Nifra

The AI-native TypeScript framework.
Typed APIs and full-stack SSR on five UI libraries, one app across Bun, Node, Deno, and the edge -
built so both humans and coding agents can change it safely.

Documentation · Playground · Benchmarks · vs. other frameworks

Most code is now written with an AI agent in the loop - and agents drift. They call an endpoint that moved, expect a response shape that changed, or hand-roll fetch with types that rot. Nifra removes that class of bug at the framework level: the client is inferred from the server's TypeScript (drift is a compile error), the docs are a live MCP server (agents read the real API, not stale memory), and a route-assurance gate fails the build when any route - human- or agent-written - ships without its required security evidence.

Quick start

bun create nifra my-app            # full-stack app: pick framework, runtime, DB, auth, CI

or start with just a typed API:

bun add @nifrajs/core @nifrajs/schema @nifrajs/client

The core loop

Routes are typed automatically from their path literals, handler context, and return values. Add a Standard Schema when you need runtime validation/coercion or an explicit request/response contract:

// server.ts
import { server } from "@nifrajs/core/server"
import { t } from "@nifrajs/schema"

export const app = server()
  .get("/users/:id", (c) => ({ id: c.params.id })) // params + response inferred from the route
  .post("/users", { body: t.object({ name: t.string() }) }, (c) => {
    // c.body is validated + typed - invalid input is a structured 422 before this runs
    return { id: crypto.randomUUID(), name: c.body.name }
  })
  .listen(3000)

// anywhere.ts - fully typed from the server, zero codegen
import { client } from "@nifrajs/client"
import type { app } from "./server"

const api = client<typeof app>("http://localhost:3000")

const res = await api.users({ id: "42" }).get()
if (res.ok) res.data.id     // typed from the route - tsc fails the moment the route changes
else res.error              // failures are returned, never thrown

Change a route and every caller stops compiling until it's updated. That one property is what keeps agent-edited codebases correct. For a decoupled, versionable surface, use defineContract + implement.

Agent-native, by construction

Register the MCP server and any coding agent reads your live routes, fetches version-checked examples, runs real requests against the app it just edited, and gates its own drift:

claude mcp add nifra -- bunx nifra mcp     # Claude Code (Cursor/VS Code: same command in mcp.json)
nifra init-agents                          # or: write .mcp.json + AGENTS.md + CLAUDE.md for you

The loop covers live project context and routes, verified docs/examples/types, checks with structured fixes, real requests and SSR renders, request inspection, tests, assurance, and verification levels. nifra_context and nifra_example are version-aware; nifra_run, nifra_render, and nifra_inspect verify what the edited app actually does; nifra_check and nifra_assure close the drift and security gates. Full tool list →

Agents that read skills get the conventions too - the same four skills on every surface:

pi install npm:@nifrajs/skills                                     # Pi
/plugin marketplace add nifrajs/nifra && /plugin install nifra@nifra  # Claude Code

Not in a Nifra repo? The docs tools are also hosted - add https://mcp.nifra.dev to Claude, Cursor, or ChatGPT and it learns Nifra from the same verified corpora, no checkout. One MCP, two transports (the same hosted-plus-local pairing Supabase, Stripe, and GitHub use): project tools run only on your machine over stdio - your code never reaches our servers.

Build and host agents

The same public contracts also cover applications that are agent products:

bun add @nifrajs/coding-agent @nifrajs/pi bunx nifra-agent --backend pi --message "run the checks and explain failures"
--verify-after-turn check --max-repair-attempts 2

Provider credentials, durable state, authorization, and approval policy are application ports rather than hidden framework state. The local process adapter contains crashes and accidents but is not a hostile-code sandbox; use OS-level isolation for untrusted code. A2A and AG-UI mounts likewise require the host application to add authentication and authorization at its route boundary.

Proof, not promises

Three CI gates turn security posture into build failures:

$ nifra assure
✖ POST /notes (authenticated-write) is missing nifra.authenticated

• nifra assure - a policy file classifies every route by reflection and fails CI naming exactly what evidence is missing: authentication on a write, a rate limit, CSRF, a body cap. No other framework ships this.
• nifra capabilities check - routes declare effect tokens ({ capabilities: ["db.write"] }); the check compares what a route says against what its module graph can actually reach, pinned in a lockfile. A GET that can reach a domain write is an error.
• nifra manifest diff - one hash-verified artifact of contracts + assurance + effects + response sensitivity; deploy promotion fails closed on breaking contracts, lost assurance, or newly exposed sensitive fields.

Security & hardening → · Effect provenance → · Verification ladder →

Full-stack, five UI libraries

The same routes, loaders, actions, streaming SSR, defer()/<Await> progressive rendering, islands, and typed data layer work on React, Vue, Solid, Svelte, or Preact - switching is one adapter import, not a rewrite. File routing, SSG/ISR, progressive-enhancement forms, query cache, and server functions whose bodies never ship to the browser.

bun create nifra my-app --framework svelte   # or react | vue | solid | preact

Frameworks → · Rendering → · Server functions →

StyleX and Tailwind migration

Nifra includes a conservative Tailwind → StyleX codemod for static JSX class lists:

nifra migrate --from tailwind --to stylex          # inspect the proposed changes
nifra migrate --from tailwind --to stylex --write  # apply safe changes

The codemod rewrites supported className="..." attributes to stylex.props(...) and a local stylex.create(...) table. It understands responsive breakpoints and element-local pseudo-classes. Dynamic class expressions, arbitrary values, parent-dependent variants, and unknown utilities are left untouched with file/line diagnostics for manual review. Use --dir <path> to scan a subdirectory.

StyleX compilation is built into Nifra's Bun pipeline. Install the runtime and optional compiler peers, then register both browser and SSR transforms in nifra.config.ts:

bun add @stylexjs/stylex
bun add -d @babel/core @stylexjs/babel-plugin @babel/plugin-syntax-flow \
  @babel/plugin-syntax-jsx @babel/plugin-syntax-typescript

import { stylexBunPlugin } from "@nifrajs/web/plugins/stylex"

export const clientPlugins = [stylexBunPlugin("dom")]
export const serverPlugins = [stylexBunPlugin("ssr")]