Overture

🔥 The Problem

Every AI coding agent today — Cursor, Claude Code, Cline, Copilot — works the same way:

What Happens Now

1. You type a prompt
2. Agent immediately starts writing code
3. You have zero visibility into what it's doing
4. You realize it misunderstood your request
5. Hundreds of lines of code need to be discarded
6. You've wasted tokens, time, and patience

Text Plans Don't Help

Some agents show plans as text in chat. But text fails to show:

• Dependencies — which tasks depend on what?
• Branch points — what alternative approaches exist?
• Context requirements — which files, APIs, or secrets are needed?
• Complexity — which steps are risky?
• Progress — what's done, what's next?

✨ The Solution

Overture intercepts your AI agent's planning phase and renders it as an interactive visual flowchart — before any code is written.

The agent doesn't write a single line of code until you approve the plan.

🚀 Installation

Overture is an MCP server that works with any MCP-compatible AI coding agent. One command to install.

Claude Code

claude mcp add overture-mcp -- npx overture-mcp

Cursor

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "overture": {
      "command": "npx",
      "args": ["overture-mcp"]
    }
  }
}

Cline (VS Code Extension)

Open VS Code settings → search "Cline MCP" → add:

{
  "mcpServers": {
    "overture": {
      "command": "npx",
      "args": ["overture-mcp"]
    }
  }
}

GitHub Copilot

Create .vscode/mcp.json in your project root:

{
  "servers": {
    "overture": {
      "command": "npx",
      "args": ["overture-mcp"]
    }
  }
}

Note: GitHub Copilot MCP requires VS Code 1.99+ and uses servers instead of mcpServers.

Sixth AI (VS Code Extension)

Add to your Sixth AI MCP settings file:

{
  "mcpServers": {
    "overture": {
      "command": "npx",
      "args": ["overture-mcp"],
      "disabled": false
    }
  }
}

Global Installation (Optional)

npm install -g overture-mcp

Verify It Works

Give your agent any task. Overture automatically opens at http://localhost:3031 with your plan ready for approval.

🎯 How It Works

The Flow:

🛠 Features

Interactive Plan Canvas

Node Details Panel

Click any node to reveal its full details:

Dynamic Fields (User Inputs)

Nodes can request input from you before execution:

Each field includes:

• Required/optional indicator
• Default values
• Help text & descriptions
• Setup instructions ("How to get an API key")

File Attachments

Attach context files to specific nodes:

• Automatic type detection — Image, code, document, or other
• Visual icons per file type
• Descriptions — add notes about why this file matters
• Delete — remove unwanted attachments

Meta Instructions

Add custom LLM instructions to any node:

"Pay special attention to error handling here" "Use the existing auth pattern from src/auth.ts" "Make sure to add tests for edge cases"

Instructions are sent to the agent right before that node executes.

Branch Detection & Selection

When the agent proposes multiple approaches:

Requirements Checklist

Before you can approve, Overture shows what's needed:

• Empty required fields — counted per node
• Branch selections — which decisions are pending
• Progress indicator — visual completion tracking
• Expandable items — click to see details
• Color coding — Green (done) / Orange (pending)

The Approve button stays disabled until all requirements are met.

Execution Controls

The approval button is smart:

• 🟢 "Approve & Execute" — plan ready, requirements met
• 🟠 "Complete Requirements" — conditions unmet
• 🔵 "Executing..." — running with spinner
• 🟢 "Completed" — all done
• 🔴 "Failed" — error occurred

Structured Output

After each node executes, see rich structured output:

Each category is expandable — drill in without visual overload.

Output Modal

Click any completed node to see full output:

• Scrollable for long outputs
• Syntax highlighted code snippets
• Close with Escape or click outside

🏪 MCP Marketplace

Browse and attach MCP servers directly from the Overture UI.

When you attach an MCP server to a node, the agent gains access to those tools only for that step.

📂 Multi-Project Support

Work on multiple projects simultaneously:

Single project? Tab bar hides automatically for a cleaner UI.

