n.codes - Changelog

February 15, 2026

Improvement

Feasibility Check Before Codegen

LLM-guided feasibility checks to avoid dead-end generations.

Added an LLM-powered feasibility step between intent parsing and code generation. It uses lightweight keyword extraction to focus on the most relevant capabilities before deciding whether the request can be fulfilled.

When requested data or actions are not supported, the pipeline now returns a clarifying question with suggested alternatives instead of producing broken UI.

February 13, 2026

Major

Agentic UI Pipeline

From fixed components to fully generated interfaces.

Replaced the DSL-based component renderer with a full agentic code generation pipeline. The widget now generates complete HTML/CSS/JS from scratch for every prompt — no pre-defined templates, no fixed component types.

The pipeline follows an agentic loop: intent parsing, code generation, QA review, and iteration. Generated UI renders inside a sandboxed iframe with a postMessage bridge for secure data fetching against the host app's APIs.

Components are now wired to real API endpoints via capability map resolution. Query bindings fetch live data; form action bindings submit to real endpoints. The server validates refs against the capability map before returning resolved endpoints.

</>

Code Generation

Full HTML/CSS/JS generated per prompt, not mapped to fixed components.

↻

QA Review Loop

A review agent evaluates generated views and feeds back for iteration.

⚡

Live Data Bindings

Queries and actions resolve to real endpoints through the capability map.

🛡

Sandboxed Rendering

Generated UI runs in an iframe sandbox with a postMessage bridge.

What changed under the hood

Server

New agentic pipeline with intent parsing, codegen prompts, and review prompts. Capability resolver validates refs and resolves endpoints. Proxy layer for cross-origin data fetching.

Widget

Removed 10 fixed component renderers. Added iframe sandbox, API bridge, and message handler. The renderer now injects generated code directly into a sandboxed frame.

Testing

400+ new tests across server, widget, and shared modules covering the pipeline, capability resolver, sandbox, code parser, and message handler.

73 files changed — 7,100 lines added, 2,200 removed. This is the architectural shift from template instantiation to true agentic UI generation.

February 1, 2026

New

Introducing Capability Map

It is the guardrails layer for n.codes.

It auto-maps what your app can do (APIs, schemas, components) and tells the LLM exactly what is allowed to generate in a production setting.

Capability Map is the foundation of that, enabling agents to securely generate new features in production, leveraging the capabilities of the existing backend and components.

Allowing users to generate their own custom features.

E

Entities

Data models surfaced as capabilities.

A

Actions

Mutations and commands the agent can use.

Q

Queries

Approved data access patterns for answers.

C

Components

UI building blocks already in your app.

How is it different from MCP or API docs?

MCP

MCP is a protocol to access tools and resources. It does not define app-specific boundaries or UI affordances.

API docs

API docs describe endpoints. They do not capture permissions, product constraints, or frontend components. They are written for humans.

Capability map

Capability map merges APIs, schemas, UI components, constraints, and RBAC into a single "allowed actions" set so users can generate features safely within guardrails.

Capability Map keeps agents inside the lines while unlocking user-generated features in production.