MDK Logo

Gateway apps

Build, publish, and install MDK apps with the mdk-app.yaml manifest, the mdk-plugin.json route table, and a UI module

Overview

The Gateway serves HTTP routes from apps. An app is a package with an mdk-app.yaml manifest, a route table (mdk-plugin.json, in plugin/ or at the package root) and one controller file per route. Each app runs in its own process under @tetherto/mdk-app-runtime: the Gateway reads only the route table and forwards each request to that runtime over HRPC, so no app code runs inside the Gateway. A Gateway serves exactly the apps your stack declares in spec.apps[] and nothing else.

The app runtime builds the app's @tetherto/mdk-client and hands it to every controller as config.mdkClient, signed in as the app's service account. No knowledge of the MDK Protocol envelope or internal message shapes is required.

Prerequisites

Apps MDK ships

The supported apps reference lists where to get each catalog app. Each is a demo to read alongside this guide. To try a Gateway app with the 0.14 CLI, start with the local fleet scaffold below.

A stack serves only the apps it declares. Any other path answers 404. If a React adapter hook reads a route, declare the app that serves it.

Third-party apps are documented by their own manifests. mdk add app <source> installs one from npm or GitHub, checks the package as mdk validate app does, and only then adds it under spec.apps.

Create an app

mdk create app fleet

This scaffolds apps/fleet, links it into the project's package.json, and adds it under spec.apps in mdk.yaml (--no-stack-entry prints the entry instead). For an app that lives in its own repo and installs with mdk add app github:<owner>/<repo>#<tag>, run mdk create app fleet --flat in the repo instead: it scaffolds at the repo root and touches no project. The scaffold carries one read-only summary route and a starter UI that shows it:

apps/fleet/
├── mdk-app.yaml               # app manifest
├── package.json               # `files` ships the manifest, plugin/, ui/, and README.md
├── README.md
├── plugin/
│   ├── mdk-plugin.json        # route table
│   ├── lib/client.js          # re-exports the app runtime's Kernel client
│   ├── types.d.ts             # response types, shared with the UI
│   └── controllers/summary.js # one module per route
├── ui-src/                    # UI source with its own package.json and Vite build, not published
└── ui/index.js                # the built UI module, committed; a placeholder until `npm run build:ui`

mdk-app.yaml describes the app to the operator: its id, version, displayName, the Kernel permissions it asks for, default config and optional background main. Every field is in the app runtime's manifest reference. Keep version equal to the package.json version; mdk validate app checks it.

The Kernel enforces spec.rbac on every call the app makes. mdk add app grants the app's service account (app:<name>) exactly these permissions, as a role in mdk.yaml the operator may narrow, so request every Kernel resource the app reads or writes: a call outside them throws ERR_KERNEL_ACTION_FORBIDDEN, which the route answers as the app runtime sends it.

Check it with mdk validate app fleet; what it checks is in the CLI reference.

Write the routes

2.1 Edit the route table

mdk-plugin.json declares the plugin identity (name, version) and a routes array. Each route needs an id, a handler path, and either an http block with a method and path, or those same method/path fields flattened to the route's top level; the agent app uses the flat form. A table that declares "routeTableVersion": "1", as mdk create app writes it, is held to the route-table schema's OpenAPI-shaped rules as errors by mdk validate app; a table without it gets them as warnings. Rather than copy a synthetic example, start from a real manifest and trim it:

  • examples/mvp-site/backend/gateway-plugins/site/mdk-plugin.json: GET routes with query parameters, and POST routes with a requestBody and path parameters
  • mdk-app-demo/mdk-plugin.json: read-only GET routes with query parameters, each annotated with safety
  • mdk-app-agent/mdk-plugin.json: auth: true routes in the flat form, one of them a stream route

The demo and agent apps live in their own repositories; the mvp-site manifest is the one in this repository.

Path parameters use {param} syntax; the loader normalizes them to Fastify's :param format. For named exports use "handler": "./controllers/foo.js#namedExport". Every field is listed in the route-table field reference, generated from the schema the app runtime publishes. The loader enforces the fields at boot; what it refuses, and why is app-runtime's ERR_PLUGIN_MANIFEST_INVALID row.

