Skip to content
VILIQ AI

Developers

A platform, not just an app

VILIQ AI is built so the orchestration layer is the product. Models change, providers change, SDKs change — the agent runtime, tool governance, routing and usage accounting are ours and stay put.

Interfaces

What the platform is being built to expose. None of these is callable yet — the honest status of each is below.

REST API

Workspace-scoped endpoints for agents, runs, flows and usage, with structured errors and request IDs.

Scoped API keys

A key belongs to exactly one workspace and carries scopes that can only narrow what its owner could already do.

MCP

VILIQ AI as a client of your Model Context Protocol servers, and as a server exposing a safe subset of itself.

Agent-to-agent

Interoperability with agent-to-agent standards, as an adapter around VILIQ AI runs rather than a replacement for them.

Webhooks

Signed delivery of run and approval events, with recorded attempts.

TypeScript SDK

A thin client over the same API, with nothing available to it that the API does not expose.

Honest status

This page describes a platform under construction. Here is exactly where it is.

  • Model gateway with multi-provider routingPrivate Beta
  • Agent runtime, budgets, approvals and tracesComing
  • Workspace model, roles and tenant isolationComing
  • Public REST API and scoped keysComing
  • MCP client and serverComing
  • Agent-to-agent interoperabilityComing
  • TypeScript SDK and webhooksComing

Nothing above is marked available until it is callable with a real key against a real workspace. If you are building against VILIQ AI and need a date, ask us and we will give you the real one.

How tools are governed

Anything an agent can reach is declared before it can be called.

  • Input is validated against a schema. Arguments a model generates are untrusted input, no different from a request body off the internet.
  • Every tool carries a risk class, so the approval rule is applied uniformly rather than each tool remembering to ask.
  • Network tools cannot reach inside. Private ranges, loopback and cloud metadata are refused, and redirects are not followed.
  • Output is bounded. An unbounded tool result is an unbounded bill and an injection surface.