notifkit/README.md

notifkit

You shouldn't have to build a notification system.

Self-hosted notification infrastructure for product notifications. One API call handles email, SMS, push, and webhooks, with preferences, quiet hours, retries, fallback, scheduling, workflows, and delivery logs built in.

Documentation · Quickstart · Examples · notifkit.dev

Sending one notification is easy

await sendEmail({
  to: user.email,
  subject: "Order Shipped",
  ...
});

Sending them reliably is the hard part. Users opt out. People are asleep. Push tokens die. Providers throw 503s. Channels fail and need fallback. Somewhere along the way you need timezone-aware quiet hours, future scheduling, deduplication, multi-channel templates, delivery logs, unsubscribe handling, multi-step workflows, and a dead-letter queue nobody wants to maintain.

notifkit is that machinery. Your app makes one typed call, and notifkit decides who gets the notification, which channel to use, when to send it, whether the user is allowed to receive it, and what to do when delivery fails.

import { notifkit } from "notifkit";

await notifkit.notify({
  user: "usr_123",
  template: "order-shipped",
  channels: ["push", "email"],
  fallback: true,
});

That call tries push, then email if push fails. Preferences, consent, quiet hours, retries, deduplication, throttling, template rendering, and delivery tracking all happen behind it.

How it runs

notifkit is an orchestration engine and a typed SDK.

flowchart TD
    App["Your Application / AI Agent<br/>Typed SDK · REST API · MCP Server"]

    App -->|"HTTP POST /v1/notify"| API

    API["Notifkit API Server<br/>Schema Validation · Auth · Multi-Tenancy<br/>Idempotency Gate · Priority Queue Ingestion"]

    API --> PG
    API --> REDIS

    PG[("PostgreSQL — Storage<br/>Users · Preferences<br/>Templates · Workflows<br/>Delivery Logs · DLQ")]
    REDIS[("Redis — Streams / ZSET<br/>Priority Queues<br/>Scheduled Sends<br/>Sliding Rate Limits")]

    subgraph WORKERS["Background Workers Pipeline"]
        direction LR
        ENRICH["Enricher<br/>(Resolve)"] --> ENGINE["Engine<br/>(Quiet Hours)"] --> DELIVER["Delivery<br/>(Rate Limits / CB)"]
        ENGINE --> SCHED["Scheduler<br/>(sendAt / QH)"]
        SCHED --> DELIVER
    end

    REDIS -->|"consume"| ENRICH
    ENRICH -.->|"read / write state"| PG
    DELIVER -.->|"delivery logs"| PG
    DELIVER -->|"Dispatch"| PROVIDERS

    PROVIDERS["Provider Transports<br/>Email: Resend · Push: Firebase (FCM) · SMS: Twilio<br/>Chat: Slack, Telegram, Discord, WhatsApp<br/>Webhooks: Custom HTTP"]

    classDef entry stroke:#6366f1,stroke-width:2px
    classDef store stroke:#0ea5e9,stroke-width:2px
    classDef work stroke:#22c55e,stroke-width:2px
    class App,API,PROVIDERS entry
    class PG,REDIS store
    class ENRICH,ENGINE,DELIVER,SCHED work

NotifkitServer runs the HTTP REST API router (/v1/notify, /health, /metrics) and the background worker pipelines: enricher, decision engine, scheduler, and delivery. NotifkitClient is the lightweight client your application uses to trigger notifications, sync templates, and manage users over HTTP.

Topologies

In a single process, the API and all workers run in the same Node.js process (services: ["all"]), which works for small and medium apps, side projects, and staging. Distributed, you run stateless API servers (services: ["api"]) behind a load balancer and scale worker pools (services: ["enricher", "engine", "delivery", "scheduler"]) horizontally across Redis Streams consumer groups.

Quickstart (needs 5 mins)

1. Install

npm install notifkit @notifkit/provider-resend
npm install -D tsx @testcontainers/postgresql @testcontainers/redis

The two @testcontainers/* packages are what notifkit uses to start throwaway PostgreSQL and Redis containers in development. They are imported lazily, only when the server is given neither a databaseUrl/redisUrl option nor a DATABASE_URL/REDIS_URL environment variable, so devDependencies is the right place for them.

