sailfork / t3code-github / .repos/ alchemy-effect/ website/ src/ content/ docs/ infrastructure-as-code/ custom-provider.mdx

Providers are Effect Layers — adding support for a new cloud or third-party API is "declare a type, implement a Layer". See Providers for the operation contract.

This guide walks through building a Stripe Product provider end-to-end: declaring props and attributes, defining the resource type, implementing the lifecycle (reconcile, delete, diff, read), bundling it into a providers() layer, and writing a test.

See Resource lifecycle for the semantics of when each lifecycle method fires.

Declare props and attributes

Every resource has two sides:

  • Input properties — the desired configuration you pass in
  • Output attributes — the values the cloud returns after creation

Start with two plain TypeScript types. Both are pure data, so they're trivial to share between the provider and call sites.

Create src/stripe/Product.ts:

// src/stripe/Product.ts
export interface StripeProductProps {
  name: string;
  description?: string;
  active?: boolean;
}

export interface StripeProductAttributes {
  productId: string;
  created: number;
}

Declare the Resource type

A Resource<Type, Props, Attributes> is a phantom type that ties a string Type to its props and attributes. The string Type (here "Stripe.Product") is what Alchemy uses to look up the provider at plan time — it must be globally unique.

// src/stripe/Product.ts
+import { Resource } from "alchemy";
+
export interface StripeProductProps {
  name: string;
  description?: string;
  active?: boolean;
}

export interface StripeProductAttributes {
  productId: string;
  created: number;
}
+
+export type StripeProduct = Resource<
+  "Stripe.Product",
+  StripeProductProps,
+  StripeProductAttributes
+>;

Declare the Resource constructor (the "tag")

Resource<T>(type) returns the value users actually call — StripeProduct("Pro", { ... }). It also doubles as the tag the provider Layer registers itself against, so by convention the type and the value share the same name.

// src/stripe/Product.ts
import { Resource } from "alchemy";

// ... props / attributes / type unchanged ...

export type StripeProduct = Resource<
  "Stripe.Product",
  StripeProductProps,
  StripeProductAttributes
>;
+
+export const StripeProduct = Resource<StripeProduct>("Stripe.Product");

You can already use this constructor in a stack — but with no provider registered, planning will fail with Provider not found for Stripe.Product. Let's fix that.

Scaffold the provider layer

A provider layer is a Layer<Provider<R>> produced by Provider.succeed(ResourceClass, service). The service is an object with reconcile, delete, and list (required) plus optional hooks like diff, read, precreate.

Start with stubs so the types compile, then fill them in:

// src/stripe/Product.ts
-import { Resource } from "alchemy";
+import * as Provider from "alchemy/Provider";
+import { Resource } from "alchemy";
+import * as Effect from "effect/Effect";

// ... props / attributes / type / constructor unchanged ...

export const StripeProduct = Resource<StripeProduct>("Stripe.Product");
+
+export const StripeProductProvider = () =>
+  Provider.succeed(
+    StripeProduct,
+    StripeProduct.Provider.of({
+      reconcile: () => Effect.die("not implemented"),
+      delete: () => Effect.die("not implemented"),
+      list: () => Effect.succeed([]),
+    }),
+  );

A few patterns worth knowing:

  • Provider.succeed wraps a ProviderService into a Layer<Provider<StripeProduct>>.
  • StripeProduct.Provider.of({...}) is a typed constructor — it forces every method's input/output to match the resource's props and attributes.
  • Provider.effect exists for services that need a dependency resolved once at construction (a logger, FileSystem) — never a live one; see the caution below.

Credentials

Handlers that call an authenticated API resolve a lazy credentials service — nothing touches the network until an operation actually runs. The Custom Auth Provider guide builds the StripeCredentials service the snippets below import; providers wrapping unauthenticated APIs skip this entirely.

Implement reconcile

reconcile is the single lifecycle method that converges the cloud's actual state to the desired state (news). It runs on every apply, so one body must handle three situations, distinguished by output and olds:

outputoldsSituation
undefinedundefinedGreenfield — nothing exists yet
defineddefinedRoutine update
definedundefinedAdoption — the engine imported it via read

Every reconciler follows the same shape:

1. Observe   — derive the physical id; read live cloud state
2. Ensure    — if missing, create it; tolerate AlreadyExists/etc.
3. Sync      — for each mutable aspect: read observed, diff vs
                desired, apply only the delta
4. Return    — fresh Attributes

:::warning[Don't branch the body on output === undefined] Writing if (output === undefined) { /* create */ } else { /* update */ } just renames the old create/update split. Write one flow that converges from any starting point — trust observed cloud state, not olds. :::

Let's build the Stripe reconciler step by step.

