Relay.app is shutting down. Export before you shop for a replacement.

Relay announced the closure on July 16. Free users received 30 days and paying customers received 60. The company says existing paid workflows will continue to run without charge through September 14, with an additional 25,000 steps and 10,000 AI credits per month during the transition.

TechCrunch reported on August 17 that founder and CEO Jacob Bank and some Relay staff are joining Google’s Chrome team. Bank said Chrome is a place to help people work with AI, but no announced Chrome feature carries a Relay workspace forward. Treat the staff move and your migration as separate facts.

For former free users, the official access window is already over. Paying customers still have a useful overlap period: the old system can run while a replacement is rebuilt and checked. Spending that overlap on product demos would be a peculiar way to use it.

A full workspace export contains a folder for each Workflow, Sequence, MCP Server and Table. Relay says those folders can include a plain-language build prompt, workflow JSON, run history as CSV, a JSON file for each run and table data as CSV. That is enough to recover the shape of the work and many of the examples that proved it worked.

There are catches worth reading before the final week. A large workspace export can take up to 24 hours to generate. The emailed download link expires after 48 hours. Workflow screenshots are not included; they must be saved separately. A run-history CSV is capped at 100,000 rows per export.

The workflow JSON is written in Relay’s own structure. Relay’s documentation says other automation products will not accept it through their native importers without interpretation. The generated build prompt is meant to help an AI assistant reconstruct the flow elsewhere. That may save typing. It does not prove that branches, field mappings or error paths survived the move.

Credentials cannot come along, and they should not. Relay will delete stored app credentials and tokens when the account closes. Every Gmail, Slack, CRM, database and payment connection must be authorized again in the new service by someone who is allowed to do it. The boxes and arrows are usually not the part that stalls a migration.

First, request the full workspace export now. Download the archive before its 48-hour link expires, open it and confirm that the important workflows, prompts, runs and tables are actually present. A successful export notification is not the same as a checked archive.

Second, save screenshots of the flows that matter. Then write one short note for each: what starts it, what it changes, which account owns each connection, who notices when it fails and what the manual fallback is. Do not document every clever detail. Document enough that a coworker could tell whether the rebuild is wrong.

Third, rank by consequence rather than complexity. A six-step lead-enrichment flow can wait if nothing bad happens when it pauses. The two-step job that sends customer messages, moves money, updates an order or creates a compliance record goes first. Migrating the prettiest canvas first is how a team ends up admiring its new setup while a boring nightly job quietly stops.

Fourth, keep one known-good run and one ugly run for each critical flow. Use the clean case to check fields and outputs. Use the ugly case—a missing value, expired token, duplicate record or declined approval—to check whether the replacement fails in a way a person can see and recover from. Run the old and new versions side by side before switching the trigger.

Ivy Chen would put a name beside every connection before rebuilding anything. The person who built a flow may not be the person allowed to reconnect the shared inbox, CRM or finance account. If those permissions are discovered on cutover day, the migration becomes a support queue with a deadline.

Noah Park would start with one routine that ran successfully last week. Feed the replacement the exported build prompt, reconnect only the apps that routine needs and replay the same input. Compare the final record, message or file—not the new tool’s claim that the run completed. Then try the failure case before turning on the live trigger.

Ivy’s concern is organizational and Noah’s is mechanical. Both beat choosing a replacement from a feature grid. A new service is useful when the right person can connect it and an ordinary run ends where the old one did.

Relay deserves credit for exporting the logic as prompts and JSON, the history as CSV and per-run JSON, and table data as CSV. It also says plainly where those files fall short. That is a respectable wind-down. It should become the minimum you ask of the next product before giving it a business-critical job.

Ask for a sample export during the trial. Check whether it includes triggers, branches, field mappings, prompts, run errors and the data stored only inside the product. Ask whether another tool can read the format without a reconstruction project. Find out who owns the connected accounts and how the job runs when the AI feature, vendor or billing account is unavailable.

This does not mean avoiding AI automation or small vendors. Relay’s users got real work done, and a large incumbent can retire a product too. It means the exit deserves one hour during setup instead of one frantic week at shutdown.

The saved hours were real. So is the migration week. Get the files before using any more of that week to shop.