FAQ

FAQ

Honest answers. Where something is not built yet, it says so.

Status

Is it production ready?

v1.1.0 As of 2026-08. Stable API, semver from here, 27 @ultimat3/* packages plus the unscoped create-ultimate — 28 in all — published to npm in lockstep, and 1.1.0 is the first release the workflow published over OIDC trusted publishing. That is exactly what the version claims — a stable API under semver, not a promise about your infrastructure.

What it does not claim:

Not claimed	Detail
A multi-node realtime result	the 50k forced-restart benchmark is measured and committed, but on one sync node over InProcessTransport — it never crossed NATS. Fanout across nodes, throughput, and per-node socket capacity are all still targets, not results (Realtime)
The two-platform deploy proof	all three build targets ship — x build --target docker, x build --target binary, x build --target static — and so do the compose files and the Helm chart. The demo app running on Compose and K8s from one image, with a rolling restart invisible to connected clients, is milestone 11's remaining item (Deployment)
The v2 set	realtime tier 3 (persist: true, local-first), the plugin API, multi-region replication, and the Redis/NATS job drivers — all behind the interfaces that ship today. The job drivers throw X_NOT_IMPLEMENTED with a runnable fix: line rather than pretending to work

What is actually finished?

All 28 packages, implemented and tested — not skeletons. The eight primitives, HTTP, rendering, caching, realtime tiers 1–2, auth, mail, storage, jobs, the AI-first surface (MCP, llm(), evals), and admin + generators + x new. Milestones 0–10 are ✅ and enforced by x verify's roadmap step; milestone 11 is 🚧, open on its two-platform deploy proof. Each milestone ends in a working demo app plus green x verify, and the same demo app grows through all twelve. docs/idea/14-roadmap.md is the source of truth for those markers.

What is left after 1.1.0?

One thing: milestone 11's two-platform deploy proof — the demo app on Compose and K8s from one image, with a rolling restart invisible to connected clients. The other item that was open at 1.0.0, the 50k-socket forced-restart benchmark, is measured and committed at 1.1.0: 50,000 sockets, forced sync restart, time-to-consistent p50 54.0s / p90 105.5s, on one node (Realtime). Milestone 11 lands when it is demonstrated, not on a date — publishing a date is how roadmaps become fiction. Everything in milestones 0–10 is shipped and gated. Scope cuts come off the back, never the middle (M4's budgets). The open defects 1.1.0 shipped with are listed on Known gaps.

The stack

Why Bun only?

Portability costs a lowest-common-denominator API. Bun-only makes Bun.sql, Bun.redis, Bun.s3, Bun.serve WebSockets, Bun.build, and bun test language features instead of dependencies — roughly a thousand transitive packages that never enter the lockfile. Target: under 40 direct dependencies for the whole framework.

Why no Node or Deno support?

Multi-runtime means every native gets an adapter, every adapter gets a second behavior, and the second behavior is where bugs live. It would also make the WebSocket cost model — the thing sync's economics depend on — a per-runtime variable. Stated cost: no native-addon packages, and long-running Bun processes are less battle-proven than Node's.

Why no GraphQL?

A second schema language, a second authz surface, and a resolver-shaped N+1 factory. Typed action + query cover the need, and openapi.json is generated from them. One authz system is the whole point (Policies and authz).

Why no Tailwind?

Tailwind plus modules plus CSS-in-JS in one repo is three token systems and three ways to theme. SCSS modules + design tokens is build-time only, zero runtime, and dark theme is a token flip. See Theming.

Why no React, and no RSC?

Wrong runtime for RSC, and the mental model taxes exactly the audience being optimized for. Solid's fine-grained reactivity makes streaming shells cost ~0 hydration, and stream + <Suspense> gets the same payoff with no new component dialect.

Why SolidJS 2 and your own router?

The router must own render mode, offline strategy, and metadata — those are framework concerns, so it cannot be a third-party dependency that disagrees with the build. As of 2026-08 SolidJS 2 is still pre-release (2.0.0-experimental.16) and the ecosystem around it is thin, which is also why the UI kit is ours.

Which schema library?

None. @ultimat3/schema ships its own dependency-free validators (vendor: 'ultimate'), exposed as t. One schema drives runtime parse, TS type, OpenAPI, and the MCP tool's JSON Schema.

Everything sits behind the Standard Schema v1 interface, so the default is swappable at the framework level — not per app. No ArkType, Zod or Valibot adapter ships; swapping to one means writing the ~40-line configureSchemaProvider() adapter yourself (packages/schema/README.md). A dependency the framework does not need is a dependency every app pays for.

Design decisions people push back on

Why is idempotencyKey required on every job?

At-least-once is the only honest guarantee any queue provides, so every handler must be replay-safe. "Remember to add a key" is exactly the instruction that gets dropped under pressure — by a human at 2am and by an agent always. Making it a required field converts a duplicate-charge incident into a red squiggle. Keys derive from input only: no timestamps, no random, no ctx.

Why is every action an MCP tool?

Because the alternative is two authorization systems. An MCP call and an HTTP call reach the same policy with the same actor resolution — no MCP permission table, no "trusted tool" mode, no broad-rights service account. Actions still opt in per-action with mcp: { expose: true }; what is not optional is that exposure reuses the existing authz.

Why does x verify block on budgets and SEO?

