Shipping today

You can ship with Trellis today. What you cannot do — by design — is point Trellis at Convex, Firebase, or Neon as the authoritative op store. That gap is the awkward middle between Trellis's thesis and the BaaS-shaped market.

This guide is the practitioner companion to ADR 0039 (no vendor kernel backends; compatibility bridge instead) and ADR 0040 (lane boundary, open engine, hosted Platform sibling) in the trellis package at docs/adr/0039-no-vendor-kernel-backends-compatibility-bridge.md and docs/adr/0040-lane-boundary-oss-and-hosted-platform.md.

The question behind the question

The answers:

1. Live app data → vendor BaaS (Jazz, InstantDB, Convex) or Trellis kernel if you need causal graph semantics — see Trellis and BaaS
2. Analytics / BI → one-way export, never the write path
3. DX on Trellis graph → typed SDK; for app-only rows, use your BaaS

Three layers (don't mix them)

Confusing sync with realtime is the most common integration mistake. See Collaboration & Sync and Realtime (ephemeral).

Paths that work now

Path A — Hosted Trellis (fastest multiplayer)

Best when: you want Jazz-like live UI without operating a server.

1. Provision a Sprite or Trellis Cloud workspace
2. Point the client at your room URL:

import { liveEntities } from "trellis/client";
import { TrellisDb } from "trellis/client/sdk";

import { Task } from "./schema";

const client = new TrellisDb({
  url: "https://myapp.sprites.app",
  apiKey: process.env.TRELLIS_API_KEY,
});

const tasks = liveEntities(client, Task);
tasks.start();

3. Use framework hooks (trellis/react/typed, etc.) in the UI
4. Mount relay for presence if needed (presenceRelay on deploy — see changelog ADR 0016)

Sovereignty check: you still own exportable .trellis state; the host relays and accelerates — it is not a substitute kernel hidden behind a vendor row store.

Path B — Local kernel + git (fastest solo / agent work)

Best when: agents, CLI, offline-first, file-centric repos.

trellis init
trellis watch          # file ops → graph
trellis issue start TRL-1

Git owns file bytes (4.0+); the op-log holds semantics. See Git bridge and Agent coordination.

Gap today: typed live reads in local embedded mode are still remote-only for WebSocket subscriptions. For local-only UIs, poll trellis query or run db serve and use Path A against localhost.

Path C — Self-hosted server

Best when: you control infra but want the same stack as Sprites.

trellis serve --port 8230
# client: TrellisDb({ url: 'http://localhost:8230' })

PartyKit or a Durable Object relay can front sync/presence without owning ops — see demo/partykit/ and Collaboration & Sync.

Path D — Relay-only third-party cloud

Best when: you already pay for Cloudflare / PartyKit and need cross-tab or cross-browser realtime without durable graph writes there.

• trellis/realtime + WebSocketRelayTransport or DurableObjectRelayTransport
• trellis/sync + PartyKitRoomTransport or IrohSyncTransport

The vendor moves bytes. Your SQLite kernel (or hosted Trellis) still authors ops.

This is the only sanctioned use of "adapters" to commercial realtime infra.

Path E — BaaS for app + Trellis for agents (recommended hybrid)

Best when: you need signup-speed app UI and auditable agent memory.

1. App UI on Jazz, InstantDB, or Convex — live rows, auth, multiplayer
2. Agents on Trellis — lanes, issues, decision traces, repo semantics
3. Link by ID — wiki-links, issue refs; no dual-write

Full pattern, decision table, and anti-patterns: Trellis and BaaS.

For a hosted Rox-shaped product (you operate the stack): see Platform brief in docs/product/platform-brief.md (ADR 0040).

What we explicitly do not ship

If a feature fails the vision test — disconnect the network; does the system still function fully? — a vendor backend made it load-bearing. Roll back.

Bridging from an existing BaaS

You are not stuck forever; you are parallel-running or migrating:

1. New features on Trellis — greenfield graph entities via typed SDK
2. Legacy reads — keep vendor app on life support; Trellis owns new domains
3. Export for analytics — warehouse ETL from Trellis snapshots/JSONL; one-way
4. Cutover — when Trellis covers the domain, retire vendor writes

There is no drop-in adapter. Plan for a domain boundary, not a flip of DATABASE_URL.

Current gaps (honest)

Decision tree

Need durable multiplayer graph state on Trellis semantics?
├─ Yes → Trellis kernel (local, self-hosted, or Sprites)
│         └─ Need live UI? → typed SDK + /realtime (remote mode today)
├─ Need fast app UI + agent audit?
│         └─ Path E: BaaS + Trellis sidecar (see trellis-and-baas)
├─ Only presence/cursors/chat?
│         └─ trellis/realtime + relay (vendor OK as pipe)
└─ Only analytics on Trellis data?
          └─ export / ETL (one-way, async)

Related

• Trellis and BaaS — decision table, sidecar pattern
• Vision — Design principles (local-first, sovereignty)
• Typed SDK
• Collaboration & Sync
• Realtime (ephemeral)
• Cloud hosting
• Git bridge