Warning

Those containers are for local development only. They are thrown away when the process exits, taking every user, template, delivery log, and queued notification with them. Before you deploy, point notifkit at a real PostgreSQL and Redis — set DATABASE_URL and REDIS_URL (or pass databaseUrl and redisUrl) and run with NODE_ENV=production, which refuses to start a container and fails loudly if either is missing.

2. Run the engine

server.ts starts the API and the worker pipelines. In development it auto-starts those containers, so Docker is the only prerequisite.

// server.ts
import { NotifkitServer } from "notifkit";
import { ResendTransport } from "@notifkit/provider-resend";

const server = new NotifkitServer({
  services: ["all"], // API + enricher + engine + scheduler + delivery
  port: 3000,
  providers: [
    new ResendTransport({
      apiKey: process.env.RESEND_API_KEY!,
      from: "notifications@yourdomain.com",
    }),
  ],
});

await server.start();
console.log("notifkit listening on http://localhost:3000");

from is required on ResendTransport — it is the sender for any template that does not name its own, and it has to be an address on a domain you have verified in Resend. A template can override it with its own from, so one transport can serve both no-reply@ receipts and marketing@ campaigns.

ADMIN_API_KEY is the root credential. It is read from the environment, it is what mints project API keys in the next step, and without it the project-management routes answer 403. Any string works locally:

ADMIN_API_KEY=supersecretkey RESEND_API_KEY=re_xxx npx tsx server.ts

Warning

supersecretkey is a local placeholder. In production this one value can mint keys for every project, so use a long random string kept in your secret store — openssl rand -hex 32 is enough.

3. Create a project and its API key

Every /v1/* route requires a project API key, and only the admin credential can mint one, so this is the single bootstrap step between a running server and your first notification:

ADMIN_API_KEY=supersecretkey npx notifkit-create-project "my-app"

Project "my-app" created. Save the API key now — it is not recoverable.

NOTIFKIT_PROJECT_ID=1ce67fa1-b4a9-4985-8046-ef6018912b2a
NOTIFKIT_API_KEY=nk_live_f57c57b76d795cef89e2dbf6b6f352a36…

The server stores only a SHA-256 hash of the key, so the nk_live_… value is printed once and never again — put it in your app's .env now. Point the script at another host with NOTIFKIT_URL, and mint further keys later with POST /v1/projects/:id/keys (role: "read_only" there gets you a key that can read but not send).

4. Dispatch your first notification

client.ts is your application code. It talks to the server over HTTP: register a template, register a user, and send.

// client.ts
import { NotifkitClient } from "notifkit";

const notifkit = new NotifkitClient({
  baseUrl: "http://localhost:3000",
  apiKey: process.env.NOTIFKIT_API_KEY!,
});

// 1. Register a template
await notifkit.syncTemplates({
  templates: [
    {
      id: "order-shipped",
      channel: "email",
      content: { subject: "Order #{{orderId}} Shipped", text: "Your order is on the way!" },
    },
  ],
});

// 2. Register a user (supports id + contacts array or object)
await notifkit.addUser("usr_123", [{ channel: "email", target: "alex@acme.com" }]);

// 3. Dispatch
await notifkit.notify({
  user: "usr_123",
  template: "order-shipped",
  channels: ["email"],
  data: { orderId: "9481" },
});

With the server still running in the first terminal, run the client in a second one:

NOTIFKIT_API_KEY=nk_live_xxx npx tsx client.ts

5. Or call the REST API directly

The Node.js SDK is optional. notifkit exposes a standard HTTP REST API, so you can dispatch notifications and manage resources from any language (cURL, Python, Go, and so on). The same project API key goes in the Authorization header (an x-api-key header works too):

curl -X POST http://localhost:3000/v1/notify \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $NOTIFKIT_API_KEY" \
  -d '{
    "user": "usr_123",
    "template": "order-shipped",
    "channels": ["email"],
    "data": { "orderId": "9481" }
  }'

Development vs. production

Locally, Docker is the only prerequisite: in development notifkit starts throwaway PostgreSQL and Redis containers for you. In production you need Node 22+, PostgreSQL, and Redis, and you run migrations by pointing drizzle-kit at node_modules/notifkit/drizzle.

Running in production

notifkit runs in production at my own company, delivering 100K+ notifications a day across email, push, and OTPs. I built it because I needed it and didn't want to spend months rebuilding distributed notification plumbing or pay a SaaS per alert. It runs on your servers, with your provider accounts and your data.

