nifra

The full-stack TypeScript framework built for AI agents - and for the humans who work alongside them.

Coding agents drift. They call an endpoint that moved, expect a response shape that changed, or hand-roll fetch with ad-hoc types that fall out of sync the moment a route changes. nifra removes that class of bug at the framework level:

The rest is a fast, contract-first full-stack TypeScript stack: routing, validated I/O, SSR, loaders/actions, auth, WebSockets, MDX, and multi-runtime deployment.

bun create nifra my-app

Use nifra from your AI assistant

nifra's own docs, runnable examples, and API types are a live remote MCP server - listed in the official MCP registry - so Claude, Cursor, Codex, and any MCP client learn nifra from the source instead of guessing from stale training data:

• Claude Code: claude mcp add --transport http nifra-docs https://mcp.nifra.dev
• Claude.ai / Desktop: Settings → Connectors → Add custom connector → https://mcp.nifra.dev
• Cursor / VS Code / other MCP clients: point them at https://mcp.nifra.dev

Inside a project, nifra mcp additionally serves your app's live routes and schemas to the agent - so it writes against the code you have, not the code it remembers.

The backend

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

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

export type App = typeof app

The typed client - the anti-drift seam

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

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

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

The client never throws - every call returns { ok, status, data, error }, so the happy path and the failure path are both in the types.

Agent tooling

nifra ships a purpose-built toolchain so coding agents stay correct as the codebase evolves.

AGENTS.md - generated per scaffold, teaches the agent nifra's non-obvious rules:

• validate every input at the boundary with t or any Standard Schema
• always call this app's own API through client<typeof app> - never hand-roll fetch
• never top-level-import server-only code into a route module

Adding nifra to an existing app? Run nifra init-agents. It writes the agent-discovery files for you - .mcp.json + .cursor/mcp.json (registering this project's nifra MCP), a CLAUDE.md MCP-first preamble, and an AGENTS.md section - no-clobber, so it never overwrites a file you've customized. (nifra check also nudges you when a project has no .mcp.json.)

nifra init-agents          # wire .mcp.json + .cursor/mcp.json + CLAUDE.md into an existing app (no-clobber)

Or connect the MCP server by hand so the agent reads your live routes, verifies endpoints, and gates drift from inside its tool loop. Run once from your project root:

# Claude Code
claude mcp add nifra -- bunx nifra mcp

# Cursor / Claude Desktop - add to .mcp.json (or claude_desktop_config.json):
# { "mcpServers": { "nifra": { "command": "bunx", "args": ["nifra", "mcp"] } } }

Once connected, the agent has fifteen tools - no setup per prompt:

No MCP? The same data is available as plain commands - paste into any prompt, or run in CI:

nifra context          # routes + schemas (+ per-route call signatures) as Markdown
nifra check            # typecheck + typed-client drift lint; --json for agents, --lints-only to skip tsc
nifra assure           # policy gate for route auth/CSRF/rate/body/idempotency evidence; --json for CI
nifra capabilities check # effect provenance + capability lockfile gate; --json for CI
nifra manifest emit    # deterministic contract + assurance + effects + classification artifact
nifra manifest diff old.json new.json # deploy-promotion breaking-change gate
nifra doctor           # undeclared imports + duplicate identity-sensitive installs
nifra sync-manifest    # regenerate a web server-manifest.ts from routes/ without a full build

Learn nifra from any assistant. The docs, example, and type tools are also hosted, project-independent, at mcp.nifra.dev - add that one URL to Claude, Cursor, VS Code, or ChatGPT and it learns nifra from the same verified corpora, no checkout. Read-only, no key.

claude mcp add --transport http nifra-docs https://mcp.nifra.dev
# Cursor / VS Code: add { "url": "https://mcp.nifra.dev" } to .cursor/mcp.json or .vscode/mcp.json
# Claude.ai / ChatGPT: Settings -> Connectors -> add the URL

See Coding agents for per-client setup.

Upgrading from 1.x? Run nifra upgrade 2.0.0 as a dry-run, then follow the Nifra 2.0 migration guide.

Install

bun add @nifrajs/core            # the lean server + router
bun add @nifrajs/client          # the typed client (browser-safe)
bun add @nifrajs/schema          # the `t` schema builder + OpenAPI (optional)
bun add @nifrajs/middleware      # CORS, security headers, rate limiting (optional)

nifra is ESM-only and Bun-native (it uses Bun.serve). It runs on Bun; the client is environment-agnostic.

Use @nifrajs/core (or @nifrajs/core/server) for the ordinary HTTP runtime. Nifra keeps the package root deliberately lean and splits everything else across documented subpaths - most apps only ever touch a handful, so start with those and reach for the rest when a concept actually comes up:

• Everyday - @nifrajs/core/server (the runtime), .../contract (defineContract + implement), .../router, .../cookies, plus @nifrajs/schema (the t builder) and @nifrajs/client (the typed client). This is the 80% API.
• Advanced, opt in when you need it - .../assurance, .../capabilities, .../idempotency, .../effect-ledger, .../durable-execution, .../causality, .../classification, .../manifest, .../reflection, .../diff, .../mcp, .../sse, .../webhook, .../budget, .../seo, .../mount. Each is a separate documented subpath, so you never pay (in bundle size or in concepts to learn) for one you don't import.

Validate input with t (and get OpenAPI for free)

@nifrajs/schema's t is a TypeBox-backed builder: it validates at the request boundary and - because a TypeBox schema is a JSON Schema - generates OpenAPI with no extra work. Bring your own Standard Schema (zod, valibot, arktype) too; they validate identically.

import { server } from "@nifrajs/core/server"
import { t, toOpenAPI } from "@nifrajs/schema"

const app = server().post("/users", { body: t.object({ name: t.string() }) }, (c) => ({
  id: "u1",
  name: c.body.name, // typed as string, validated at runtime
}))

const openapi = toOpenAPI(app) // OpenAPI 3.1 document

Invalid bodies are rejected with a structured 422 before your handler runs.