Give each route a description that says what it does and when to call it. It becomes the description of the MCP tool derived from the route, and the route's Purpose in an HTTP reference if you generate one; without one, mdk validate app warns, and a generated reference shows "Not declared".

The Gateway validates a route's schema and requestBody with type coercion: write a nullable number, string, boolean, or array as a type list ("type": ["number", "null"]), not as the anyOf that z.toJSONSchema writes for a bounded nullable, or a JSON null reaches the controller as 0, "", false, or an array.

2.2 Write a controller

The app runtime builds the app's @tetherto/mdk-client and hands it to controllers as config.mdkClient; the scaffold's lib/client.js re-exports it, and every controller in the app requires that module.

Every controller exports an async function (req):

// controllers/live.js: read live telemetry
const mdkClient = require('../lib/client')

module.exports = async function live (req) {
  const deviceId = req.query.deviceId
  const telemetry = await mdkClient.pullTelemetry(deviceId, 'metrics')

  // No telemetry: no Worker owns the device, or its Worker could not read it
  if (telemetry.error) {
    const code = String(telemetry.error)
    const statusCode = code.startsWith('ERR_DEVICE_NOT_FOUND') ? 404 : 503
    throw Object.assign(new Error(code), { statusCode })
  }

  return { deviceId, ...telemetry }
}

A write a user asks for goes through req.mdk, the request's Kernel client. It acts for the signed-in user, so the Kernel checks their own spec.rbac rules as well as the app's, and the write-command log records them via the app:

// controllers/command.js: dispatch a command for the signed-in user
module.exports = async function command (req) {
  const deviceId = req.params.deviceId
  const { mode } = req.body

  const result = await req.mdk.sendCommand(deviceId, 'setPowerMode', { mode })

  // The Kernel refused the command, such as an undeclared command or a param outside the contract's rules
  if (result.status === 'REJECTED') throw new Error(result.error)

  return {
    deviceId,
    commandId: result.commandId,
    status: result.status
  }
}

Neither call throws on a Kernel error reply: pullTelemetry and sendCommand each resolve with the reason in error. Returned as is, that reply reaches the caller as 200. Thrown with its code, it gets the status the controller rule gives that code, or the statusCode the controller attaches, as live.js does. An authorization refusal throws instead, with its code as err.code, and the same rule answers it.

The app's own rules must allow the write too, so an app with a route like this one requests { resources: [devices], verbs: [write] } in its permissions. Without it the Kernel refuses the command with ERR_KERNEL_ACTION_FORBIDDEN.

Declare and mount it

mdk create app already wrote this entry; a hand-written one has the same shape:

spec:
  apps:
    - name: fleet          # ^[a-z][a-z0-9-]*$, unique in the stack
      package: fleet       # npm name or local path
      config: {}           # optional: overrides the mdk-app.yaml config defaults
      mount: /             # optional: serve the routes at the root instead of /apps/fleet

The Gateway serves an app's routes under /apps/<name> unless mount is set, so the scaffold's GET /api/fleet/summary answers at /apps/fleet/api/fleet/summary, or at /api/fleet/summary with mount: /. Mount an app at the root only when a client calls its paths there, as the agent app needs for its UI.

Run it

If your stack has no administrative binding, run mdk rbac init before its first start. Projects created with mdk onboard already have the CLI admin binding.

mdk run             # Kernel, Workers, one runtime process per app, then the Gateway
mdk run app fleet   # or only this app, in the foreground, beside `mdk run kernel` and `mdk run gateway`

The runtime loads the route table and every handler before it serves, so a missing file or invalid manifest fails that app at startup. It accepts calls only from the keys mdk run allows, the Gateway's (.mdk/gateway.key) and the MCP server's (.mdk/mcp.key); until it is up, its routes answer 503.

No route is open, so sign in and send the access token to reach http://localhost:3847/apps/fleet/api/fleet/summary.

An app's routes become MCP tools with no separate manifest. The operator agent calls them only with strictToolGating off.