Reliability and failure testing

Every component of the pipeline is tested against failure:

flowchart LR
    S1["Redis Streams"] -->|"Kill Worker (SIGKILL)"| M1["Auto-Claim and Replay"] --> O1["Zero Lost Messages"]
    S2["Connection Loss"] -->|"Drop DB / Redis"| M2["Auto-Reconnect / Retry"] --> O2["In-Flight State Intact"]
    S3["10k+ Messages"] -->|"Burst"| M3["Concurrency and Limits"] --> O3["Flat Memory, No Leaks"]

    classDef fault stroke:#ef4444,stroke-width:2px
    classDef guard stroke:#6366f1,stroke-width:2px
    classDef result stroke:#22c55e,stroke-width:2px
    class S1,S2,S3 fault
    class M1,M2,M3 guard
    class O1,O2,O3 result

• Crash testing (tests/chaos/crash.test.ts): background worker processes are killed with SIGKILL during high-throughput message streaming. Consumer group Pending Entries List (PEL) re-claims mean no messages are lost and another worker takes over.
• Infrastructure recovery (tests/chaos/recovery.test.ts): PostgreSQL and Redis connections are severed and restored under live traffic, verifying client reconnection, worker backpressure, and durable state resumption.
• Load testing (tests/chaos/load.test.ts): bursts of 10,000+ notifications across parallel worker pools, checking queue drain speed, sliding-window rate limiters, and memory use over time.
• Race conditions and concurrency (tests/race-conditions.test.ts, tests/idempotency.test.ts): concurrent duplicate dispatches, overlapping quiet-hour boundary evaluations, atomic user updates, and 24-hour idempotency key deduplication.
• Real containers, no mocks: unit, integration, and chaos suites all run against real PostgreSQL and Redis containers via Testcontainers.

What you get

The problem you don't want to build	How notifkit solves it
"Should this user receive it?"	User preferences, topic opt-outs, and consent gates
"Is this a bad time to send?"	Timezone-aware quiet hours that defer non-urgent sends
"What if push fails?"	Ordered multi-channel fallback (push, then email, then sms)
"What if my worker crashes?"	Redis Streams consumer groups, retries, and durable idempotency
"What if an event fires twice?"	24-hour deduplication via idempotency keys
"Can I send this later?"	Priority scheduling with sendAt and cancellation before dispatch
"Can I send this 3 days after signup?"	Stateful multi-step workflows with wait and waitForEvent
"How do I know what happened?"	Queryable delivery logs, Prometheus metrics, and campaign reporting
"What happens when a provider goes down?"	Circuit breakers, exponential backoff, and DLQ replay
"What about bounces and spam complaints?"	RFC 8058 one-click unsubscribe and automatic hard-bounce suppression
"What if I don't want another SaaS holding my data?"	Fully self-hosted on your PostgreSQL and Redis

You decide what to say. notifkit gets it there.

Scope

notifkit is the durable notification layer that runs inside your own stack. It is not a marketing automation suite, and it does not replace Customer.io, OneSignal, or SendGrid. You bring your own provider accounts and pay them directly.

First-party providers cover Resend, Firebase Cloud Messaging, Slack, Twilio, Telegram, Discord, and WhatsApp. Anything else is a Transport class with a send() method.

Agent-operable

https://github.com/user-attachments/assets/4dff98bb-37d3-44b4-bf46-9607c1cd89b5

An AI agent can operate notifkit directly. Connect the notifkit MCP server (@notifkit/mcp) to Claude Code, Cursor, Claude Desktop, Gemini, or any MCP-compatible agent:

npx -y @notifkit/mcp

Ask your agent

You: Why didn't usr_9182 receive their password reset?

Agent: The notification was suppressed because usr_9182's email
       address has a hard-bounce suppression from yesterday.

