below). The remote URL is all most clients need:

https://agentstack.tech/mcp

2. Authenticate. Either sign in when your client opens the OAuth window, or create an API key at agentstack.tech/user/profile?tab=api and export it:

export AGENTSTACK_API_KEY=ask_...

3. Ask your agent: "Use AgentStack: show my profile and projects."

Or let the CLI write the config for you:

npx -y @agentstack/mcp init cursor            # or: claude-desktop | vscode | windsurf
npx -y @agentstack/mcp doctor                 # read-only connectivity / auth check

Install by client

Deeplink (config is base64 of {"url":"https://agentstack.tech/mcp"}):

cursor://anysphere.cursor-deeplink/mcp/install?name=agentstack&config=eyJ1cmwiOiJodHRwczovL2FnZW50c3RhY2sudGVjaC9tY3AifQ%3D%3D

Or add to ~/.cursor/mcp.json (global) or .cursor/mcp.json (project):

{
  "mcpServers": {
    "agentstack": { "url": "https://agentstack.tech/mcp" }
  }
}

With an API key from your environment instead of OAuth: examples/cursor/mcp.with-api-key.json, or npx -y @agentstack/mcp init cursor --with-key. Full plugin (rules, skills, commands): agentstacktech/cursor-plugin.

# OAuth sign-in on first use
claude mcp add --transport http agentstack https://agentstack.tech/mcp

# …or with an API key
claude mcp add --transport http agentstack https://agentstack.tech/mcp \
  --header "X-API-Key: $AGENTSTACK_API_KEY"

Then run /mcp inside Claude Code. Project-scoped .mcp.json: examples/claude-code/.mcp.json. Plugin: agentstacktech/claude-plugin.

Use the stdio bridge. Edit claude_desktop_config.json (macOS ~/Library/Application Support/Claude/, Windows %APPDATA%\Claude\, Linux ~/.config/Claude/):

{
  "mcpServers": {
    "agentstack": {
      "command": "npx",
      "args": ["-y", "@agentstack/mcp"],
      "env": { "AGENTSTACK_API_KEY": "ask_REPLACE_WITH_YOUR_KEY" }
    }
  }
}

Or run npx -y @agentstack/mcp init claude-desktop and replace the placeholder (--embed-key copies $AGENTSTACK_API_KEY in and sets the file to 0600). Restart Claude Desktop. Alternatively add https://agentstack.tech/mcp as a custom connector in Claude's settings where your plan supports it.

vscode:mcp/install?%7B%22name%22%3A%22agentstack%22%2C%22type%22%3A%22http%22%2C%22url%22%3A%22https%3A%2F%2Fagentstack.tech%2Fmcp%22%7D

Or .vscode/mcp.json (the key is asked once and stored as a secret):

{
  "inputs": [
    { "type": "promptString", "id": "agentstack-api-key", "description": "AgentStack API key", "password": true }
  ],
  "servers": {
    "agentstack": {
      "type": "http",
      "url": "https://agentstack.tech/mcp",
      "headers": { "X-API-Key": "${input:agentstack-api-key}" }
    }
  }
}

npx -y @agentstack/mcp init vscode [--with-key] writes this for you. Extension: agentstacktech/vscode-plugin.

~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "agentstack": {
      "serverUrl": "https://agentstack.tech/mcp",
      "headers": { "X-API-Key": "${env:AGENTSTACK_API_KEY}" }
    }
  }
}

Or npx -y @agentstack/mcp init windsurf --with-key. Drop headers to use OAuth.

~/.gemini/settings.json:

{
  "mcpServers": {
    "agentstack": {
      "httpUrl": "https://agentstack.tech/mcp",
      "headers": { "X-API-Key": "$AGENTSTACK_API_KEY" }
    }
  }
}

Extension: agentstacktech/agentstack-gemini-extension.

• ChatGPT connectors: add https://agentstack.tech/mcp as a custom connector with OAuth. AgentStack supports dynamic client registration (/mcp/.well-known/oauth-register), scopes openid tools:read tools:invoke.

• Responses API: use a remote mcp tool with server_url: "https://agentstack.tech/mcp" and an X-API-Key header.

Details and a request example: examples/openai/. Plugin: agentstacktech/gpt-plugin.

curl walkthrough: examples/generic/.

Example prompts

• "Use AgentStack to show my profile and list my projects."

• "Create a new AgentStack project called pet-shop and give me its project id."

• "Publish this site/ folder as a static site on AgentStack and give me the live URL."

• "Add a CRM contact for jane@example.com and move her deal to the Qualified stage."

• "Set up a subscription plan with a 7-day trial and gate the pro feature behind it."

• "Ingest docs/ into a knowledge base and answer: how do refunds work?"

• "When a new order is paid, send a Telegram message to the shop owner."

• "What should I work on next in this project?" (Work Graph: agents.work_next)

The agent discovers the right actions itself (discovery.search, live catalog at agentstack.tech/mcp/actions).

The bridge CLI

npx -y @agentstack/mcp                  stdio bridge → https://agentstack.tech/mcp (default)
npx -y @agentstack/mcp init <client>    cursor | claude-desktop | vscode | windsurf
      --mode remote|stdio  --with-key  --project  --embed-key  --config <path>  --dry-run
npx -y @agentstack/mcp doctor           read-only check: initialize, tools/list, auth
npx -y @agentstack/mcp links            print one-click install links

The bridge is a transparent pipe: tools, prompts, resources and notifications come straight from the server. init merges into existing configs, keeps a .bak backup and refuses to touch invalid JSON.

Security

• Prefer OAuth: URL-only configs contain no secret; your client stores the token.

• Keep API keys in environment variables or client secret inputs, never in committed files. .env*, .cursor/mcp.json and .vscode/mcp.json are git-ignored here by default.

• Use project-scoped, least-privilege keys; rotate at agentstack.tech/user/profile?tab=api.

• The connector never prints key values. Report vulnerabilities privately: see SECURITY.md.

Pricing

AgentStack has a free tier; paid business and personal plans add higher limits. See agentstack.tech/pricing for current plans.

FAQ

Is this the official AgentStack MCP server? Yes. This is the official connector for the AgentStack MCP server. The server itself is hosted at https://agentstack.tech/mcp.

Is the AgentStack core open source? No. The AgentStack platform and MCP server core are closed-source, proprietary software. This repository (MIT) contains only the connector: the stdio bridge, the init/doctor CLI, registry manifests, example configs and docs.

Do I need to install anything? Not for clients with remote MCP support (Cursor, Claude Code, VS Code, Windsurf, Gemini CLI, ChatGPT). Only stdio-only clients (e.g. Claude Desktop config files) need npx -y @agentstack/mcp.

Why only one tool? agentstack.execute takes a list of steps ({id, action, params}) plus a context (e.g. project_id), so one call can run several actions (up to 50 steps per call). Listing tools works without a key; calling them needs OAuth or an API key.

How many actions are there? 700+ public actions across dozens of domains; the live numbers are at /mcp/actions/summary.

Where are the docs? agentstack.tech/mcp-docs and github.com/agentstacktech/AgentStack.

Contributing & license

Issues and PRs for the connector are welcome: CONTRIBUTING.md · CHANGELOG.md.

Connector code: MIT. The AgentStack service, name and logos are proprietary and not covered by the MIT license.