🤖 Self-Integrating AI Agent Example

A complete, runnable example of an AI agent that builds its own integrations at runtime

No prebuilt connectors. No manual wiring. Just intent → action.

📖 Read the Self-Integration Manifesto

Quick Navigation:
🚀 Quick Start · 🧠 How It Works · 🛠 Tools · 🖥 UI Components

⭐ Star this repo and make your agents compatible with anything your customers need

---

What is a Self-Integrating Agent?

A self-integrating AI agent can:

• Connect to external applications (CRM, task managers, communication tools, etc.)
• Discover and use available actions (create tasks, send messages, fetch data)
• Automatically build missing integrations via Membrane Agent delegation

This example provides a working implementation using:

• OpenCode - AI agent framework
• Next.js + React - Frontend UI
• Membrane - Integration platform

🚀 Quick Start

⏱️ Time to first integration: ~2 minutes

1. Install Dependencies

npm install

2. Configure Environment

Copy .env.example to .env.local and fill in your credentials:

cp .env.example .env.local

Required variables:

• MEMBRANE_WORKSPACE_KEY - Get from Membrane
• MEMBRANE_WORKSPACE_SECRET - Get from Membrane
• OPENROUTER_API_KEY - Get from OpenRouter. You can also use other LLM providers (Anthropic, OpenAI, Azure) - see the "Configuring LLM Providers" section in opencode-workspace-template/opencode-template.json.

3. Run the Application

npm run dev

This starts:

• OpenCode server at http://localhost:1337
• Next.js app at http://localhost:3000

4. Try It Out

Open http://localhost:3000 and try:

• "Create a task in Linear"
• "Get my recent deals from HubSpot"
• "Send a message in Slack"

---

🧠 Self-integration Skill

The integration capability consists of three parts: a system prompt, custom tools, and UI components.

When implementing this skill for your agent, you need to:

• Add the system prompt to your agent's system prompt.
• Add tools to your agent's list of available tools.
• Implement the custom UI components in your agent UI.

System Prompt

The agent's behavior is defined in opencode-workspace-template/AGENTS.md. It explains to agent how to use tools mentioned below and instructs it to:

• Use Membrane for interacting with external apps: create connections and use tools (actions) to perform user requests.
• Find existing integrations and build new ones if existing ones are not sufficient.
• Interact with Membrane Agent to build integrations on the fly.

---

🛠 Tools

Integration skill uses the following tools:

Connecting External Apps

Tool	Description
list-connections	List all existing connections for the current customer
search-connectibles	Search for things that connections can be created from (integrations, connectors, external apps)
request-test-connection	Request OAuth connection from user (interactive)

Using Tools (Actions)

Tool	Description
search-actions	Search for actions by intent (e.g., "create task", "send message")
run-action	Execute an action by ID

Dynamically Building Integrations

Tool	Description
create-agent-session	Start a Membrane Agent session to build connectors/actions
get-agent-session	Check Membrane Agent session status (supports long polling)
send-agent-message	Send follow-up message to Membrane Agent session
abort-agent-session	Stop a running Membrane Agent session

💡 All tools are implemented in opencode-workspace-template/.opencode/tool/ and can be easily ported to any other tool framework.

---

🖥 UI Components

To work with Membrane, this agent implements a few custom UI components:

Component	Description	File
Request a Connection	Connection UI triggered by request-test-connection tool call when agent needs user to create a connection	connection-request.tsx
Membrane Agent Session Summary	Status widget showing session state (busy/idle/error) right in the agent chat UI	membrane-agent-session.tsx
Membrane Agent Session Details	Sidebar showing full Membrane Agent session logs (when user clicks "Show details" button)	membrane-agent-sidebar.tsx

---

If you're building agents that need to operate in the real world —
this is the missing integration layer.

⭐ Star this repo if you want agents that don't stop at "I can't do that yet".