Best Relay.app Alternatives for Small Business (2026): Where to Move Before Shutdown

Export Relay.app before choosing its replacement. For a small business, shortlist Zapier for managed app connections and native approval/data-collection steps, Make for visual routing and data transformation, n8n for technical control and AI-tool approvals, and Activepieces for an integrated flows/agents/tables workspace with run-based billing. None is a proven one-click replacement for your Relay workspace.

Deadline verified September 14, 2026: Relay’s shutdown notice says paid accounts and associated data will be permanently deleted after September 14, 2026 at 23:59 Pacific Time. That is September 15 at 06:59 UTC. Free-account deletion was scheduled after August 15. If you are reading after the paid deadline, work from already-downloaded exports; do not assume Relay remains available.

The immediate decision is what to preserve. The buying decision comes next: which platform can reproduce your actual triggers, approvals, data and recovery behavior with an operating burden your team can handle?

Original AI-generated editorial diagram; not a product screenshot or evidence of a completed migration.

Export Relay.app now: the deadline and exact checklist

Request the full workspace export immediately, but do not rely on it as your only deadline-day backup. Relay’s export documentation says generation can take up to 24 hours, and the emailed download is available for 48 hours. Those periods are not a published extension of account access or the deletion deadline. Contact Relay support about an unfinished export; do not assume an exception will be granted.

Work through this preservation checklist:

1. Request every workspace export. Use Export workspace data from the dashboard and follow the settings flow. Download the finished ZIP, unzip it and verify that it opens. Save a second access-controlled copy outside Relay.
2. Capture critical workflows individually now. Use the workflow overflow menu’s Export workflow and Copy build prompt. The full ZIP is documented to contain a folder for each Workflow, Sequence, MCP Server and Table, with applicable structural JSON, build-prompt Markdown, run-history CSV, per-run JSON and Table CSV.
3. Inventory Sequences separately. Record enrollment, current position, waits, stop conditions, scheduled messages and business owner. An exported definition does not prove that in-flight enrollment or timers can resume elsewhere. Record unresolved work explicitly.
4. Inventory MCP servers. Record tool names, input/output schemas, allowed actions, connected clients and ownership. Rebuild the tools and reconnect clients against new endpoints; exporting a server definition does not keep the old endpoint alive.
5. Preserve prompts and execution evidence. Save instructions, expected output schemas, model settings where available, representative successes, failures and pending approvals. Relay’s individual run JSON can include completed-step and AI-credit metrics, selected inputs/outputs, errors and AI chat history when available.
6. Check history completeness. Individual workflow run-list CSV exports are capped at 100,000 rows per export and are truncated beyond that. Compare counts and earliest/latest dates; retain important per-run JSON separately. Do not describe one capped CSV as a complete archive.
7. Check Tables independently. Compare exported row counts, field names, stable IDs and sample values. Preserve linked-record meanings, dates, null values and attachment references. A CSV reference to a file is not necessarily a copy of the file itself.
8. Save visual evidence separately. Save workflow as screenshot captures the overall workflow and individual steps. Relay explicitly says these screenshots are not included in the workspace export.

Treat the archive as sensitive business data. Redact customer information before using exported prompts with a general AI builder; never paste credentials into a build prompt. Do not manually delete the workspace while you still need access.

Relay JSON is migration evidence, not a destination import format. Relay explicitly says its JSON will not work through other tools’ native importers. An AI-assisted rebuild still needs human review, fresh connections and execution tests. Neither prompts nor historical run files establish that the replacement will behave identically.

Best Relay.app alternatives by small-business scenario

Choose the platform that passes your hardest workflow, not the one with the longest feature list. These are editorial fit judgments based on current vendor documentation, not hands-on migration results.

Zapier: shortlist for managed approvals and app coverage

Zapier is a strong first candidate for an operations-led migration. Its Human in the Loop integration documents Request Approval and Collect Data, including reviewer, timeout and reminder settings. Request Approval can let a reviewer edit content before the workflow continues. That is more relevant to a Relay approval workflow than a generic claim that a product “supports humans.”

