On 23 July 2021 I pushed the first commit to a repository called Awesome Integration. It was a curated list of system integration software, patterns and resources: 27 flat sections and about 260 links that I did not want to lose in my browser bookmarks.
Five years later it holds around 700 links across 41 categories, and about 500 distinct tools. I have been editing it almost every month since then, which means I have an unusual record of what our field decided to care about.
A curated list is not a market forecast. It is slower and more honest. A category only appears when several serious tools exist and practitioners actually argue about which one to pick. So the diff between the 2021 file and today's file is a reasonable picture of what really changed.
Here is what I see in that diff.
1. A new layer appeared in front of models
In 2021 there was no AI category in the list. Today there is one, and it is not a list of models. It is a list of gateways: LiteLLM, Portkey, Bifrost, Envoy AI Gateway, Kong AI Gateway, plus MCP-oriented ones such as agentgateway, Docker MCP Gateway, IBM ContextForge and MCPJungle.
What is interesting is how familiar the problems are. Routing across providers. Failover. Rate limiting. Cost tracking. Caching. Observability. Guardrails, which are the new name for policy enforcement. This is the API gateway problem restated for a different backend.
That is why I placed AI Gateway next to API Gateway, not in a separate "AI" area. If you have run an API gateway in production, you already understand most of the operational model. What is new is that the backend is non-deterministic, priced per token, and occasionally slower by an order of magnitude than anything you had behind a gateway before.
The Model Context Protocol belongs to the same shift. Strip away the naming and MCP is a connectivity standard: how a client discovers a tool, describes its contract and calls it. That is integration work, with a new audience.
2. Data movement became continuous
In 2021 the list had one section called ETL. Today it has four related sections: Data Integration, Change Data Capture, Stream Processing and Schema Registry.
That split was not a taxonomy exercise. It happened because the tools separated in practice. Debezium, PeerDB and Sequin turn a database log into an event stream. Flink, Arroyo, RisingWave, Bytewax and Redpanda Connect do stateful computation on the stream. Confluent Schema Registry, Apicurio and Karapace hold the contract that keeps producers and consumers compatible while the schema evolves.
The schema registry category is the one I find most telling. It is pure integration governance: versioning, compatibility rules, breaking-change detection. Five years ago that discipline lived mostly around REST and OpenAPI. Now it is applied to event payloads as well, because events crossed the boundary from "internal implementation detail" to "published interface".
3. Webhooks became infrastructure
Webhook Infrastructure is a small category, only Convoy, Hookdeck and Svix, and I almost did not create it. I did because the same conversation kept repeating: someone builds webhook delivery in an afternoon, then spends two quarters on retries, signing, replay, fan-out, endpoint management and delivery logs.
That is a classic integration story. The transport is trivial. Everything around reliability is not. The fact that it became a product category, rather than a utility class in every codebase, is a healthy sign.
4. The ESB did not die, it was repackaged
The most quietly interesting change is in the oldest section.
The ESB category still exists. Almost none of the 2021 entries survived under their original names. Red Hat Fuse became the Red Hat build of Apache Camel. Software AG webMethods became IBM webMethods. WSO2 Enterprise Integrator became WSO2 Integrator: MI. ArcESB became CData Arc. Dell Boomi became simply Boomi. Apache ServiceMix left the list entirely.
Read the marketing and you will conclude the ESB is dead. Read the list and you see something more accurate: the runtimes were rebuilt for containers and Kubernetes, the products were renamed, the owners changed, and the underlying job of routing, transformation and mediation stayed exactly where it was.
Categories in this field rarely disappear. They get new packaging and a new price model.
5. What did not change at all
Three things stayed constant across five years of edits.
Enterprise Integration Patterns are still the most valuable thing a new integration engineer can read. The book is from 2003. The vocabulary still describes tools released last month.
Message brokers are still the backbone. Kafka, RabbitMQ, NATS, Pulsar. The list around them grew, the position did not move.
And the hard part is still data mapping and error handling. Not the protocol, not the transport, not the deployment target. Every year I add faster ways to move a message between two systems. I add very few things that make it easier to agree what a customer record means in both of them.
On maintaining the list itself
One practical note for anyone running a similar project. A curated list decays the moment you stop looking at it. Since August 2021 a GitHub Action has checked every link on a schedule, and since 2023 the star counts refresh automatically. That automation is the only reason the list is still credible after five years and 478 commits.
The rest came from other people: 16 contributors, plus everyone who opened an issue to say that a tool was archived or that a description was wrong.
If you know someone who works with integrations, a repost helps this list reach them.
And if the list is useful to you, a star on GitHub helps others find it: https://github.com/stn1slv/awesome-integration
Which category do you think will be added in the next five years? Tell me in the comments, or open a pull request and add it yourself.
Top comments (0)