Overview
Canvas Connectors gives agents one controlled way to discover business apps, connect user or organization accounts, read indexed workspace knowledge, and take app actions through MCP-compatible tools.
The mental model
Section titled “The mental model”Canvas Connectors has three surfaces. Knowing which one you’re on answers most “can I do this?” questions:
- Public MCP — external agents connect to the
/mcpendpoint and can search and read your indexed Sketch context. Read-oriented. This is what Claude Code, Cursor, and your own agents connect to, and what most of these docs cover. - Sketch agent runtime — Sketch’s own built-in agent. It runs the Claude Agent SDK with guarded built-in tools plus Sketch’s MCP tools (indexed search, file content, and more).
- Canvas live actions — the Sketch runtime can broker live actions in your connected apps (post to Slack, create a Linear issue) without exposing credentials to the agent.
Live app writes are not available through public MCP today. Public MCP is read-oriented; live actions run only through the Sketch runtime’s brokered Canvas path. See App integrations.
Where do I start?
Section titled “Where do I start?”| I want to… | Go to |
|---|---|
| Connect Claude Code, Cursor, the OpenAI SDK, or LangChain to search my indexed data | Connect over MCP |
| Create a workspace and run my first MCP call | Getting started |
| Self-host Sketch on my own infrastructure | Deployment |
| Have Sketch agents take live actions in apps | App integrations |
Public MCP vs the agent runtime
Section titled “Public MCP vs the agent runtime”The two access surfaces are separate. Most of these docs are about the first one.
- Public MCP access — the
/mcpendpoint exposed by Sketch today. Token-authenticated, and currently lists four public tools for search, entity context, and indexed file content. - Agent runtime access — the internal Sketch agent runtime. It runs the Claude Agent SDK with guarded built-in tools, the Sketch MCP tools, optional external MCP servers, and optional Canvas broker access for live app actions.
Public MCP (/mcp) |
Agent runtime (internal) | |
|---|---|---|
| Who connects | external agents / clients | Sketch’s own agent |
| Auth | PAT or OAuth bearer token | in-process |
| Tools | 4 sketch_* tools |
built-in tools + Sketch MCP tools |
| Writes | read-oriented | brokered Canvas live actions |
| Docs | Public MCP tools | Runtime built-in tools |
Why the tool counts differ
Section titled “Why the tool counts differ”The public endpoint and the runtime expose different toolsets, and the runtime’s built-ins are counted more than one way:
- The public
/mcpendpoint exposes foursketch_*tools — search, entity search, entity context, and indexed file content. - The runtime’s built-in tools are counted three ways: nine are guarded by
canUseTool(the permission allowlist), seven are enabled by default, and eight are grantable in the team UI. Runtime built-in tools breaks down each list. - The runtime also loads Sketch MCP tools (indexed search, file content, and more) on top of the built-ins.
The four public tools are documented in Public MCP tools.