📜 Plan History & Persistence

Never lose your work:

Resume Information

When resuming, you get complete context:

• Current node — where execution stopped
• Completed nodes — with their outputs
• Pending nodes — what's left to do
• Failed nodes — with error messages
• All configurations — field values, branches, attachments
• Timestamps — when created, when paused

✏️ Dynamic Plan Modification

Modify plans even during execution:

Plan Diff View

When a plan changes, see exactly what's different:

• Added nodes — highlighted green
• Removed nodes — highlighted red
• Modified nodes — yellow with before/after comparison
• Edge changes — added/removed connections

🔌 MCP Tools (For Agent Developers)

Overture exposes 11 MCP tools for agents to interact with:

🔄 Real-time WebSocket Communication

19 server-to-client message types:

connected • plan_started • node_added • edge_added • plan_ready • plan_approved • node_status_updated • plan_completed • plan_failed • plan_paused • plan_resumed • nodes_inserted • node_removed • project_registered • projects_list • history_entries • plan_loaded • resume_plan_info • plan_updated

16 client-to-server message types:

approve_plan • cancel_plan • rerun_request • pause_execution • resume_execution • insert_nodes • remove_node • register_project • subscribe_project • unsubscribe_project • get_history • load_plan • get_resume_info • save_plan • request_plan_update • create_new_plan

Relay Mode

When the WebSocket port is already in use, Overture automatically operates as a relay client, forwarding messages through the existing server. Multiple agent instances can share a single UI.

⚙️ Configuration

Setting Environment Variables

claude mcp add overture-mcp -e OVERTURE_HTTP_PORT=4000 -e OVERTURE_AUTO_OPEN=false -- npx overture-mcp

{
  "mcpServers": {
    "overture": {
      "command": "npx",
      "args": ["overture-mcp"],
      "env": {
        "OVERTURE_HTTP_PORT": "4000",
        "OVERTURE_WS_PORT": "4001",
        "OVERTURE_AUTO_OPEN": "false"
      }
    }
  }
}

{
  "servers": {
    "overture": {
      "command": "npx",
      "args": ["overture-mcp"],
      "env": {
        "OVERTURE_HTTP_PORT": "4000",
        "OVERTURE_WS_PORT": "4001",
        "OVERTURE_AUTO_OPEN": "false"
      }
    }
  }
}

⌨️ Keyboard Shortcuts

🤝 Supported Agents

Each agent has custom-tailored prompts for optimal plan generation.

💪 Why Overture?

For Users

• Transparency — See exactly what happens before code is written
• Control — Approve, reject, or modify any plan
• Context — Attach files and instructions to the right steps
• Choice — Compare approaches and pick your path
• Visibility — Real-time progress with rich output
• Safety — Pause, resume, or re-run at any time
• History — Resume any past plan instantly
• Efficiency — No wasted tokens on rejected approaches

For AI Coding

• Trust — Makes agents predictable and controllable
• Interpretability — See AI reasoning before execution
• Universal — Works with any MCP-compatible agent
• Extensible — MCP Marketplace for tool discovery
• Open Source — MIT licensed, community-driven
• Self-Contained — No cloud dependencies
• Works Offline — Fully local execution
• Multi-Project — Manage multiple workspaces

🧑‍💻 Development

# Clone the repo
git clone https://github.com/SixHq/Overture.git
cd Overture

# Install dependencies
npm install

# Build all packages
npm run build

# Start MCP server (in one terminal)
cd packages/mcp-server && npm start

# Start UI dev server (in another terminal)
cd packages/ui && npm run dev

Tech Stack

🤝 Contributing

Overture is open source and we welcome contributions!

• 🐛 Report bugs at GitHub Issues
• 💡 Suggest features at GitHub Discussions
• 📖 Improve docs — PRs welcome
• 🔧 Contribute code — see CONTRIBUTING.md

All contributions are appreciated, no matter how small.

📄 License

MIT License - see LICENSE for details.

Star History