The current pricing reference includes Tables and Forms, paid multi-step Zaps, Paths, Code, AI by Zapier and MCP. Professional starts at US$19.99/month billed annually for 750 tasks, or US$29.99 billed monthly. Team begins at 2,000 tasks, at US$69/month billed annually or US$103.50 monthly, and includes 25 users, shared connections and SAML SSO. Taxes and separate add-ons are additional considerations.

Avoid choosing solely on the entry price. Successful actions consume tasks, some AI/code/programmatic actions have variable rates, and standalone Agents has separate pricing. Broad app coverage does not establish attachment, custom-field or approval parity. This is a managed-service choice, not a full self-hosted runtime replacement.

Make: shortlist for visual branching and transformations

Make fits a team with someone comfortable maintaining a detailed scenario canvas. Routers, filters, iterators and data stores offer building blocks for multi-branch work. Make Code supports JavaScript and Python on paid plans; it is not accurate to describe Make as unable to run code.

Make MCP server can expose active, on-demand scenarios with defined inputs/outputs. Scenario-run scopes are available across plans; management scopes require a paid plan. Its documentation directs users who need more granular scenario selection toward MCP toolboxes. An MCP connection needs its own permission review.

The limitation is reconstruction effort. Plan a persistent approval record, notification, authenticated response and resume path where required; do not assume Relay’s approval experience transfers automatically. Likewise, an on-premise connector agent is not the same as self-hosting the whole platform.

Make uses credits, generally one per ordinary non-AI operation, with exceptions for AI and code. Its base-price page could not be retrieved reliably for this review, so we are not repeating an unverified entry price. The credit rules and published pricing adjustments were accessible: extra credits carry a 25% premium relative to included credits. Confirm the selected tier at Make pricing.

n8n: shortlist for technical control and AI-tool review

n8n is the stronger shortlist when a technical owner needs to control custom logic and deployment. Its human-in-the-loop tool review pauses an AI Agent before selected tools execute, shows the proposed parameters and lets a reviewer approve or deny. That is selective tool oversight, not merely a final-output review.

Data tables support CSV import/export and workflow access. Current documentation describes a default 200 MiB total instance limit, project-level access and failure behavior when storage is full. This is useful internal workflow storage, not a promise that every Relay table relationship survives import.

Do not self-host without an operator. Someone must own updates, backups, restoration, access control and incident response. Cloud avoids much of the infrastructure work. The pricing page separates cloud Starter/Pro, self-hosted Business and custom Enterprise features; SSO, environments and extended retention are plan-sensitive. Numeric subscription prices did not appear in the retrieved page, so no current price is asserted here.

Paid-plan quotas use production executions, not one charge for every canvas node. That makes long workflows worth evaluating, but it does not remove model-provider bills, infrastructure costs or capacity limits.

Activepieces: shortlist for an integrated workspace with run billing

Activepieces deserves a proof-of-concept when you want flows, agents, tables and MCP together. Its current pricing page lists Free with 100 credits per day, Plus with 10,000 credits/month at $16/month billed yearly, and Team with 50,000 credits/month at $166/month billed yearly. The page displays dollar prices; confirm invoice currency and tax at checkout. Plus includes up to five users and bring-your-own AI keys; Team includes 25 users, SSO and standard roles.

The current page says one credit covers a whole run, however many steps, and advertises cloud or self-hosted deployment at the same plan price. Do not use older “pay per active flow” comparisons as current pricing evidence. Also do not read a run allowance as unlimited model usage; verify agent and AI billing for your selected configuration.

Two limitations deserve special attention. The Approval piece is labeled Legacy and offers Wait for Approval/Create Approval Links. Verify its suitability for a new deployment rather than assuming a modern Relay-equivalent inbox. Separately, agent-tool documentation identifies audit gaps for MCP tools, code and email; “has audit logs” does not mean every agent action is covered.

Relay feature-to-replacement matrix

Map behavior and state, not just feature names. “Rebuild” below means implementation and testing are required; it is not a claim of native Relay compatibility.

Use the products’ actual review flows to compare who can approve, whether edits are allowed, whether rejection stops the write, and what happens when the reviewer is unavailable. Approval to publish a workflow is a different governance feature from approval to execute a business action.