Resolve the client

+import * as Redacted from "effect/Redacted";
+import Stripe from "stripe";
+import { StripeCredentials } from "./Credentials.ts";
+
return StripeProduct.Provider.of({
-  reconcile: () => Effect.die("not implemented"),
+  reconcile: Effect.fn(function* ({ news, output }) {
+    const { apiKey } = yield* yield* StripeCredentials;
+    const stripe = new Stripe(Redacted.value(apiKey));
+  }),
  delete: () => Effect.die("not implemented"),
});

The double-yield resolves credentials lazily, inside the handler — the provider layer builds without ever touching the auth chain.

Observe

 reconcile: Effect.fn(function* ({ news, output }) {
   const { apiKey } = yield* yield* StripeCredentials;
   const stripe = new Stripe(Redacted.value(apiKey));
+
+  let product = output?.productId
+    ? yield* Effect.tryPromise(() =>
+        stripe.products.retrieve(output.productId),
+      ).pipe(Effect.catchAll(() => Effect.succeed(undefined)))
+    : undefined;
 }),

If a previous run cached a productId, look up its live state — a failed retrieve just means "missing". output is a cache of identifiers, never proof the resource still exists.

:::tip[Use Effect.tryPromise, not Effect.promise] Effect.tryPromise surfaces SDK errors as Effect failures the lifecycle machinery can report cleanly; Effect.promise turns rejections into defects. :::

Ensure

   let product = output?.productId
     ? // ... observe ...
     : undefined;
+
+  if (!product) {
+    product = yield* Effect.tryPromise(() =>
+      stripe.products.create({
+        name: news.name,
+        description: news.description,
+        active: news.active,
+      }),
+    );
+  }
 }),

Create only when observation came back empty — greenfield creates and "the cached id points at nothing" recover through the same line.

Sync

   if (!product) {
     // ... ensure ...
   }
+
+  if (
+    product.name !== news.name ||
+    product.description !== (news.description ?? null) ||
+    product.active !== (news.active ?? true)
+  ) {
+    product = yield* Effect.tryPromise(() =>
+      stripe.products.update(product!.id, {
+        name: news.name,
+        description: news.description,
+        active: news.active,
+      }),
+    );
+  }
 }),

Compare each mutable field of the observed product against news and write only when something drifted — a no-drift deploy makes zero write calls.

Return fresh Attributes

   if (/* drifted */) {
     // ... sync ...
   }
+
+  return {
+    productId: product.id,
+    created: product.created,
+  };
 }),

The returned Attributes are persisted to state and become the next run's output.

Why this shape works

The body is idempotent by construction: Alchemy can retry it after a failed state write and it re-converges. Adoption needs no special case — retrieve finds the imported resource and the sync step rewrites whatever drifted from news.

:::note[Bindings] reconcile also receives bindings — data attached by upstream policies and event sources (IAM grants, queue subscriptions). Read them here if your resource is a binding target; see Binding. :::

Implement delete

delete runs when the resource is removed from code, replaced, or when alchemy destroy runs.

:::tip[Treat 404 as success] Like reconcile, delete must be idempotent. Alchemy may retry it if persistence fails between the cloud call succeeding and state being written. Catch "not found" errors and return Effect.void. :::

return StripeProduct.Provider.of({
  reconcile: Effect.fn(function* ({ news, output }) { /* ... */ }),
-  delete: () => Effect.die("not implemented"),
+  delete: Effect.fn(function* ({ output }) {
+    const { apiKey } = yield* yield* StripeCredentials;
+    const stripe = new Stripe(Redacted.value(apiKey));
+    yield* Effect.tryPromise(() =>
+      stripe.products.del(output.productId),
+    ).pipe(
+      Effect.catchAll((cause) =>
+        cause instanceof Error && cause.message.includes("No such product")
+          ? Effect.void
+          : Effect.fail(cause),
+      ),
+    );
+  }),
});

Implement diff (optional)

Some property changes can't be applied in place. For Stripe products the name is mutable but (hypothetically) the description is not — changing it requires recreating the product. Implement diff to tell Alchemy which kind of change to plan.

diff runs at plan time, before reconcile, and returns one of:

