Overture MCP
An MCP server that draws the agent's plan as an interactive flowchart in your browser and holds the agent until you approve it
- 642 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The server starts a local web interface and WebSocket on ports 3031 and 3030
- Plan history in `~/.overture/history.json` keeps all field values, and the field types include secret ones for API keys
- Files and instructions attached to nodes enter the agent's context before the step runs
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add mcp/overtureDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The server starts a local web interface and WebSocket on ports 3031 and 3030
- Plan history in `~/.overture/history.json` keeps all field values, and the field types include secret ones for API keys
- Files and instructions attached to nodes enter the agent's context before the step runs
This is third-party code. Review the repository files before installing.
What it does
Overture takes the planning phase away from an AI agent and shows it as a flowchart on a local page. The agent submits the plan through the submit_plan tool, steps appear on the canvas as they are generated, and the get_approval tool keeps the agent waiting until you approve. On the page, steps can be opened and extended. Each node shows its complexity, expected output, risks and edge cases. You attach files to a node, write your own instructions ("use the auth pattern from src/auth.ts"), and where the agent proposes several routes you compare the options with their pros and cons and pick one. Nodes can ask for values: a string, a number, a toggle, a select, a file or a secret field for an API key. The approve button stays disabled until the required fields are filled and the branches are chosen. Once approved, nodes light up as they run, and each step's result is collected into structure: changed and created files, installed packages, searches, tool calls. The space bar pauses the run, a failed node or everything from a chosen step can be re-run, and steps can be inserted or removed mid-run. Plans are saved to `~/.overture/history.json`, and any past plan can be reopened and resumed.
Who it is for. For developers working with Claude Code, Cursor, Cline or GitHub Copilot who want to review the agent's plan before it spends tokens on code.
Good fit when
- A large task such as "build a REST API with auth", and you want to see the steps and their dependencies before the first line of code
- The agent proposes several approaches and you want the pros and cons side by side so you can choose
- A step needs an API key, a file or a project name, and a node's form is easier than typing it into chat
- One node failed and you want to re-run it, or the whole chain from there, without starting over
- You work on several projects at once: Overture keeps a tab for each and shows where updates arrived
Not a fit when
- A one-line edit in a single file: a plan on a flowchart takes longer than the change
- Your agent cannot connect MCP servers: Overture works only through MCP
- You work on a remote machine without a browser: the interface opens on local port 3031
- A text plan in chat is enough for you: the flowchart pays off with many dependencies, branches or input values
Example request
Build a REST API with auth. Submit the plan to Overture first and wait for my approval before you start writing codeMCP
- Transport
- stdio
- Authentication
- not required
Environment variables
OVERTURE_HTTP_PORT- Web interface port, 3031 by default
OVERTURE_WS_PORT- WebSocket port, 3030 by default
OVERTURE_AUTO_OPEN- Open the browser on start, true by default
Limitations
Node.js is required; the server starts with `npx overture-mcp`. The web interface opens on port 3031 by default, the WebSocket runs on 3030, and the browser opens by itself on start. The ports and auto-open are set with OVERTURE_HTTP_PORT, OVERTURE_WS_PORT and OVERTURE_AUTO_OPEN.
GitHub Copilot needs VS Code 1.99 or newer, and its configuration lives in `.vscode/mcp.json` with a servers key instead of mcpServers.
The package holds a separate set of instructions for each agent, and the flowchart appears when the agent follows them. The package is at version 0.1.8.
How to disable. In Claude Code, run `claude mcp remove overture-mcp`. In Cursor, Cline and Sixth AI, delete the overture block from the MCP settings file; in Copilot, from `.vscode/mcp.json`. If you installed the package globally, remove it with `npm uninstall -g overture-mcp`. Plan history stays in `~/.overture/history.json` and can be deleted by hand.
FAQ
How do I install Overture in Claude Code?
Run `claude mcp add overture-mcp -- npx overture-mcp`. Then give the agent any task: Overture opens `http://localhost:3031` with a plan waiting for approval. Environment variables go in with the -e flag, for example `-e OVERTURE_HTTP_PORT=4000`.
How do I connect Overture to Cursor or Cline?
Add an overture block with command npx and args overture-mcp to `~/.cursor/mcp.json`. For Cline, the same block goes into the MCP settings in VS Code. GitHub Copilot uses `.vscode/mcp.json` with a servers key, and Sixth AI reads its own sixth-mcp-settings.json.
Which MCP tools does Overture give the agent?
Eleven: submit_plan, get_approval, update_node_status, plan_completed, plan_failed, check_rerun, check_pause, get_resume_info, request_plan_update, create_new_plan and get_usage_instructions. The first three carry the main loop: the agent submits the plan, waits for approval and reports step status.
What happens if the WebSocket port is already taken?
Overture switches to relay mode and forwards messages through the server that is already running. This way several agent instances share one interface.
Can the plan change after execution starts?
Yes. With the request_plan_update tool the agent inserts, removes and replaces nodes, and the page shows the difference by colour: green for new nodes, red for removed ones, yellow for changed ones with a before and after comparison.
Discussion
No questions or reviews yet.
Related
MCP serversA skill that keeps the task plan in files on disk and re-injects it every turn so work survives context clears, compaction and crashes
A planner for Claude Code, Codex and OpenCode: one goal becomes a task plan and independent tasks run in parallel
MCP server of the self-hosted Synaplan AI platform: semantic search over your documents, long-term memory and chat
Agents, skills and templates for structured AI development in OpenCode, Codex and Claude Code: docs, plans, review, checkpoints
A CLI that runs parallel Claude Code sessions from your plan, each in its own git worktree, with stage checks and a knowledge base
A prompt template store with search and variable filling, plus a script that runs Claude agents across a whole project.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| Overture (this card) | no | not required | 642 | |
| Plannotator | no | 10,168 | annotates the agent's finished plan and diffs in the browser and sends the notes back | |
| MCP Shrimp Task Manager | no | not required | 2,151 | splits a task into steps with dependencies and keeps context between sessions, without a visual canvas |
README in short
The README opens with the problem: agents such as Cursor and Claude Code start writing code right away, and a text plan in chat does not show dependencies, branches or complexity. The fix is described as a visual layer between the request and the code, followed by install steps for five agents, a walkthrough from request to execution control, and detailed feature tables: canvas, node panel, dynamic fields, attachments, branches, history, mid-run plan changes. It ends with the MCP tools, WebSocket messages, environment variables, shortcuts and the tech stack.