Graduate to a contract - handlers unchanged

When you want a decoupled, versionable API surface, lift the same routes into a contract. Handlers written inline lift over unchanged.

import { defineContract, implement } from "@nifrajs/core/contract"
import { t } from "@nifrajs/schema"

const contract = defineContract({
  getUser:    { method: "GET",  path: "/users/:id", response: t.object({ id: t.string(), name: t.string() }) },
  createUser: { method: "POST", path: "/users",     body: t.object({ name: t.string() }), response: t.object({ id: t.string(), name: t.string() }) },
})

const app = implement(contract, {
  getUser:    (c) => ({ id: c.params.id, name: "ada" }),
  createUser: (c) => ({ id: "new", name: c.body.name }),
})

The client can now be built from the contract alone (client(contract, url)) - no dependency on the server's source. This is the shape agents reference: nifra context emits the live contract; nifra check enforces it.

Harden it

import { server } from "@nifrajs/core/server"
import { cors, securityHeaders, rateLimit, MemoryStore } from "@nifrajs/middleware"

const app = server()
  .use(securityHeaders())
  .use(cors({ origin: ["https://app.example.com"], credentials: true }))
  .use(rateLimit({
    store: new MemoryStore(),
    max: 100,
    windowMs: 60_000,
    key: (req) => req.headers.get("x-user-id") ?? "anonymous",
  }))
  .get("/", () => ({ ok: true }))

// Graceful shutdown, request timeout, body-size cap, redacting logger are built in:
server({ requestTimeoutMs: 5_000, gracefulSignals: true })

Official hardening modules also publish route evidence. Add a nifra.assurance.ts policy and run nifra assure in CI to fail when a new route is unclassified or misses required authentication, CSRF, rate-limit, body-limit, idempotency, IP, or security-header enforcement. The proof is built from route reflection, so it adds no request-path work. See Security & hardening.

Routes may also declare effect tokens ({ capabilities: ["db.read"] }). Add capability definitions and approved/forbidden import provenance to the same nifra.assurance.ts: nifra check then blocks raw effect imports and declaration/evidence drift, while nifra capabilities check also compares the deterministic capabilities.lock.json. GET/HEAD domain writes fail unconditionally; mutating effects must carry the request-idempotency or durable-command evidence required by their definition. Disabled apps retain the existing request hot path; enabled routes pay only when they call useCapability. For owned effects, prefer executeCapability(c, id, metadata, run): it assigns an effectId, records intent plus exactly one automatic terminal outcome, forwards c.signal, and supports token-only async aroundCapability() admission policies with fail-closed denial, timeout, and abort behavior. The original synchronous beacon remains available for adapter hot paths. For long-running or crash-sensitive effects, @nifrajs/core/durable-execution adds a signed, single-use approval coordinator (tenant/principal/operation bound), a durable effect journal, reconciliation reports, and a typed saga engine with reverse compensation and persisted retry state. Crash-ambiguous executions and compensations stop in manual review; an operator can apply a provider-confirmed outcome with resolveAmbiguity() (bound to the exact stored effect ID), then call resume() or compensate() without replaying an unknown effect. These require an explicitly durable store in production; the saga store owns encrypted business input and compensation arguments, while the sealed ledger and @nifrajs/otel/effects remain token-only. An approved provenance import is the explicit trust boundary: its provider internals are not scanned, while every unapproved local wrapper remains transitively scanned for raw-effect bypasses.

For deployment promotion, nifra manifest emit combines those schemas and proofs with field-level response sensitivity (classified(schema, "pii")) in one deterministic, hash-verified artifact. Operator code may sign it with Ed25519 through a KMS/HSM callback; nifra manifest diff fails closed on breaking contracts, lost assurance, expanded effects, or increased response sensitivity.

Runs on the edge, too

Bun is the first-class runtime (app.listen()), but the whole lifecycle is app.fetch(Request): Promise<Response> with zero Bun APIs - so the same app deploys to Cloudflare Workers (export default app), Deno (Deno.serve(app.fetch)), or Node (via the @nifrajs/node adapter). See Deployment and Edge & bindings.

Principles (enforced, not aspirational)

• Reject invalid input at three boundaries - compile-time (types), boot-time (config throws loudly), request-time (Standard Schema → structured 422). "Genuine fallback" is a documented whitelist; everything else rejects.
• Tests everywhere, six kinds - unit, type-level (*.test-d.ts), property/fuzz, mode-conformance, benchmark-regression, security-guardrail.
• Speed is a measured goal - tracked with the oha HTTP matrix (bun run bench:http) across Bun, Node, and Deno against raw runtime handlers plus representative API framework baselines.
• Production-grade by default - graceful shutdown, redacting logs, idempotent guards, integer-money discipline; nothing is "we'll fix it later".

Packages

Examples

Runnable, type-checked apps live in examples/:

bun run examples/inline-server.ts
bun run examples/contract-client.ts
bun run examples/schema-openapi.ts
bun run examples/hardened.ts
bun run examples/edge.ts        # app.fetch as a universal handler

Develop

bun install
bun run check          # lint + typecheck (incl. type-level tests) + tests w/ coverage
bun run build          # emit dist/ (js + d.ts) for all packages
bun run check:publish  # build + publint + arethetypeswrong
bun run bench:http     # oha HTTP matrix across Bun/Node/Deno

MIT licensed.