  • { action: "noop" } — change is trivial, skip reconciling
  • { action: "update", stables?: [...] } — apply in place
  • { action: "replace", deleteFirst?: boolean } — destroy and recreate
  • undefined / void — fall back to default (plan an in-place update)

:::note[diff runs during planning] Props may still be unresolved Outputs when diff is called — they're only fully resolved at apply time. Always guard with isResolved(news) and return undefined to defer to the default path when you can't decide. :::

+import { isResolved } from "alchemy/Diff";
+
// ...

return StripeProduct.Provider.of({
+  diff: Effect.fn(function* ({ news, olds }) {
+    if (!isResolved(news)) return undefined;
+    if (news.description !== olds.description) {
+      return { action: "replace" } as const;
+    }
+    return undefined;
+  }),
  reconcile: Effect.fn(function* ({ news, output }) { /* ... */ }),
  delete: Effect.fn(function* ({ output }) { /* ... */ }),
});

For attributes that are immutable across all updates (e.g. the Stripe productId, an ARN), declare them in stables at the top level:

return StripeProduct.Provider.of({
+  stables: ["productId"],
  diff: Effect.fn(function* ({ news, olds }) { /* ... */ }),
  // ...
});

Implement read (optional, for recovery and adoption)

The engine calls read whenever a resource has no prior state, both to recover from interrupted reconciles and to import pre-existing cloud resources into a fresh state store. Returning undefined tells Alchemy the resource doesn't exist and should be created.

return StripeProduct.Provider.of({
  stables: ["productId"],
  diff: Effect.fn(function* ({ news, olds }) { /* ... */ }),
+  read: Effect.fn(function* ({ output }) {
+    if (!output?.productId) return undefined;
+    const { apiKey } = yield* yield* StripeCredentials;
+    const stripe = new Stripe(Redacted.value(apiKey));
+    const product = yield* Effect.tryPromise(() =>
+      stripe.products.retrieve(output.productId),
+    ).pipe(
+      Effect.catchAll(() => Effect.succeed(undefined)),
+    );
+    if (!product) return undefined;
+    return { productId: product.id, created: product.created };
+  }),
  reconcile: Effect.fn(function* ({ news, output }) { /* ... */ }),
  delete: Effect.fn(function* ({ output }) { /* ... */ }),
});

This example only finds resources by an ID we previously saved — without a prior output it returns undefined, and that's a fine default.

Ownership-aware reads

import { Unowned } from "alchemy/AdoptPolicy";

read: Effect.fn(function* ({ id, olds }) {
  const live = yield* lookupByName(olds.name);
  if (!live) return undefined;
  const attrs = { productId: live.id, created: live.created };
  // Compare tags/owner against this stack/stage/id.
  return ownsResource(id, live.tags) ? attrs : Unowned(attrs);
}),

If your provider can find an existing resource from props alone (tags, deterministic naming), brand foreign ones with Unowned so the engine gates the takeover behind --adopt. The full three-way return semantics are in Providers › read.

Implement list

list enumerates every product in the account, returning the same Attributes shape as read — it powers account-wide operations like alchemy unsafe nuke and must paginate exhaustively (see Providers › list):

return StripeProduct.Provider.of({
  stables: ["productId"],
-  list: () => Effect.succeed([]),
+  list: Effect.fn(function* () {
+    const { apiKey } = yield* yield* StripeCredentials;
+    const stripe = new Stripe(Redacted.value(apiKey));
+    const products = yield* Effect.tryPromise(() =>
+      stripe.products.list({ limit: 100 }).autoPagingToArray({ limit: 10_000 }),
+    );
+    return products.map((p) => ({ productId: p.id, created: p.created }));
+  }),
  // ...
});

Resources with no enumeration API (singletons, existence-only resources) keep the Effect.succeed([]) stub instead.

Bundle into a providers() layer

Users expect the same one-line ergonomics as the built-ins (Cloudflare.providers(), AWS.providers()). Bundle the resource collection and every provider implementation into a single layer:

// src/stripe/Providers.ts
import * as Provider from "alchemy/Provider";
import * as Layer from "effect/Layer";
import { StripeProduct, StripeProductProvider } from "./Product.ts";

export class Providers extends Provider.ProviderCollection<Providers>()(
  "Stripe",
) {}

export const providers = () =>
  Layer.effect(
    Providers,
    Provider.collection([StripeProduct]),
  ).pipe(Layer.provide(StripeProductProvider()));

Layer.provide wires each resource provider to the collection privately. A provider with no credentials service is done here.