The app's UI module

An app's routes are one half of it, and this is the other. ui/index.js default-exports the result of defineMdkApp from @tetherto/mdk-react-adapter, an identity function that types the contract a shell loads:

// ui-src/index.tsx, built to ui/index.js
export default defineMdkApp({
  id: 'com.example.fleet',
  displayName: 'Fleet',
  routes: [{ path: '', component: Overview }],
  panels: [{ id: 'drift', slot: 'home', component: DriftCard }]
})

mdk create app scaffolds ui-src/ with one page and one Home panel, plus a placeholder ui/index.js that npm run build:ui replaces.

Share response types with the UI

Controllers are CommonJS JavaScript: the app runtime loads .js and .cjs from plugin/, so a TypeScript controller must be compiled to one first. Type a route's response once, in plugin/types.d.ts, and check both halves against it. The controller names the type with JSDoc under // @ts-check, and the UI imports it with import type, which the UI build erases:

// plugin/controllers/summary.js
// @ts-check
/** @typedef {import('../types').AppRequest} AppRequest */
/** @typedef {import('../types').Summary} Summary */

/**
 * @param {AppRequest} req
 * @returns {Promise<Summary>}
 */
module.exports = async function summary (req) { /* ... */ }
// ui-src/data/summary.ts
import type { Summary } from '../../plugin/types'

AppRequest in the same file types the req every controller receives, so new controllers do not re-type it. The scaffold wires the summary route this way, and npm run typecheck checks both halves, so renaming a field fails the type check in the controller and in every component that reads it.

FieldContains
idThe app's reverse-DNS id. Must equal id in mdk-app.yaml: the shell checks it against the installed app's id, so a package cannot load under another app's identity
displayNameRepeats displayName from mdk-app.yaml, which is what the shell shows; the shell checks only that this one is set. A panel with no title falls back to it, as do some unlabelled sidebar entries
routesPages the shell mounts under /app/<name>/. path is relative, '' being the landing page; label sets the sidebar label; section groups the page with the others naming it; hidden routes a deep link without listing it
sectionsOptional. An icon for a route section: label is the section its routes name, icon a component the shell renders with no props. A section works with no entry here, which only adds what a label alone cannot
panelsComponents the shell composes into pages it owns. slot picks where: every MDK shell provides MDK_APP_PANEL_SLOT.HOME, and a shell ignores slots it does not know

The sidebar lists only routes that are not hidden and have no : or * in their path, because a deep link needs a value only another page can give it. With one such route, the app gets a single entry, labelled with the route's label or else displayName, whatever its path. With several, the app gets a group named displayName, where an unlabelled entry shows its raw path, or displayName for the landing page. When the app is the only one installed and its routes use section, its sections and other pages are listed at the top level, without the group.

The dashboard serves its app list at /app-ui.json, written from the stack's mdk.yaml by mdk run dashboard. It imports each app's ui/index.js, checks it against the contract, and adds its stylesheets to the page, so installing an app needs no dashboard change: restart the dashboard and the app's pages and panels are there. An app whose UI fails to load is reported on its own instead of taking the dashboard down.

A built UI module imports React and the MDK packages by bare specifier, and the dashboard resolves those to its own instances through an import map. A shell that does not map them fails to load every app UI.

Publish and install an app

An app is an npm package, so shipping one is npm publish or a Git tag. What makes it installable is what the tarball carries and what the consumer runs.

Package it

package.json files has to ship mdk-app.yaml, plugin/, and ui/. Commit the built ui/ rather than building on install: a GitHub install runs no build, so an unbuilt ui/ reaches the consumer as a missing module.

{
  "files": ["mdk-app.yaml", "plugin/", "ui/", "README.md"]
}

Run mdk validate app <name> before publishing or tagging.

An app you plan to list also has to meet the listing criteria, which add smoke checks against a mock Kernel.

