MCP server

Assign runs a remote Model Context Protocol server. With it, an AI assistant can read and change your Projects, Tasks, Documents and Knowledge. It works under the same permissions you have in the web app.

https://mcp.assign.so/

Add that URL as a remote MCP server in your client. The client opens Assign in a browser, where you choose a Workspace and approve read access, and write access when you need it. You don't need a client ID, client secret or API key.

What an assistant can do

The catalog doesn't expose any of the following: deleting a Workspace, Project or Document, permanently purging a Task, billing, member administration, credential management, arbitrary HTTP, SQL, filesystem access or shell access.

Break down larger work Upcoming skill update

With the updated Assign skills loaded, ask your assistant to break a roadmap or large Task into work you can resume later. It reviews existing work, proposes a compact structure and saves an approved batch through the available MCP tools.

• Milestones group Tasks around a meaningful delivery checkpoint. The assistant shows the proposed scope and asks for your confirmation before creating one.
• Tasks describe independently assignable, verifiable outcomes.
• Subtasks track parts of one outcome that need their own owner, Status or blocker.
• Checkboxes capture steps or acceptance checks that share the Task's ownership and lifecycle.

For a large roadmap, the assistant outlines the whole scope and details ready work first. Later scope and the next batch stay in a linked plan so another client can continue. You can also request a full breakdown, saved in manageable batches. Each Task keeps a short outcome, essential constraints and completion checks; detailed requirements stay linked.

For example: “Break this reporting roadmap into manageable work. Reuse existing Tasks, propose Milestones for my confirmation, and show the first batch before creating it.”

Skills guide the assistant; connecting MCP alone does not install them. Tool availability and permissions still apply. Saved plans and created Tasks are not completed work.

How access works

• The scope is a ceiling. assign:read and assign:write cap what a connection can do. Every call still checks your current Workspace membership, role, Project visibility, Workspace AI policy and the resource's state.
• Changes are safe to retry. Every write takes an idempotency key. Task and Document updates also take the current revision, so a retry can't overwrite a newer change made by someone else.
• Revoking access works immediately. Disconnecting a client, disabling AI access for the Workspace or losing membership blocks the next call, including tools the client has cached.
• Your Assign credentials stay in Assign. A client gets its own revocable MCP credential. The Assign CLI, the web session and service credentials are never shared with it.

Supported clients

Assign identifies clients through their published Client ID Metadata Document (CIMD). It does not offer open dynamic client registration, so a client needs to support remote servers with OAuth and CIMD.

Codex is the client with a documented, verified setup path. Instructions for other clients will be added here once each one has been tested against Assign.

The next deployment enables a compact single-Task preview for task_get. Task lists and clients without interactive previews use structured results, text and links.