Ours isn't — the handlers' StripeCredentials requirement is still unmet. The Custom Auth Provider guide finishes this bundle by Layer.provideMerge-ing in the credential bridge and the auth-provider registration.

Re-export the public surface, dropping the redundant service prefix (the namespaced-export convention: callers write Stripe.Product):

// src/stripe/index.ts
export { StripeProduct as Product } from "./Product.ts";
export { Providers, providers } from "./Providers.ts";

Either way, users plug your providers in like any built-in — no API key in sight:

// alchemy.run.ts
import * as Alchemy from "alchemy";
import * as Effect from "effect/Effect";
import * as Stripe from "./src/stripe";

export default Alchemy.Stack(
  "MyApp",
  { providers: Stripe.providers(), state: Alchemy.localState() },
  Effect.gen(function* () {
    const pro = yield* Stripe.Product("Pro", {
      name: "Pro plan",
      description: "Everything in Free, plus...",
    });
    return { productId: pro.productId };
  }),
);

Users connect the provider with alchemy profile edit --add Stripe (CI reads env vars instead); an unconfigured deploy fails with that exact command to run. See Auth Providers.

To mix with another cloud, merge the layers:

import * as Layer from "effect/Layer";

providers: Layer.mergeAll(Cloudflare.providers(), Stripe.providers()),

Test the lifecycle

Full test.provider reference and patterns: Testing Providers.

Alchemy's test harness (@alchemy.run/test/Vitest or @alchemy.run/test/Bun) configures providers + state once at the top of the file, then exposes test.provider(name, (stack) => ...) for provider-level tests. Each test.provider body receives a fresh in-memory scratch stack with .deploy(effect) and .destroy() helpers.

:::note[Three things the harness gives you]

  • Test.make({ providers }) — wires your provider Layer into every test in the file so the resource type resolves.
  • stack.deploy(effect) — runs a sub-stack with an in-memory state store. The same logical ID across two stack.deploy calls triggers an update (or replace, depending on diff).
  • stack.destroy() — replans against an empty desired state to tear everything down. Call at the end of the test. :::

Create test/Product.test.ts:

// test/Product.test.ts
import * as Test from "@alchemy.run/test/Vitest";
import { expect } from "@effect/vitest";
import * as Effect from "effect/Effect";
import Stripe from "stripe";
import * as StripeProvider from "../src/stripe";

// Configure providers once per file. Credentials resolve through the
// Auth Provider — see /environments/custom-auth-provider.
const { test } = Test.make({ providers: StripeProvider.providers() });

const stripe = new Stripe(process.env.STRIPE_API_KEY!);

test.provider(
  "create, update, delete a product",
  (stack) => Effect.gen(function* () {
    // Create
    const created = yield* stack.deploy(
      Effect.gen(function* () {
        return yield* StripeProvider.Product("TestProduct", {
          name: "v1",
          description: "first version",
        });
      }),
    );
    expect(created.productId).toBeDefined();

    const live1 = yield* Effect.tryPromise(() =>
      stripe.products.retrieve(created.productId),
    );
    expect(live1.name).toBe("v1");

    // Update (in place)
    const updated = yield* stack.deploy(
      Effect.gen(function* () {
        return yield* StripeProvider.Product("TestProduct", {
          name: "v2",
          description: "first version",
        });
      }),
    );
    expect(updated.productId).toBe(created.productId);

    const live2 = yield* Effect.tryPromise(() =>
      stripe.products.retrieve(updated.productId),
    );
    expect(live2.name).toBe("v2");

    // Destroy
    yield* stack.destroy();
  }),
);

To verify replacement semantics, change a stables field (or the field your diff flags as replace) and assert that updated.productId !== created.productId.

Reference implementations

If you'd rather start from a real provider:

Where next

  • Local Providers — give your resource a second, dev-mode implementation that emulates it locally (ProviderLayer.dual, LocalProvider.make).
  • Custom Auth Provider — build the StripeCredentials service and login flow this guide consumes.
  • Auth Providers — how credentials resolve: lazy Effects, Profiles, auto-refresh.
  • Actions — deploy-time work that isn't a resource.
  • Testing Providers — the harness patterns behind test.provider.
  • Providers — the operation contract this guide implements.
  • Bindings — connecting Resources to Functions.