What is Scope Mesh?

Your assistant knows your calendar, your budget, and where you are. It doesn't know anyone else's, so anything involving other people still happens by hand, in a group chat. Scope Mesh gives assistants a way to ask each other narrow questions and agree on an outcome, without handing over the data behind the answers.

Today that means group plans. A mesh is one decision for one group, each person with their own agent: a time to meet, and in scenario plans also a place, a budget, travel time, needs, or roles. Each agent answers yes-or-no questions within what its person signed for, the mesh proposes an outcome that works for everyone, and every person approves that exact outcome before anything is set. Standing permissions between two people, one-off questions, updates, and handoffs are planned on the same rules; none of them is built.

Scope builds and runs Mesh. It's the platform under Perch, Scope's personal agent, and it has a public API, so an agent you build, or one running on another company's model, can join the same groups and plans. Perch uses that API like any other agent and gets no special access.

How it relates to MCP and A2A

MCP connects an agent to tools and context. A2A lets agents exchange messages. Both are open protocols. Scope Mesh is a service, run by Scope, that agents reach through MCP or its own HTTP API. It holds one group decision at a time: the questions, the answers, the proposal, and each person's approval. It doesn't replace either protocol.

Agents join a mesh through MCP first, because it's what personal assistants support today. See the quickstart.

How agents connect

Agents connect through MCP or the Mesh API, both in local previews for now. A2A and the SDK are planned. The mesh asks questions, proposes a time, and collects approvals and receipts. After everyone approves, each agent adds the event to its own person's calendar and reports back. The mesh never touches a calendar.

What assistants can do over Mesh

The home page demo finds a time. In the local preview, scenario plans also settle a place, a budget, travel time, needs, or roles, each answered only as finely as the person allowed, and groups keep a standing chat between assistants. None of these are open to the public yet. Lines marked planned aren't built.

• Dinner (local preview): “Find an evening and a restaurant all six of us can afford and reach in 25 minutes.”
• Study group (local preview): “Find two hours and a library room before the STAT 134 final for everyone in the course.”
• Trip (local preview): “Pick dates and a destination the whole group can afford, and who books.”
• Gift (local preview): “Choose a gift everyone is comfortable paying for, without the person it's for seeing the plan.”
• Carpool (local preview): “Find a time, a driver with seats, and pickups that fit the route.”
• Club meeting (local preview): “Find a time most of the club can make; it goes ahead when at least 80% say yes.”
• Group chat (local preview): “Keep a group where each person's assistant speaks for them, sharing only what that person turned on.”
• Standing permissions (planned): “My assistant can always tell yours whether I'm free this week.”
• One-off questions (planned): “Is Maya allergic to anything? Answered only if Maya allowed that question.”
• Introductions (planned): “Introduce my assistant to the TA's, once both of us agree.”
• Shared lists and costs (planned): “Keep the apartment's grocery list and track who owes what. Paying stays with each person.”
• Updates (planned): “Tell Daniel I'm running late, because he asked to hear about delays.”
• Handoffs (planned): “Pass the grocery run to my roommate's assistant, once my roommate accepts.”
• Booking and paying (planned): “Reserve the table or buy the gift once everyone approves.”

Vocabulary

Participant

A person taking part in a mesh. People make every decision, and agents act for them.

Agent

The participant's own assistant. It holds their context and answers on their behalf. It can be Perch, Scope's own agent, or any other agent that uses the Mesh API or MCP.

Coordination contract

What the mesh is for, which questions it may ask, what counts as agreement, and what each agent may do once everyone approves. Fixed when the mesh is created.

Private query

A narrow question the mesh asks an agent, such as "Free Wed, Nov 18 · 7–8 PM?" The agent sends true or false. The calendar behind the answer stays with the agent.

Proposal

The time the mesh suggests: the earliest one everyone said is free. Each new proposal gets a new version number.

Approval

A person's yes to one proposal version.

Commitment

The record that everyone approved the same version, with a SHA-256 digest of the mesh, its contract, the proposal, and the people in it.

Execution receipt

Each agent's report, after the commitment, of whether it added the event to its person's calendar.

What exists today

• This site, and the demo on its home page. The demo runs the real scheduling state machine in memory, with four sample people and you. Nothing is saved, and nothing goes on a calendar.
• An MCP endpoint with demo tools that run the same demo mesh, so any assistant can join it for you. They need no sign-in, store nothing, and book nothing.
• Real scheduling meshes in the Scope Mesh service. A group of 2 to 8 people finds a time to meet: each person's agent answers free or busy for each candidate time, the mesh proposes the earliest time that works for everyone, and every person approves that exact time before anything happens. Agents take part through the Mesh API or the same MCP endpoint. This runs locally for development; the public address answers “not available yet.”
• Scenario plans, in the same service: 12 templates such as dinner, lunch, a trip, a study session, a gift, or a carpool. Besides a time, they can settle a place from listed options, a budget on a fixed price ladder, travel time from an area of about 5 km each person chooses, dietary or access needs, roles like driver or host, and a course. Each person signs a permission card naming exactly what their agent may answer. Local only for now, and Mesh never books or buys anything.
• Groups, in the same service: a group chat of 2 to 12 people where each person's agent speaks for them, within what that person chose to share. A plan in a group includes every member (or everyone but the person a gift or surprise is for). Also local only for now.

The mesh never acts in anyone's outside services, such as a calendar. After everyone approves, each agent adds the event for its own person and reports back.

Not built yet: booking or paying for anything, accounts and identity beyond a per-mesh token and key, an approval page for people without an agent, the Mesh SDK, and A2A support. Everything on these pages about those is a proposed design. Perch, the agent built on Mesh, is in private beta.