Low code workflow automation tools, sorted by what actually writes the workflow file

It means the workflow definition is something a non-developer can read, write, and modify without writing a programming language. In SaaS orchestrators that surface is a node graph in a browser canvas. In RPA studios it is a sequence of recorded actions plus a properties panel. In observation-based AI agents like Mediar it is a small YAML or JSON spec the recorder writes, plus a UI that lets a person tweak step names, arguments, and error strategy in plain fields. The line between 'low code' and 'no code' is fuzzy in practice; 'low code' usually implies you can drop into a code node when you need to, while 'no code' implies you cannot. All three categories on this page allow both modes.

At the API boundary. Zapier, Make, Workato, and n8n connect cloud services that publish a REST or RPC API. SAP GUI is a Windows desktop client that talks to its server over the SAP RFC protocol; there is no public REST API for SAP GUI. Jack Henry, FIS, and Fiserv core banking systems are Windows desktop terminals that talk to a private host. Epic and Cerner expose FHIR for some patient data but not for the in-clinic desktop screens a medical assistant fills out. When one end of a workflow is a desktop app, the orchestrator-only categories cannot reach it; the standard pattern is to fire a webhook from the orchestrator into a desktop runtime that does the desktop work.

Two places. Maintenance and time to production. UiPath, Power Automate Desktop, and Automation Anywhere ship a developer studio where a user authors selectors and conditional branches by hand; the studio is a real IDE and the workflow files (xaml in UiPath, ppr in Power Automate) become real artifacts the team has to version, review, and patch. When a target app's UI changes (a relabeled button, a moved tab, a new modal), the selector breaks and a human edits the workflow. Documented industry numbers suggest the maintenance tax on a mature RPA estate runs 30 to 50 percent of total RPA cost. Time to production is the second drag: a typical UiPath project takes 12 to 16 weeks from kickoff to first deploy, mostly spent in studio.

The author is the observer, not the human. The user demonstrates the workflow once on a Windows machine, the recorder captures meaningful events through OS accessibility APIs (clicks, typed text, hotkeys, application switches), an LLM synthesis pass turns the event stream into a structured workflow with named steps and inputs, and the result is a small YAML spec the executor can replay. The 'low code' surface is no longer a builder canvas; it is the YAML and a small UI for editing step descriptions, retry counts, and error strategy. The user only touches the spec when they want to override what the recorder picked, which on simple flows is never.

Thirteen top-level fields on WorkflowSequence and nine fields on WorkflowStep, defined in crates/executor/src/models/workflow.rs lines 41 to 80 of the open-source Terminator runtime at github.com/mediar-ai/terminator. The full top-level set is steps, variables, selectors, inputs, stop_on_error, include_detailed_results, cron, start_from_step, end_at_step, follow_fallback, execute_jumps_at_end, scripts_base_path. Each step has id, tool_name, group_name, arguments, description, retry_count, timeout, on_error, fallback_id. The on_error enum has four values: Stop, Continue, Retry, Fallback. There is no graph, no canvas, no separate 'orchestration layer' file. One file is the workflow.

Conditionals and loops are step-level, not spec-level. A condition becomes a group with an early-exit step that checks state and returns; a loop is implemented by a step whose tool is execute_sequence with a different sequence. The spec stays small because branching is expressed in the step graph rather than in the wrapper. The trade-off is honest: deeply nested branching is harder to read in this shape than in a visual UiPath flow, and we recommend either a code workflow (TypeScript, same SDK) or a multi-file split for genuinely complex flows. Most automations a back-office team needs end up small enough that the YAML stays readable.

Match the tool to where the data lives. If both ends are SaaS apps with public APIs, a SaaS orchestrator (Zapier, Make, n8n) is the cheapest path. If one end is a SaaS app and the other is a Windows desktop with no API, the right composition is a SaaS orchestrator on the SaaS side and an RPA studio or AI agent on the desktop side. If both ends are Windows desktops with no APIs (a banking core to a CRM, a healthcare EHR to a billing system, an SAP GUI to an internal Access database), the SaaS orchestrator drops out and you are choosing between an RPA studio and an AI agent for the entire flow. The AI agent is faster to ship and self-heals on UI changes; the RPA studio gives you a heavier IDE and a longer audit trail by default.

Either. Some pilots replace a stalled UiPath project end to end (an LG-customer F&B chain moved off UiPath onto Mediar and reported 70 percent cost savings to their board). Others run Mediar inside an existing UiPath estate as the agent that handles the workflows where the UiPath selector kept breaking. Mediar's executor exposes webhook and cron triggers, so an existing Orchestrator schedule can call into it. The interesting question for a team with a Center of Excellence is usually not 'which platform' but 'which workflows belong on which platform,' and observation-based agents tend to win the long-tail of small variable workflows that are not worth a 12-week studio build.

Yes. The desktop agent's executor crate, the queue processor, the workflow parser, the MCP transport, and the WorkflowSequence model this page describes all live under github.com/mediar-ai/terminator. The hosted dashboard at app.mediar.ai/web (the no-code recorder, the workflow library, the run history, the auth and billing) is the proprietary surface. A team that wants to run the executor on its own infrastructure and skip the hosted dashboard can do that, and the workflow file format stays the same.

SaaS orchestrators charge per task or per operation. Zapier's tiers run a few cents per task at low volume; Workato is closer to enterprise iPaaS pricing on annual contracts. RPA studios charge per bot license and per developer seat; UiPath public pricing has historically run in the low five figures per attended bot per year and higher for unattended, with separate Studio and Orchestrator costs on top. Mediar charges $0.75 per minute of executor runtime with no per-seat license, and a $10,000 turn-key program fee that converts to credits and effectively prepays usage. A 30-second SAP post costs about 38 cents in Mediar regardless of how many users kicked off the workflow.