A convention that isn't a build error doesn't exist. A missing description, a blown JS budget, or a drifted migration is discovered in production otherwise — usually by someone who did not write the change. x verify --json names the route, the metric, the actual value, and the import chain that caused it.

Is x verify really the only gate?

Yes. CI runs exactly x verify — no bespoke pipeline steps, because a check that lives only in CI is a check you cannot run locally. Seventeen steps, in this order: typecheck, lint, boundaries, filesize, package-shape, errors, unit, contract, live, job, e2e, eval, drift, contract-diff, budgets, manifest, roadmap. No --only, no --skip. See Testing.

Why are there only eight primitives?

entity, policy, action, mutator, query, job, route, task. If a feature doesn't fit one of them, it doesn't ship. A push notification is a job; a share target is a route with a policy; an admin screen is a query plus actions. No PWA-specific or realtime-specific concept escapes into your mental model.

Running it

Do I need Docker to develop?

No. x dev uses embedded Postgres, in-process NATS, and a local directory in place of S3. bunx create-ultimate myapp && cd myapp && x dev — no Docker install, no .env scavenger hunt. The compose files exist for parity debugging and CI jobs that want real services (Deployment).

Do I need Redis or NATS?

No. Postgres covers the job queue (SELECT ... FOR UPDATE SKIP LOCKED) and pubsub, and a local volume covers files. Redis is a cache tier you opt into; NATS is fanout you need once sync runs on more than one node. Both are drivers behind one interface — job code and query code never change.

Can I deploy serverless or to the edge?

No, by design. Containers only. Edge runtimes, function-per-route, and vendor KV/queue/cron primitives would each need a second implementation of a framework primitive, and the second implementation is where behavior diverges. Requirement: something that runs containers, plus Postgres.

How much infrastructure does a small app need?

One container running ROLE=web and one Postgres — there is no ROLE=all, and x dev is what co-locates roles locally. Split roles when a signal tells you to: RPS for web, connection count for sync, queue depth for worker. scheduler and replicator are fixed at one replica — a second is a warm standby, not throughput.

Can I use it without the realtime tiers?

Yes. realtime.tier: 1 with transport: 'memory' is the default, and a tier-1 app needs no sync role, no replicator, and no NATS. That is a complete product without any realtime at all: entities, actions, a typed client, five render modes, a 0kb static path, and durable jobs.

Risk

What happens if the sync engine doesn't work out?

It is roughly 70% of total effort and the single largest risk. Tiers 1–2 shipped in milestone 6 and are under semver; tier 3 local-first is v2. The reconnect benchmark that gated topology — 50k sockets, a forced sync restart, time-to-consistent and DB load — is measured at 1.1.0: all 50,000 reconnected, 49,981 consistent, p50 54.0s / p90 105.5s, 156,851 connect attempts shed before any query path (Realtime). It was run on one node, so multi-node fanout is still unproven. If the incremental matcher turns out to be the bottleneck, wrapping an existing protocol (Zero's) is an accepted fallback.

Why ship realtime last if it's the differentiator?

A half-built sync engine is worth nothing: it cannot be shipped partially, demoed honestly, or tested against a real app, and it consumes the attention everything else needs. Starting there is how this project would die with 40 packages and no users.

What if Bun has a problem under sustained load?

Stated risk, not a hidden one. As of 2026-08 long-running Bun processes are less proven than Node's, so memory profiling under sustained socket load is explicit roadmap work. See docs/idea/15-risks.md.

Scope

Where do plugins fit?

Nowhere in 1.0 — the plugin API is v2. Semver covers the documented surface, not internals, and a plugin API freezes internals permanently. Fork the blessed path if you need something else; extension points earn their existence from real forks, not from speculation.

Will you add an adapter for my host or my ORM?

No. Multi-ORM means two migration stories and two sets of generated SQL for an agent to learn. Vendor deploy adapters violate the containers-only axiom. Removing an alternative is treated as a feature.

Who is it for?

Primary user: an AI agent — one correct way per task, machine-readable errors, generated facts to read. Secondary: a tired senior engineer, working through their own AI agent and AI reviewer — no glue code, no config archaeology, no 3am pager. Both want the same thing: fewer decisions with consequences.

How is this different from Rails, Next.js, Meteor, or Phoenix?

Compared to	Difference
Rails	same philosophy — conventions, generators, one blessed path — on Bun + Postgres + Solid, with types end to end and an agent as the primary reader
Next.js	render modes are borrowed; jobs, realtime, authz, mail, storage, admin, and MCP are in-box, and there are no platform primitives
Meteor	realtime by default is borrowed; one authz system that live queries reuse per delivered row is the part Meteor's generation got wrong
Phoenix	server-authoritative realtime is borrowed; different runtime, and durable step workflows are a first-class primitive rather than a library choice
Inngest / Zero	their shapes appear as job steps and mutators — assembled into one runtime with one authz system, which is the actual bet

Nothing here is novel in isolation. See Home for the full "steal explicitly" table.

Getting unstuck

Where do I report a bug?

Run x verify --json and x doctor --json, attach both, and open an issue. The JSON is the report. Security issues go through SECURITY.md, never a public issue.

An error code is unclear — where do I look?

x errors explain <CODE> --json, then Error codes. Every X_* code carries a cause, an exact fix command, and a docs URL, rendered identically in the terminal, the browser overlay, and --json. Symptom-first: Troubleshooting.