Publish it

Either source works, and an app does not have to reach npm at all:

git tag v0.2.0 && git push --tags   # install from GitHub at that tag
npm publish --access public         # or publish under your own scope

Install it

A consumer adds the app to their stack from either source, then starts it:

mdk add app @your-org/mdk-app-fleet@^1
mdk add app github:your-org/mdk-app-fleet#v0.2.0 --name fleet
mdk run

mdk add app installs with npm install --ignore-scripts --save-exact, so no install script of an unchecked package runs and the exact version is recorded. For a GitHub source, that's the commit the lockfile resolved, written into package.json as github:<owner>/<repo>#<sha>. It then checks the package as mdk validate app does and only then writes the spec.apps entry, keeping the stack file's comments and formatting. A refused package leaves mdk.yaml, package.json and the lockfile as they were. The CLI reference covers --name, repeat installs, and the worker equivalent.

A GitHub source cannot be added with --no-install: its package name is only known once npm has fetched it.

Declare what it needs

An app states its own config defaults under config in mdk-app.yaml, and the operator overrides them per stack. Supplying per-app config covers the layering.

Start apps from code

Outside the CLI, pass each app to startGateway() from @tetherto/mdk-core as apps (the complete list, nothing else is mounted) and run each one with createAppRuntime(), allowing only the Gateway's key and passing the app's service-account credential, which mdk add app writes to .mdk/credentials/apps/<name>.json (mdk run also writes it at boot when it is missing). mdk create app doesn't write it, so for a local fleet app already linked by mdk create app, run mdk add app fleet --no-install once first. The boot script also needs two packages that neither mdk create app nor mdk add app installs:

npm install @tetherto/mdk-core @tetherto/mdk-app-runtime

Then boot the Kernel, the app, and the Gateway:

const path = require('path')
const { getKernel, startGateway } = require('@tetherto/mdk-core')
const { createAppRuntime } = require('@tetherto/mdk-app-runtime')
const fleet = require('./.mdk/credentials/apps/fleet.json') // written by mdk add app

const main = async () => {
  // Registers app:fleet and grants it what its routes call, as mdk-app.yaml permissions would under mdk run
  const kernel = await getKernel({
    rbac: {
      enforcement: 'deny',
      roles: [{
        name: 'app-fleet',
        rules: [{ resources: ['workers'], verbs: ['list'] }, { resources: ['devices'], verbs: ['read', 'write'] }]
      }],
      bindings: [{ role: 'app-fleet', subjects: [{ kind: 'ServiceAccount', name: 'app:fleet' }] }]
    },
    serviceAccounts: [{ name: fleet.name, publicKey: fleet.publicKey }]
  })

  const appDir = path.join(__dirname, 'apps/fleet')

  const runtime = createAppRuntime({
    name: 'fleet',
    appDir,
    kernelKey: kernel.getPublicKey(),
    credentialFile: './.mdk/credentials/apps/fleet.json',
    storeDir: './run/apps/fleet',
    keyFile: './run/fleet.key',
    allowedKeyFiles: ['./run/gateway.key'],
    kernelKeyFile: './run/fleet.kernel.pub'
  })
  await runtime.start()
  kernel.allowKey(runtime.getKernelPublicKey())

  await startGateway({
    kernel,
    port: 3000,
    gatewayKeyFile: './run/gateway.key',
    apps: [{ name: 'fleet', appDir, keyFile: './run/fleet.key' }]
  })
}

main().catch((err) => { console.error(err); process.exit(1) })

Without network options, the Kernel, runtime, and Gateway join the public HyperDHT. On a machine that cannot reach it, pass the private DHT options to each of them.

Under mdk run, the CLI registers each app's account and applies its permissions for you. From code, pass serviceAccounts and rbac in the boot script so the app can authenticate and make its calls, unless the Kernel's stored policy and accounts already cover the app.

rbac and serviceAccounts are the Kernel's complete policy and service-account list, applied at every start: the policy replaces the stored one, and any service account not listed is removed. Users stay but lose their bindings. On a Kernel that already has a policy, users, or accounts, merge the app-fleet role, binding, and account into the ones it already has, or everyone else loses access. The getKernel options define both.

To create a user on this Kernel, call await kernel.identity.upsertUser({ name, password }) from your boot script using a password supplied by the operator. Bind the user to a role in the same rbac policy with the access they need, then sign in and call a route. mdk user add doesn't reach this Kernel: it reads the Kernel key from .mdk/kernel.key, which mdk run writes, while getKernel() writes its key to ~/.mdk/.kernel-key by default.

