No MCP installation required

DiffBeam Direct Agent Bridge

An AI agent that can send HTTPS POST requests can read and preview a user's live DiffBeam workspace without installing a native MCP server. The user creates the workspace and temporary permission inside the planner, then copies one generated prompt into the AI chat.

One endpoint

The prompt contains a workspace-specific endpoint of the form https://api.gaussian-beam.com/api/agent/workspaces/{public_workspace_id}. Every command is JSON.

Temporary authority

The generated workspace_credential is capability-bound, expires, and stays in the POST body. It must never be placed in a URL, public page, model configuration, or source repository.

Live preview first

preview_workspace_design validates and solves a candidate, then streams the calculated preview into the open planner. Applying a revision remains a separate authorized action.

When an AI agent should use this bridge

Use the Direct Agent Bridge when the user has clicked Connect your AI in an existing DiffBeam planner workspace and pasted the generated connection prompt into the conversation. The bridge is the fastest default for general internet-capable agents because it needs no MCP installation. Use native MCP only when the user explicitly prefers an MCP client or when the client already supports remote MCP tools.

A passive chatbot that cannot make authenticated HTTPS requests cannot use the bridge. It can still explain the design or send the user to the DiffBeam planner, but it must not claim to have published a preview.

Command contract

All commands use Content-Type: application/json. The public workspace ID is part of the endpoint; the temporary credential is always part of the request body. Placeholder values below are documentation only and grant no access.

{"operation":"get_workspace","workspace_credential":"<temporary credential from the user>"}
{"operation":"preview_workspace_design","workspace_credential":"<temporary credential>","spec":{"wavelength":{"value":689,"unit":"nm","fixed":true},"optical_chain":[]}}
{"operation":"apply_workspace_design","workspace_credential":"<temporary credential>","expected_revision":3,"spec":{"wavelength":{"value":689,"unit":"nm","fixed":true},"optical_chain":[]}}

The bridge also accepts list_workspace_revisions and restore_workspace_revision when the issued capability contains those scopes. Unknown fields and malformed commands are rejected. Responses are marked no-store, and stale writes return a revision conflict instead of overwriting somebody else's work.

Fast preview workflow

  1. Call get_workspace once and start from the returned snapshot.
  2. Build the simplest physically valid candidate that satisfies the stated constraints.
  3. Call preview_workspace_design immediately. This call already performs validation and deterministic solving.
  4. Tell the user in one sentence what appeared in the planner. Refine and preview again only when useful.
  5. Do not call apply_workspace_design unless the copied prompt grants apply authority and the user has approved the change.

For native tool discovery, use the DiffBeam MCP server. For exact request and response surfaces, see the contracts page. DiffBeam does not host or proxy a language model; it supplies deterministic Gaussian beam calculations and capability-bound workspace operations.