Run an AI agent as one HTTP step from Zapier, n8n or a script

Programmable — Thu, 05 Nov 2026 17:34:50 GMT

I build Programmable, a tool that turns one sentence into a sales agent. A question I kept getting was how to run that agent from somewhere else: a Zap, an n8n flow, a cron job. The answer ended up being boring on purpose. Every agent is one HTTP step.

Here is how that works, and the two problems it had to solve along the way.

The call

An agent has an address. You give it the company from whatever step came before and get people, checked emails and research back:

GET https://programmable.info/x/live?request=Find the head of sales and write an opener about their hiring&format=json&domains=acme.com
Authorization: Bearer YOUR_API_KEY

The first time a sentence runs, it's read once into a plan. Every call after that runs the stored plan, so the same sentence doesn't get re-interpreted differently on Tuesday than it did on Monday.

The response is one row per person (name, title, checked email, the opener) plus a summary with the credits used.

In Zapier, n8n and Make

It's the same request in each tool:

Start with a trigger that has the company in it: a new row in Google Sheets, a new company in HubSpot, a form answer.
Add an HTTP step (in Zapier, Webhooks by Zapier with event GET) pointing at the agent's link.
Map the company's website into domains, and add the Authorization: Bearer header.

One gotcha: Webhooks by Zapier is a Premium app, so it needs a paid Zapier plan. n8n and Google Sheets don't have that limit.

Problem one: Zapier waits about 30 seconds

A quick lookup finishes well inside that. A longer run (finding new companies, AI research on each one) doesn't. If the step just hung, Zapier would give up on it while the work was still running.

So the agent's link always answers within 25 seconds. If the results are ready, you get them. If not, you get this, and the run carries on:

{ "status": "started", "run_id": "...", "status_url": "..." }

The status_url shows progress, and once the run is done, the summary and rows.

Problem two: getting rows from a long run back into Zapier

Polling from inside a Zap is awkward. The cleaner pattern is a second zap:

Make a zap that starts with Webhooks by Zapier, Catch Hook, and copy its URL.
Paste it as your webhook in the app (Developers, Webhooks).
Rows arrive at the Catch Hook in batches as they finish (event sheet.rows), then a summary (sheet.completed).

Every batch is signed with X-Programmable-Signature (HMAC-SHA256) and carries an X-Programmable-Id that stays the same across retries, so you can drop duplicates with one filter step.

What a link can and can't change

A deliberate limit: the link only takes the trigger's details. That means company or domains, or a person (name plus company, or a LinkedIn URL), a lower limit, a campaign_name, and format=json. It can't change the columns, the AI questions, where results are sent or whether sends need approval. Those stay with the agent. A link pasted into a public Zap can't be made to spend more or send somewhere new.

The rest of the surface

POST https://programmable.info/api/sheet for the full API, with CSV by default, JSON with format=json, or NDJSON events with Accept: application/x-ndjson.
An MCP server at https://programmable.info/api/mcp for Claude, Cursor and other assistants, with the same key and limits.
Plain errors: 401 unknown key, 402 out of credits, 428 a sentence not yet confirmed, 429 over the hourly limit.

New accounts start with 500 credits included, no card. Full reference: https://programmable.info/developers

If you wire this into something I haven't thought of, I'd like to see it.