The runtime's Kernel key must be admitted before its controllers call the Kernel, through kernel.allowKey(), as the example does, or the trusted-keys directory. The credential is the app's service account, app:fleet, as mdk add app writes it to .mdk/credentials/apps/<name>.json; a runtime with a kernelKey and no credential logs a warning, and the Kernel refuses every call it makes. startGateway()'s apps reference lists the validation errors, and @tetherto/mdk-app-runtime the runtime options. startGateway() throws ERR_GATEWAY_EXTRA_PLUGIN_DIRS_UNSUPPORTED when passed extraPluginDirs.

The req object

A controller's only argument: params, query, body, and headers from the request, _info with the identity the Gateway verified, and mdk, the request's Kernel client. The app runtime's reference describes how req.mdk and req._info are assembled.

The app's context module

require('@tetherto/mdk-gateway/plugin') resolves, inside an app's controllers, to that app's own frozen context. The scaffold's lib/client.js reads config.mdkClient from it:

FieldTypeContains
configobjectThe mdk-app.yaml config defaults with the app's spec.apps[].config layered over them key-by-key, plus mdkClient: the app runtime's Kernel client, signed in as the app's service account (app:<name>), so every call is checked against spec.rbac
loggerobjectThe app runtime's logger, namespaced mdk:app:<name>
onReadyfunctionRegisters a callback that runs once the app runtime is serving, or straight away if it already is. Errors thrown inside it are caught and logged, not fatal

Supplying per-app config

That config isn't declared in mdk-plugin.json. Defaults live under config in mdk-app.yaml, and the stack spec overrides them per app with spec.apps[].config (createAppRuntime({ config }) from code):

spec:
  apps:
    - name: custom-metrics
      package: ./apps/custom-metrics
      config:
        apiKey: ${METRICS_API_KEY}

config.mdkClient is already built, so require the scaffold's lib/client.js from every controller that needs it. There is no per-request Kernel access to guard, only the client's own connect failures:

The client connects on first use and memoizes the connection. A failure arrives with err.code set to ERR_MDK_CLIENT_UNAVAILABLE and a message that adds the cause, and the next call dials again. Guard the call on err.code, not on the message and not on a null client:

let telemetry
try {
  telemetry = await mdkClient.pullTelemetry(deviceId, 'metrics')
} catch (err) {
  // The Kernel is down, not the caller's fault: answer 503
  if (err.code === 'ERR_MDK_CLIENT_UNAVAILABLE') throw Object.assign(new Error('ERR_KERNEL_UNREACHABLE'), { statusCode: 503 })
  throw err
}

// A Kernel error reply resolves instead of throwing: answer it as `live.js` does
if (telemetry.error) {
  const code = String(telemetry.error)
  const statusCode = code.startsWith('ERR_DEVICE_NOT_FOUND') ? 404 : 503
  throw Object.assign(new Error(code), { statusCode })
}
return telemetry

Read hardware data

Call the client directly for live device data: pullTelemetry, getCapabilities, and listWorkers are documented with their return shapes in the client's own reference.

A fleet-wide number, hashrate across every miner on site say, must never start a second hardware polling loop. Workers whose contract declares telemetryPollIntervalMs already poll their devices and store the series, so read the fleet from those stores in one call with pullSnapshot, or let the Workers compute the number with pullAggregate:

const { workers } = await mdkClient.listWorkers()
const workerIds = workers.map((w) => w.workerId)

const snap = await mdkClient.pullSnapshot({ workers: workerIds, metrics: ['hashrate_avg'] })
const values = Object.values(snap.devices).map((d) => d.metrics.hashrate_avg?.value).filter(Number.isFinite)
const totalHashrate = values.reduce((sum, v) => sum + v, 0)
// Workers that could not answer, such as one that does not poll, are not in the total
const leftOut = Object.keys(snap.errors ?? {})

const agg = await mdkClient.pullAggregate({ workers: workerIds, ops: { hashrate_total: { op: 'sum', src: 'metrics.hashrate_avg' } } })
// agg.ops.hashrate_total, agg.deviceCount, agg.errors (Workers that could not answer)

Each Worker names its metrics in its own contract, and hashrate_avg is the Avalon Worker's name. pullSnapshot still returns a device whose Worker names hashrate differently, with no error and no hashrate_avg in its metrics, so count values, not snap.devices entries; the example's Number.isFinite filter drops the missing value. pullAggregate reports that Worker in agg.errors as ERR_AGGREGATE_SRC_UNKNOWN.

