I gave my coding agent an n8n/Zapier-style automation platform over MCP

Most MCP servers get used one tool call at a time, inside a chat. When the chat closes, the work stops.

I wanted something else: tell my coding agent "every weekday at 9am, do X and email me" and have it build a real workflow that keeps running on its own. So I built WebArm24. It's a hosted automation platform in the same space as n8n, Zapier and Make, except it's designed to be driven by an agent over MCP instead of dragged together on a canvas.

What it looks like

The prompt in Claude Code was:

Every weekday at 9am, fetch USD rates for EUR, GBP and JPY, have AI summarize them and email me.

The agent drafted the pipeline, created it, ran it once to check, and the email arrived 15.5 seconds later. After that it runs on its cron schedule with no chat open.

Connect it in one line

claude mcp add --transport http webarm24 https://webarm24.online/mcp \ --header "Authorization: Bearer wak_..." 

Create the wak_... key with the MCP button at webarm24.online. New accounts get free credits. It's a plain streamable-HTTP endpoint, so Cursor, Windsurf, VS Code and other MCP clients work too.

Discovery (initialize, tools/list) doesn't need a key, so you can inspect all the tools before signing up.

What the agent can do

The server exposes about 60 tools, grouped like this:

• Pipelines: pipeline_create, pipeline_update, pipeline_run, pipeline_run_wait, pipeline_test_step, pipeline_debug_chat, plus groups, version history and restore of deleted pipelines.
• AI builder: pipeline_builder_chat drafts a pipeline from a plain-English description.
• Triggers: cron, a one-shot run_at time, an inbound webhook URL, or a manual run.
• Storage: persistent per-user JSON/text stores and files that pipelines can read and write.
• Direct tools: email, JSON analysis, and a browser for sites that have no API.

A pipeline is a list of steps, and each step can use the earlier ones' output through templates like {{prev_output}}, {{step_2_output}} or {{step_1.json.items}}. The step types are:

MCP servers as building blocks

The mcp step is my favorite part. Your pipeline can call other people's MCP servers on a schedule, in parallel, and feed one server's output into another's input.

Every Monday, this pipeline:

1. Calls TidyData MCP public_holidays three times in parallel (US, UK, DE)
2. Merges the results into 33 rows sorted by date in a JS step
3. Publishes them as a shared table through rows.page MCP push
4. Emails the team the link

The whole run takes 5.3 seconds, and the step config is just:

{ "type": "mcp", "mcp_url": "https://rows.page/mcp", "mcp_tool": "push", "mcp_arguments": "{\"rows\": {{prev_output}}, \"title\": \"2027 public holidays\"}" } 

For servers that need auth, headers can come from pipeline variables: "Authorization": "Bearer {{TOKEN}}".

What I learned building it

Agents are better at editing workflows than at drawing them. Making failures readable was the hard part. Per-step outputs, a "test one step" tool and a debug chat let the agent find and fix its own mistakes instead of guessing.

Undo stops being optional once agents edit things unattended. Every change, whether it comes from the UI, the AI builder or an agent over MCP, is saved as a revision tagged with its source. Deleted pipelines can be restored too.

Use LLMs at build time and keep run time deterministic. The mcp step has an agent mode where an LLM chains a server's tools itself. For scheduled jobs, though, fixed call_tool steps were far more predictable.

MCP servers: please return structuredContent. A friendly sentence plus JSON in one text block reads well in chat but is painful to use downstream. Clear inputSchema descriptions with examples (like "year (e.g. 2027)") also save agents a lot of wrong guesses.

Pricing

WebArm24 uses pay-as-you-go credits with no subscription. Creating and editing pipelines is free. Runs, AI builder calls and direct browser or email tool calls cost credits, and new accounts start with free ones.

Try it

• Website: webarm24.online
• MCP endpoint: https://webarm24.online/mcp
• Install docs and demos: github.com/leovegas/webarm24-mcp
• Also on the official MCP registry as online.webarm24/webarm24

I'd love feedback, especially if you run n8n or Zapier in production: what would stop you from letting an agent maintain your workflows?