Connect an agent

Pair MCP, experimental WebMCP, or a host-supplied model with one authorized application session and Region.

Add an agent only when your product wants language-driven UI. Varendo itself stays deterministic: your code can send the same intent with no model.

Pick a way to connect

Browser agent

Enable experimental WebMCP in the public Playground. A browser agent calls the page's tools; manual controls stay available.

Try WebMCP →

MCP client

Run the standalone local example and connect an MCP client through HTTP or stdio.

MCP setup →

Your model

Run a bounded tool loop in your Node host. Configure the endpoint, protocol, model, and authentication; keys stay on the server.

BYOK setup →

The runtime does not depend on a model provider. MCP and WebMCP connect agents through tools; ToolModelPort connects a host-selected model protocol. The supplied Chat Completions and Responses adapters require those wire formats. A different protocol can implement the same port without changing your data or rendering code.

Know the three tools

Tool	Purpose	Cannot do
`aeliqo_context`	Discover resources, fields, meanings, views, and actions visible to this session.	Read undisclosed rows or acquire grants.
`aeliqo_render`	Submit an intent through validation, evaluation, recipe selection, and Region commit.	Send HTML, JavaScript, SQL, or module URLs.
`aeliqo_act`	Preview or request execution of a registered business action.	Confirm itself or bypass server authorization.

Read a tool call

aeliqo_render accepts the same intent envelope your code sends. This call asks the paired Region to browse permitted People rows:

aeliqo_render input

JSON example

{
  "version": "1",
  "id": "browse-engineering",
  "kind": "browse",
  "resource": "people",
  "fields": ["name", "team"],
  "filter": { "op": "compare", "field": "team", "comparison": "eq", "value": "Engineering" }
}

The tool returns a trusted receipt. renderer-ready means the browser acknowledged the commit. denied means current authority rejected the request.

Trust the boundary

The agent proposes. The runtime validates. The Region commits. Your agent never emits executable UI, and you confirm consequential actions yourself.

Discovery can show that a requested concept is unavailable. Answer truthfully instead of forcing unrelated data on screen. The local runner labels model-only prose as a draft and leaves the Region unchanged.

Pick a way to connect#