A Worker that does not poll has no store, so both calls report it under errors as ERR_TELEMETRY_STORE_DISABLED and snap.devices holds none of its devices. Return leftOut with totalHashrate, so the caller can tell a partial total from the fleet's, then read such a Worker live with pullTelemetry per device, or ask its operator to set telemetryPollIntervalMs. The miner Workers on the Starter site declare no poll interval, so on that site both calls report every miner under errors, and the fleet read falls back to pullTelemetry per device.

An op groups or filters on a metric, on deviceId, or on a label the operator gave the device in mdk.yaml under config.devices[].info (group: 'info.container'): the labels are the only part of a device's configuration an app can read this way.

Read deviceCount and errors before you treat a number as a fleet total. deviceCount counts the devices whose stored point held a reading, so a listed Worker that is registered with no devices answers, adds nothing to deviceCount and leaves every op at its empty value (a sum of 0). ops is {} only when no listed Worker answered at all, and each of those is then under errors. A Worker with no devices resolves to no device name, so a caller whose devices read rule is narrowed with resourceNames is refused for a list that names it (ERR_KERNEL_ACTION_FORBIDDEN); only an unscoped rule covers it.

Before you swap sum for avg, read what a merged avg answers: it may not be the figure your question needs.

The Kernel checks devices read for every device of each listed Worker, so an app that reads the fleet requests it in mdk-app.yaml beside the scaffold's workers list:

permissions:
  - { resources: [workers], verbs: [list] }
  - { resources: [devices], verbs: [read] }

Without the rule the call throws ERR_KERNEL_ACTION_FORBIDDEN, which the route answers as the app runtime sends it, with that message. A Worker that collects its own telemetry is left out of the Kernel's scheduled pull, so nothing polls its devices twice.

pullTelemetry(deviceId, 'metrics') still reads one device live when the question is about that device right now.

demo/controllers/summary.js is a small worked example of a fleet read, worth starting from if you're authoring your own app.

There is no separate Gateway-side store for historical or aggregated data, either. Read the series from each Worker's own persisted tail-log with pullTelemetry, addressed to a device on a Worker that keeps one, such as a container:

const history = await mdkClient.pullTelemetry(containerId, { type: 'logs', key: 'stat-1D', tag: 't-container', start, end })

// A refusal resolves, such as ERR_UNKNOWN_QUERY_TYPE from a Worker that keeps no log
if (history.error) throw new Error(history.error)

The Kernel authorizes a pullTelemetry call as devices read on the device it addresses and does not inspect the query; what comes back is decided by the Worker. These store-backed query types exist on a Worker that runs the Worker Runtime's store services (in this repo, the container and temperature Workers); a WorkerRuntimeV2 Worker keeps its history in its telemetry store and answers pullSnapshot with from instead. mdk run boots every Worker on WorkerRuntimeV2, so a Worker on a CLI stack answers only list of these query types, with every device it runs. It refuses the others: a read addressed to one device gets ERR_UNKNOWN_QUERY_TYPE, unless the Worker's contract declares a telemetry channel of that name, and a read with no device addressed, as pullWorkerQuery sends, gets ERR_DEVICE_ID_REQUIRED. A host that starts a store-services Worker itself, as the Antspace development script does, can serve them.

