Maestro MCP: plan from anywhere
Use any MCP-capable client (claude.ai, Claude Desktop, Cursor, ...) to plan work into Vibetight as yourself, then let agents execute.
Updated 2026-05-19
Vibetight is project management for coding agents, but the planning step is its own job. Maestro MCP lets you do that planning from any MCP-capable client, with your favorite assistant acting as you, and have the resulting tasks land straight on your Vibetight board.
What it actually is
Maestro is a standard MCP server, the same connector format claude.ai, Claude Desktop, and Cursor already speak. It exposes the operations a planner needs:
- Create tasks (with titles, descriptions, dependencies, priorities)
- Group tasks under an epic
- Set the agent assignment, status, and dependencies
- Move things around as the plan evolves
The MCP client authenticates as you. Tasks it creates show up in audit logs under your name. It’s exactly as if you’d typed the cards yourself, just delegated to a chattier interface.
Why this matters
Coding agents are now genuinely good at breaking work down. They’re less good at executing on huge sprawling projects in one shot. Maestro bridges that:
- You describe what you want in a long-form chat with your favorite MCP-capable assistant.
- It calls Maestro tools to lay the plan out as a structured set of tasks on your Vibetight board, with dependencies, assignees, the works.
- Your Vibetight agents (auto-pickup enabled) start executing the leaf nodes in the dependency graph.
- You review, steer, merge.
The planner and the executor are different roles. They don’t have to be the same agent, the same provider, or even the same machine.
Auth
Standard one-click OAuth. You paste the server URL into your client, a browser window opens, you approve once, and the client is connected. You never copy-paste a credential.
The endpoint URL is https://api.vibetight.com/api/mcp/maestro/v1.
Open Settings → Connected apps in Vibetight. The page hosts a
“Copy URL” button and per-client setup guides for claude.ai, Claude
Desktop, and Cursor.
For clients that don’t speak OAuth-MCP (the claude CLI is the main
one), mint a Personal Access Token in Settings → Connected apps and
paste it as the bearer credential. Vibetight stores only the hash, so
copy the token at create time, or revoke and mint a new one.
What it can do
Roughly the things you’d do in the web UI when you’re planning, exposed as MCP capabilities your assistant can call:
- Create tasks and group them under epics, with titles, descriptions, priorities, dependency edges, and parent epics.
- Plan a batch of work in one shot, laying out several tasks with the dependency graph already wired up.
- Move tasks around as the plan evolves: change status, assignee, dependencies, position within a lane.
- Delete or reorder mis-created cards as the plan firms up.
- Read the current state of a project, its agents, and its tasks so the planner has context before adding more.
- Schedule work onto an agent by posting
@agent ...into a task’s chat, the same dispatch path the web UI uses. Slash commands (/start,/interrupt, …) drive the agent lifecycle without leaving the conversation.
Tasks created via MCP get the same defaults as web-UI tasks: agent assignment, dependency-aware scheduling, auto-pickup if the agent has it enabled.
Mental model that’s load-bearing
The MCP client acts as the user, not as an agent. It plans. It does not execute code on your repo.
Your Vibetight agents (the ones that show up on your board as assignees) are the executors. They pick up tasks and do the work.
The split matters because: (a) you can plan from anywhere, with any client, without paying compute on a planner that doesn’t need to execute; (b) the audit trail correctly shows you as the person who scheduled the work; (c) you can have one provider plan and another execute.
Use cases
- Triage a backlog at 7am. Open claude.ai, paste your incident notes, have it plan a sprint into Vibetight. Walk into the office, your agents have already picked up the unblocked items.
- From a Cursor side-chat. While you’re coding, ask Cursor to file the follow-ups you keep noticing as you go. They land in your todo lane for an agent to pick up tonight.
- Replan from your phone. Mid-meeting, have a quick conversation with your assistant that re-prioritizes the board. No laptop required.