Atsign AI Architect MCP Server

What is Atsign AI Architect MCP Server?

Atsign AI Architect MCP Server connects AI coding assistants to Atsign AI Architect, a visual tool for designing secure-by-design applications and agentic systems, so the assistant can build, validate and export architecture blueprints.

AI Architect is made by Atsign. Teams use it to draw a blueprint of an application before code is written: its components, AI agents, identities, permissions, data flows and communication paths. The blueprint is then exported as structured context for an AI coding assistant, which generates code that uses the Atsign Platform, Atsign's service for encrypted, identity-based communication with no exposed inbound ports. Since June 2026, the MCP connection lets assistants such as Claude or Cursor draft a blueprint from a plain-language request and later read the approved version back for code generation.

How to use Atsign AI Architect MCP Server?

The server is a remote Streamable HTTP endpoint that accepts anonymous MCP clients, so no API key or header is needed to connect:

https://aiarchitect.atsign.com/mcp

To see blueprints on the visual canvas, open AI Architect in your browser (Atsign offers a free trial on its product page). The tools that read from or write to the canvas, request_current_blueprint and push_blueprint, need the stream session ID of your open browser session; the assistant will ask you for it.

Claude Code

claude mcp add --transport http atsign-ai-architect https://aiarchitect.atsign.com/mcp

Claude Desktop and claude.ai

Go to Settings > Connectors, choose Add custom connector, and enter https://aiarchitect.atsign.com/mcp as the URL.

Cursor

Add to ~/.cursor/mcp.json or the project's .cursor/mcp.json:

{
  "mcpServers": {
    "atsign-ai-architect": {
      "url": "https://aiarchitect.atsign.com/mcp"
    }
  }
}

VS Code

Add to .vscode/mcp.json:

{
  "servers": {
    "atsign-ai-architect": {
      "type": "http",
      "url": "https://aiarchitect.atsign.com/mcp"
    }
  }
}

Then ask for a design, for example: "Build a CRM platform with customer workflows, analytics pipelines and restricted-access AI agents" (the example prompt from Atsign's announcement).

Key features of Atsign AI Architect MCP Server

• Official MCP connection to Atsign AI Architect, announced by Atsign in June 2026.
• Scaffolding tools that produce schema-valid blueprint, node and edge JSON, so the assistant does not hand-write the format.
• Catalogs of allowed node types and connection types, including a scoping principle for what should be a node at all.
• Schema validation that returns errors and warnings before anything is pushed or exported.
• Two-way link with the browser canvas: pull the blueprint you have open, or push a validated one so the canvas updates immediately.
• Export for implementation that bundles the blueprint with the Atsign Platform SDK guide and node roles, intended for use in IDEs such as Cursor and Claude Code.

Atsign AI Architect MCP Server tools

The server exposes 22 tools, which fall into five groups:

• Catalogs and rules: list_node_types, list_connection_types and canonical_edge_style (for the connection types default, async, sync, notification, rpc, stream and tcp) tell the assistant what a blueprint may contain. get_implementation_rules returns the Atsign Platform rules for code generation once a blueprint validates.
• Scaffolding: scaffold_blueprint creates an empty blueprint with one page; scaffold_node and scaffold_edge build complete node and edge JSON with unique IDs; pick_edge_handles picks the sides an edge should leave and enter from.
• Layout: suggest_grid_position, suggest_horizontal_position, suggest_vertical_position and get_layout_constants calculate well-spaced positions.
• Editing: delete_node, delete_edge, update_node_position, update_node_label, update_node_notes and update_edge_notes fix or refine a blueprint, streaming the change to the UI when a session ID is given.
• Validation, canvas and export: validate_blueprint checks the JSON against the schema; request_current_blueprint asks permission to read the blueprint loaded in your browser; push_blueprint replaces the canvas with a validated blueprint; export_for_implementation returns the same output as the web app's "Export Prompt" button.

Use cases of Atsign AI Architect MCP Server

• Starting from a blank canvas: "Design a support system with a triage agent, a billing agent and a human escalation path, then push it to my canvas."
• Improving an existing design: "Load my current blueprint, find nodes without notes and add short descriptions."
• Fixing validation errors: have the assistant run validate_blueprint, then delete or update the offending nodes and edges.
• Going from design to code: in Cursor or Claude Code, call export_for_implementation on the approved blueprint and generate code that follows the Atsign Platform implementation rules.

Troubleshooting Atsign AI Architect MCP Server

• push_blueprint or request_current_blueprint fails. Both need an active stream session. Open AI Architect in your browser and give the assistant the stream session ID.
• The canvas doesn't change after edits. Edit tools stream updates to the UI only when a session ID is provided; otherwise they just return updated JSON.
• Validation errors after hand-edited JSON. Ask the assistant to rebuild elements with scaffold_node and scaffold_edge instead of writing JSON by hand, as the tool descriptions recommend.