Deployment topologies

The word “server” hides several boundaries. Start with who initiates work and what protocol crosses the process edge.

Decision table

Need	Use	Status	Boundary
Product code starts a workflow	local `Nika`	Published npm 0.120.2 (bundled engine)	native process
The project clock starts work	`nika serve` (no `--bind`)	Live	`nika.yaml arm[]` to the resident firer
An application admits jobs over HTTP	`Nika` + `nika serve --bind`	Published SDK · public engine	Bearer + OpenAPI · cancellation requests action
The host scheduler starts work	`nika arm --emit`	Live	launchd or systemd to `nika arm fire`
The engine needs local inference	`nika model serve`	Live	Loopback OpenAI-compatible model API

Request-driven application

const run = await nika.run('workflows/support.nika', {
  inputs: { ticket: request.id },
})

for await (const event of run.events()) render(event)
const result = await run.result()

The application owns the user experience and cancellation. The engine owns workflow admission, execution and the receipt. inputs carries literal typed values — see configuration for why the deprecated vars operator channel is not the app door.

Clock-driven project

container / bare VM ── nika serve ──────┐
                                         ├── one firer ── run + ledger
managed host ───────── launchd / systemd ┘

Both lanes consume the same arm: registry. Choose based on who should own process lifetime and wakeups, not because one lane changes workflow semantics.

Inference provider

nika model serve --model Qwen/Qwen3-0.6B-GGUF --port 8712

This process serves a model on loopback for the engine to call. It does not accept Nika workflow jobs.

Loopback HTTP (opt-in)

nika serve --bind is not nika model serve, and it is not bare nika serve. The pinned OpenAPI exposes cancellation as an action; the runtime’s settlement determines the result. Artifact routes, job listing and POST /v1/run remain absent.

Architecture view

PROCESS TOPOLOGY · NAME THE EDGE

Continue

Server surfaces

The compact truth table for similarly named commands.

Resident server

Run the stable project clock.

Remote client

Typed client. Measure it against the OpenAPI of the binary you hold.

Project setup

Place control, intent, application code and state.

Overview

Start

Project

Local API

Runtime

Remote API

Operations

Reference

Deployment topologies

Decision table

Request-driven application

Clock-driven project

Inference provider

Loopback HTTP (opt-in)

Architecture view

Continue

Server surfaces

Resident server

Remote client

Project setup

​Decision table

​Request-driven application

​Clock-driven project

​Inference provider

​Loopback HTTP (opt-in)

​Architecture view

​Continue

Server surfaces

Resident server

Remote client

Project setup

Decision table

Request-driven application

Clock-driven project

Inference provider

Loopback HTTP (opt-in)

Architecture view

Continue