Substructure vs. n8n for AI agents

n8n is a visual workflow automation platform that has leaned hard into AI agents: an agent node on a canvas, a chat model and tools wired in as sub-nodes, and 500+ integrations to reach into. Ops people can build with it, the template library is enormous, and one platform can own all of your company's glue, not just the agents.

Substructure is a developer tool with a different center of gravity. An agent is a file in your repo, the engine runs it in Slack, authorizes its MCP servers, keeps its history, and streams its output. The file is a default agent. When you want control, your webhook receives every step of the loop as JSON and can build or override any of it, in any language. Your agent logic is a stateless function: state in, decision out, and the engine persists the rest. Full loop customization is the point, not a workaround.

The short answer

Pick n8n when the people building automations are not all developers, when the job is broad automation with some AI in it, and when 500+ drag-in integrations beat writing tools. Pick Substructure when the agent is the product: a teammate in Slack that developers define in git, wire to real systems over MCP, and customize down to individual loop steps when it matters.

What the agent is

In n8n the agent is a workflow JSON living in n8n's database, built on the canvas around the LangChain-based AI Agent node. Version control for workflows is real but gated to the Business tier, which starts at several hundred euros a month.

In Substructure the agent is substructure.toml in your repo. It reviews like code because it is in git next to your code, and subs apply deploys it. There is no canvas, which is a real limitation for non-developers and a real feature for teams who want agents in code review.

The Slack story

Building a Slack bot in n8n means creating your own Slack app, wiring OAuth scopes and event subscriptions to a webhook, then assembling the loop: trigger node, agent node, memory store, and a post-message node, with thread context persisted in Redis or Postgres yourself. It works, and the templates help. There is friction the docs are open about, like one webhook per Slack app, which makes test and production instances fight over the same app.

In Substructure, subs slack connect opens the consent screen and the bot is live. Threads are sessions, DMs and mentions route to agents per channel, and approval steps happen in Slack. That is the difference between an integration and a surface.

Durability

Table stakes, but the models differ. n8n's unit of execution is one run of the whole workflow. If the instance dies mid-run, the run is marked crashed, and retrying is a button or an error-workflow you build. Deliberate pauses survive restarts; a crash mid-agent-loop does not resume where it stopped. Queue mode adds resilience for scale, and requires Postgres plus Redis.

Substructure's engine saves every step before it runs and resumes on its own after a crash or a redeploy. One binary, SQLite, nothing to configure to make that true.

Models, MCP, and licensing

Credit where due: n8n has bring-your-own-key support for around 18 providers, and MCP in both directions, client and server. The licensing nuance is worth knowing: n8n is fair-code under the Sustainable Use License, free to self-host for internal use but not open source by n8n's own description. Substructure's engine is open source, and the self-hosted engine is the full product.

On price, n8n's cloud meters executions, and a chatbot burns one execution per message, so a busy channel is a busy meter. Substructure is one flat price with no per-message accounting, and you bring your own LLM key.

Where n8n is the right call

If your team includes non-developers who need to ship automations, nothing here competes with the canvas plus 10,000 templates. If the work is mostly integration glue, syncing systems, cron jobs, notification plumbing, with AI sprinkled in, n8n covers all of it in one tool. And the community is massive, so every question has an answer somewhere.

But if the agent is the thing, defined in git, living in Slack, customized at the loop level by developers, that is Substructure. The quick start takes about five minutes.