Query typeWhat a read addressed to one device returns
logs with key: 'thing-5m', logs_multi, historical_logsThe addressed device's own snapshot series, alerts, and info changes; the tag and deviceIds fields cannot point it at another device
logs with a stat-* keyThe series the Worker stores for tag, aggregated over every device carrying that tag; a tag the addressed device does not carry is refused with ERR_LOG_TAG_FORBIDDEN
list, stats, thing_configThe addressed device only
count, config, settings, ext_dataWorker-level facts: the device count, the contract, the settings, or the pool data, whichever device you address

To read a Worker's own records rather than one device's, ask the Worker: pullWorkerQuery runs the query on each listed Worker with no device addressed and answers per Worker, and the Kernel checks devices read on every device of each listed Worker, as it does for pullSnapshot:

const { workers: perWorker, errors } = await mdkClient.pullWorkerQuery({ workers: [workerId], type: 'list', status: true, limit: 200 })
// perWorker[workerId].things: the Worker's devices, up to `limit`, each with its latest snapshot (a WorkerRuntimeV2 Worker answers `devices`); errors?.[workerId] when it could not answer
const series = await mdkClient.pullWorkerQuery({ workers: [workerId], type: 'logs', key: 'stat-1D', tag: 't-container', start, end })

A stat-* log is shared: a response can include figures from devices the caller cannot address, and a caller scoped to one device still reads the shared total (what resourceNames does and does not bound). Do not treat one device as standing for its whole Worker: a device-addressed read answers for that device. A Worker with no current devices still has its retained history; read it with pullWorkerQuery, which for a Worker with no devices only an unscoped devices read rule allows.

demo/controllers/history.js reads history a different way: it pulls a history telemetry channel that the Worker's contract declares (pullTelemetry(deviceId, { type: 'history', limit })), with an optional per-device filter and a Kernel-unavailable fallback.

Both live and historical calls are network calls through the client, so guard both the same way. An unguarded pullTelemetry throws ERR_MDK_CLIENT_UNAVAILABLE straight through to the caller as a 400 the moment the Kernel drops. When only a Worker drops, pullTelemetry does not throw: it answers { error } (ERR_WORKER_UNREACHABLE, ERR_WORKER_NOT_ROUTABLE, or ERR_DEVICE_NOT_FOUND), so check res.error before the field you read. A Worker that hangs instead can still time the request out with ERR_MDK_REQUEST_TIMEOUT.

Send a command

sendCommand dispatches via the Kernel to the Worker that owns the device. The command must be declared in the Worker's mdk-contract.json, and the command controller shows the pattern. The return value is the Kernel's acceptance, { commandId, status: 'QUEUED' } or { error, status: 'REJECTED' }, not the command's outcome; read that with getCommandStatus.

Caching

Add a "cache" array of dot-path strings to a route to enable request-level caching. The key is the app name, the route id and those values; on a route the Gateway authenticates, every route under spec.rbac or an auth: true route under authVerifier, it adds a hash of the caller's identity, so a per-caller answer needs no identity path of its own. The auth check runs before the cache lookup. Who may force a refresh with ?overwriteCache=true is in the Gateway's security model, and the route-table field reference lists the field.

Stream routes

Add "stream": true to a route to write the response yourself, through a second argument that behaves like a ServerResponse over the runtime's channel, instead of returning a plain value: for SSE or any other response Fastify shouldn't serialize. Stream backpressure covers the mechanism. A handler that throws before it sends headers answers a JSON error body; after that the channel closes without an end, so the caller sees a truncated body. mdk-app-agent is a shipping example: its message route streams text/event-stream this way, and the agent Gateway-deployment guide covers the consumer side.

The HTTP caller's pace doesn't reach your handler: raw.write() reports only the app runtime's own connection to the Gateway. When an HTTP caller falls more than the Gateway's maxStreamBufferBytes (default 16 MiB) behind the handler, the Gateway closes the stream and logs ERR_APP_STREAM_OVERFLOW at warn: the caller has the status the handler sent and part of the body, and the connection closes before the response completes. Write a stream a slow caller can keep up with, and have the caller treat an incomplete response as a failure.

