MCP Protocol

MCP and Agents

How external agents drive and consume Agentable through the Model Context Protocol.

Agent tool surface

External agents interact with Agentable through a focused set of MCP tools. These tools expose the full lifecycle of panels while respecting human-in-the-loop boundaries.

list_panels

Discover available panels in the current workspace.

open_panel

Open an existing panel on the canvas.

fill_panel

Populate a panel with structured data.

compose_panel

Create new panels at runtime using validated specs.

patch_panel

Apply targeted updates to existing panels.

run_panel_action

Trigger actions defined within a panel.

describe_panel

Retrieve the full schema and current state of a panel.

Capability classes

Every tool belongs to one of four capability classes that define the scope of agent permissions.

read

View panel state and canvas content without modification.

ui

Open, close, and navigate panel interfaces.

mutate

Modify panel data through human-approved patches.

job

Execute long-running panel actions and workflows.

Multi-agent coordination

Agentable provides built-in primitives for safe multi-agent operation on a shared canvas.

Leases

Time-bound exclusive access to panels and canvas regions.

Attribution

Every change is attributed to the originating agent with full audit history.

Per-agent HITL queues

Separate human-in-the-loop approval queues for each agent to prevent conflicts.

Canvas-over-MCP

The entire workspace is exposed as an MCP resource. External clients connect under strict scope enforcement, allowing agents to read and act on the canvas while respecting all governance rules defined by the host.

The framework-native registry remains the source of truth. All panel schemas, capabilities, and permissions are validated server-side before any MCP operation is executed.

CopilotKit as optional transport

CopilotKit can be used as an optional transport layer for agent communication. However, the framework-native registry remains the single source of truth for all panel definitions and capabilities.

Example tool call

Agents interact with the system using standard MCP tool call format.

{
  "name": "compose_panel",
  "arguments": {
    "type": "career-profile",
    "position": { "x": 420, "y": 180 },
    "data": {
      "name": "Alex Rivera",
      "role": "Senior Product Designer"
    }
  }
}

Agents draw and see

Agents can draw and annotate directly on the canvas using read_canvas and screenshot_canvas operations. These high-level primitives allow agents to understand spatial layout and visual context without needing low-level pixel access.

All drawing operations flow through the same governed panel system, ensuring consistency with human-authored content.