Your application and your AI agents use the same notification infrastructure. Through MCP an agent can:

• Send one-off notifications or campaigns to users, lists, and segments (send_notification, send_campaign)
• Diagnose delivery issues by inspecting message histories, provider responses, and quiet hours (get_delivery_logs, get_notification)
• Schedule future sends and cancel pending notifications (list_scheduled, cancel_notification)
• Check delivery, open, click, bounce, and complaint metrics (list_campaigns, get_campaign_stats)
• List, preview, and update templates with sample data (list_templates, preview_template, upsert_template)
• Look up users, contacts, preferences, and segment membership (list_users, get_user_preferences, update_user_preferences)
• Trigger workflows and inspect workflow runs (create_workflow, trigger_workflow, get_workflow_run)
• Manage bounce suppressions, check system queues, and replay dead-letter messages (list_suppressions, get_dead_letters, replay_dead_letter)

The same task, with and without an agent

Without an agent	With the notifkit MCP server
Query the database for contact info, open Twilio or Resend or write a throwaway script, format the payload, check the user's timezone by hand, send it, and hope it delivered.	You: "Send an urgent update to alex@acme.com that his package was lost in transit and support is rushing a replacement. Text him if push doesn't deliver." Agent: Looks up alex@acme.com, renders the template, dispatches push with SMS fallback, bypasses quiet hours because the send is urgent, tracks delivery status, and confirms it reached his phone.

MCP documentation

AI-assisted migration

Already have notification code scattered across your application? Point your coding agent at:

https://notifkit.dev/llms-full.txt

It can read notifkit's API from there, find ad-hoc notification code in your repository, and refactor it into notifkit calls.

Feature matrix

Channels	email, sms, push, webhook, telegram, discord, whatsapp, slack
Targeting	A user, a list of users, a segment, or a topic
Priorities	low, normal, high, critical, on separate stream lanes
Scheduling	Future sends with sendAt, quiet-hours deferral, cancellation
Preferences	Per-user channel and topic opt-outs, quiet hours, contact-level overrides
Workflows	Multi-step sequences with wait, waitForEvent, and notify steps
Reliability	Redis Streams, 24h idempotency, retries, DLQ, provider circuit breakers
Templates	{{var}} interpolation with destination-aware escaping
AI	Optional LLM augmentation before render via the Vercel AI SDK
Multi-tenancy	Projects with isolated keys, data, and rate limits
Consent	RFC 8058 one-click unsubscribe; complaints and hard bounces suppress automatically
Reporting	Campaign labels with delivery and engagement totals
Agent operation	MCP server for sending, triage, campaigns, templates, workflows, and system operations
Observability	Prometheus /metrics, /health, /live, /ready, and queryable delivery logs

Providers

Bring your own provider accounts. First-party packages:

• @notifkit/provider-resend: transactional email via Resend
• @notifkit/provider-fcm: push notifications via Firebase Cloud Messaging
• @notifkit/provider-slack: Slack messages via Incoming Webhooks or the Web API
• @notifkit/provider-twilio: SMS via Twilio, with signature-verified delivery status callbacks
• @notifkit/provider-telegram: messages via a Telegram bot
• @notifkit/provider-discord: messages via a Discord webhook
• @notifkit/provider-whatsapp: messages via Meta's WhatsApp Cloud API

For anything else, implement a Transport:

class MyTransport implements Transport {
  async send(message) {
    // Send through SES, Postmark, APNs,
    // SendGrid, a custom webhook, or anything else.
  }
}

The keys, the billing, and the deliverability stay yours.

Documentation

Everything lives at notifkit.dev/docs.

Quickstart	Install to first delivered notification
How it works	Core concepts and the notification pipeline
Channels & fallback	Multicast, ordered fallback, and custom transports
Preferences & quiet hours	Preference, consent, and timing rules
Templates & AI	Interpolation, escaping, and per-channel content
Segments & scheduling	Fan-out, priority lanes, sendAt, and idempotency
Workflows	Multi-step sequences, recurring sends, and digests
Examples	Runnable projects
Architecture	Streams, delivery guarantees, topologies, and data model
Deployment	Docker, Compose, and production topologies
Operations	Health, metrics, DLQ, key rotation, and shutdown
Reference	API, payloads, and SDK methods
MCP server	Operate notifkit from an AI agent

Why build this?

Notification infrastructure looks simple until you're responsible for it. Queues, retries, provider adapters, preference systems, quiet-hour logic, workflows, suppression handling, and operational tooling take months to build well. notifkit is what I built instead, and it's what I run.

Contributing

Issues and pull requests are welcome. Stars help other people find the project.

npm install
npm run build
npm test

The test suite starts its own PostgreSQL and Redis containers, so Docker is the only thing you need running.

Contact

Questions, bugs, or ideas: contact.devkitshq@gmail.com, or open an issue.

License

MIT. Do what you like with it, including commercially. See LICENSE.