Auth and permissions

Every route of an app is checked against spec.rbac at the Gateway: the caller needs a bearer access token (POST /auth/login issues one) and a rule granting the route's resource and verb. Without a token the Gateway answers 401 ERR_AUTH_UNAUTHORIZED, and without a rule 403 ERR_AUTH_FORBIDDEN, before body validation, the cache, and your controller. The Gateway reference lists every error an app route can answer, in the order the checks run. Who may call the route is the operator's to set in mdk.yaml; the manifest only says how the route is governed:

Route fieldStatusDescription
verbOptionalget, list, create, update, or delete. Without it, the verb comes from the HTTP method, as the Gateway's security model maps it
resourceNameOptional{ "param": "<name>" }, { "query": "<name>" }, or { "body": "<name>" }: where the instance a rule's resourceNames matches comes from

A route's auth and permissions fields aren't read here: they're input to the authVerifier fallback, so a permissions list gates nothing under spec.rbac. mdk validate app warns about a route that sets permissions, and the Gateway logs each such route at boot.

The resource is <app>/<path>: the app's name in spec.apps and the route's path template, so the route /miners/:id of ops-center is ops-center/miners/:id. A rule for it:

spec:
  rbac:
    roles:
      - name: rack3-tech
        rules:
          - resources: [ops-center/miners/command]
            verbs: [create]
            resourceNames: ["rack-3-*"]      # matched against the route's resourceName field
    bindings:
      - role: rack3-tech
        subjects: [{ kind: User, name: tech@example.com }]

The verified identity reaches the controller as req._info, { kind, sub, email?, groups? }, so identity never comes from the request body. The Gateway's security model lists the rest of the defaults.

Call a route

Sign in for an access token, then send it as a bearer token. The user to sign in as is the first operator that mdk user add <name> --admin creates:

TOKEN=$(curl -s -X POST http://127.0.0.1:<gateway-port>/auth/login \
  -H 'content-type: application/json' \
  -d '{"username":"<user>","password":"<password>"}' | jq -er .accessToken)

curl -s -H "Authorization: Bearer $TOKEN" \
  http://127.0.0.1:<gateway-port>/apps/<app-name>/api/<domain>/<resource> | jq .

The token lasts minutes and the refresh token rotates on every use, so a long-running client signs in again rather than holding one token. A browser dashboard uses passwordAuth() instead of signing in this way.

A Gateway started with no Kernel and no credential falls back to an authVerifier for routes marked "auth": true:

await startGateway({
  kernelKey: false,
  gatewayKeyFile: './run/gateway.key',
  apps: [{ name: 'fleet', appDir, keyFile: './run/fleet.key' }],
  // route is { id, method, path, app, permissions } from the manifest
  authVerifier: async (request, route) => {
    const token = request.headers.authorization?.replace('Bearer ', '')
    if (!token) return false
    return { kind: 'user', sub: await validateToken(token) }
  }
})

The controller reads whatever the Gateway verified: { kind, sub, email?, groups? } under spec.rbac, or exactly what authVerifier returned:

module.exports = async function protectedRoute (req) {
  const { sub } = req._info   // who asked; under spec.rbac, req.mdk acts for a user, but as the app for a service account
  const row = await findRowFor(sub)
  if (!row) throw Object.assign(new Error('ERR_ROW_NOT_FOUND'), { statusCode: 404 })
  return row
}

The Gateway forwards that identity to the app runtime with the request, so req._info is set in the app's process too.

A controller's return value always goes out as 200, as JSON: it crosses from the app runtime to the Gateway serialized, so return plain data (a Buffer arrives as { type: 'Buffer', data: [...] }; send bytes from a stream: true route). A thrown error's status and message follow the app runtime's rule, and the Gateway sends it as { statusCode, error, message } (errorResponse). Start the message with an ERR_* code when the caller should read it; any other message reaches the caller as the status text, and the app runtime logs the original. Attach .statusCode when the default status doesn't fit: 400 for an ERR_* error, since most are input errors (403 for ERR_KERNEL_ACTION_FORBIDDEN), and 500 for any other. Always attach it to an ERR_* error that isn't the caller's fault, such as 503 for a dependency that's down. protectedRoute attaches 404 here, and the bundled agent app uses the same pattern for its 503, 404, and 409 responses. Fastify's own 4xx errors (validation, content type, body size) keep their message.

Manifest validation errors

The app runtime validates the route table and every handler at startup, and the Gateway the route table, and each throws if anything is wrong. The loader's error codes list every case.

Next steps

On this page