Zapier AI Actions (NLA) Deprecated: What It Actually Means

If you built a custom GPT, an agent, or an integration on Zapier AI Actions (the Natural Language Actions API) and just saw "no longer being developed and supported," you have one urgent question: is there a shutdown date? No, none has been published. But a different Zapier change does have a hard date, and mixing the two up is how teams panic early or break late.

Dated vs undated: two different Zapier changes#

Search results and forum threads blur two separate announcements into one "Zapier AI shutdown." They are not the same change, and only one of them has a date. Use this split before you plan anything.

If your exposure is Zap steps built on the Assistants API, start with the dated August 26, 2026 stop and the Responses API step mapping. This post covers the other track: AI Actions and NLA.

What Zapier actually said about AI Actions#

AI Actions is no longer being developed and supported. This documentation is available as a reference for current users. For a fully supported experience, explore Zapier MCP — it's built to offer even more flexibility and reliability for your workflows.

— Banner on Zapier's AI Actions documentation, accessed June 21, 2026

Notice what the banner does not say: no end date, no countdown, no "stops working on" sentence. Zapier's Assistants-step deprecation article names August 26, 2026 explicitly, so the missing date here is a real difference, not an oversight.

The infrastructure signal is louder than the banner. Both legacy entry points, nla.zapier.com/support/ and actions.zapier.com, now return HTTP 301 permanent redirects to mcp.zapier.com (live-fetch verified June 21, 2026). And in Zapier's community forum, AI Actions failure threads get resolved by pointing the user to Zapier MCP, not by fixing the AI Actions path.

Why teams get this migration wrong#

A dated shutdown creates its own urgency: someone puts August 26 on a calendar and work happens. An undated wind-down creates none. Nobody owns it, quarters pass, and the unsupported surface runs right up until it does not. Failure mode one: no date, so no owner.

Failure mode two is the quota math on the other side. Zapier's MCP usage docs state: "Each successful tool call through your MCP server consumes two tasks. This is a fixed rate." Failed calls do not consume tasks, and MCP draws from the same plan task allowance as your Zaps. On the Free plan's 100 tasks a month, that is roughly 50 tool calls.

How to migrate AI Actions to Zapier MCP#

1. 1

   Inventory every AI Actions dependency

   List each custom GPT, agent, and integration that calls AI Actions or NLA, including anything hitting actions.zapier.com or nla.zapier.com directly. Give each one a named owner.

2. 2

   Create the MCP server

   At mcp.zapier.com, click + New MCP Server, pick your AI client from the dropdown, and name the server. The server URL is https://mcp.zapier.com/api/v1/connect.

3. 3

   Connect and authenticate your client

   Supported Streamable-HTTP clients include ChatGPT, Claude (web, Desktop, and Code), Cursor, Microsoft Copilot Studio, VS Code, and Windsurf. For unlisted clients, use a connection token in an Authorization: Bearer header; tokens can be rotated.

4. 4

   Re-add tools and re-test every call

   Click + Add tool, type the app name, and select each action your old integration used. Then run real test calls from the exact client and surface you use in production.

5. 5

   Re-budget task usage before switching traffic

   Count expected tool calls per run, multiply by two tasks each, and compare against your plan allowance. A single integration that becomes a chatty agent can multiply task burn and hit plan limits mid-month.

The auth traps: what breaks in practice#

The most common migration failure is not tool logic, it is authentication. One community report from a Claude Code user shows the OAuth flow finishing with a success screen while the token exchange fails behind it, leaving the server "connected" but not authenticated:

HTTP 500: Invalid OAuth error response:
SyntaxError: JSON Parse error: Unrecognized token '<'

The token endpoint returned an HTML error page instead of JSON. Zapier also publishes a dedicated troubleshooting article for sign-ins that finish in the browser but never reach the AI client, naming Claude Code on WSL, terminal-based clients, and older Cursor and Windsurf sign-in bugs (an RFC 9728 OAuth issue) as triggers. Run these checks before calling the migration done:

Common mistakes with the AI Actions wind-down#

• Inventing a deadline. AI Actions has no published end date. Planning around an assumed one either burns an emergency sprint now or creates false calm later.
• Assuming August 26, 2026 covers AI Actions. That date applies only to the OpenAI Assistants API Zap steps, which have their own migration path.
• Skipping the task re-budget. Two tasks per successful tool call, from the same allowance as your Zaps, changes the economics of chatty agents.
• Testing in one client, shipping in another. The reports above show the same server passing in one client and failing with 401s in another.
• Leaving the migration ownerless. An undated change never becomes urgent on its own. Without a named owner and a scheduled re-check, it sits until it breaks.

That last mistake is a process gap, not a Zapier gap, and it repeats with every vendor. A vendor API sunset readiness process closes it: keep affected surfaces in a retained inventory with named owners and readiness states, and let a scheduled watcher resurface undated items so they cannot age out of memory.

Frequently asked questions#