History is also not portable execution state. Make’s history exposes credits, outputs and CSV logs, with retention depending on plan. Its automatic recovery applies to specified errors and configured incomplete executions, not every business failure. Zapier documents paid Autoreplay. Activepieces Runs exposes step inputs/outputs. Test each system’s error path before treating a green run as complete business evidence.

Compare billing with three normalized workloads

Normalize the business outcome, then keep each vendor’s unit separate. The examples below are design estimates, not observed bills, migration benchmarks or exchange rates. Assume a 30-day month, no retries and no unexpected fan-out unless specified. External app subscriptions, AI-provider charges, storage, hosting and tax are outside the figures.

Relay columns use explicitly hypothetical export counters, not an asserted Relay charging rule. Replace them with the completed-step and AI-credit metrics in your exported runs. There is no new Relay subscription decision: the shutdown notice says wind-down usage will not be charged.

Example 1: 1,000 leads, three ordinary app actions each

A simple lead workflow exposes the difference between action and run billing. Each lead triggers a CRM upsert, spreadsheet append and internal notification. Assume one Make trigger operation and three ordinary one-credit modules; three standard chargeable Zapier actions; one top-level n8n execution; one Activepieces flow run.

Polling, lookup helpers, duplicate checks or separate callback runs change these totals. Zapier’s retrieved standard tiers would put 3,000 tasks above its 2,000 tier; its 5,000-task Professional tier is US$89/month billed annually, before other usage. That is a tier illustration, not an all-in recommendation.

Example 2: 300 AI drafts, human review and two app writes

Approval and AI costs must be modeled explicitly. Each item requires one AI draft, one human decision and two downstream writes. Assume illustrative Relay totals of 1,200 completed steps and 600 AI credits; these are sample ledger inputs, not a claim that each AI draft always consumes two Relay credits.

For Make, model a one-credit trigger, one ordinary external-AI module with your own provider connection, and two one-credit writes: 1,200 credits plus approval infrastructure. If the notification/state/callback design adds four one-credit operations per item, the total becomes 2,400 credits, plus the external model bill. Make-managed AI would need a different token-credit calculation.

For Zapier, the two standard writes contribute 600 tasks, plus 300 × the applicable AI-step rate, plus approval/helper task usage. The pricing reference points to a variable AI rate card that was unavailable during this review; no flat AI task rate is assumed. Standalone Agents usage is a separate ledger, not silently included.

For n8n, a design that waits and resumes within the same top-level workflow starts with 300 production executions, plus the model bill. For Activepieces, a single flow run per item starts with 300 run credits; confirm how the chosen agent step and approval resume are billed. A design using separate triggered workflows can increase run counts. Prove this in usage history.

Example 3: a daily batch processing 100 records

Batching changes the denominator, but it does not guarantee the same capacity or reliability. Run 30 daily batches; each processes 100 records with four ordinary app actions and no AI. Assume the records arrive together and no extra fetch/iterator modules are required for the illustration.

n8n’s quota rules say scheduled runs count when triggered, empty polling checks do not, and sub-workflows called through Execute Sub-workflow do not add to the parent quota. Independent trigger invocations are different. Large batches still face runtime, memory, concurrency and third-party rate limits.

Cost-check rule: replay the same safe sample, inspect each platform’s usage ledger, then multiply by expected volume and add retry/headroom assumptions. Never convert Relay AI credits into Make credits, Zapier tasks or Activepieces run credits by matching their names.

Prove each connector before choosing a platform

Build one complete representative workflow in each of your two finalists. The following is a proposed acceptance test, not a report that we ran it. Use a lead-intake → CRM → AI draft → review → draft-email → audit-table path, replacing the apps with your actual stack.

AI-generated acceptance-test illustration, not a vendor UI or test result.

For every row, record the candidate’s exact action/module/node/piece name and version, plan requirement and whether implementation is native or a custom API call. A custom HTTP workaround is a maintenance commitment, not connector parity. Include pagination, app-plan API restrictions and OAuth renewal.

Test a valid record, duplicate, missing required field, revoked connection, rate-limit response, approval rejection and timeout. Assert the expected destination record and absence of unintended side effects—not just “run succeeded.” Keep the failed cases and their fixes alongside the export archive.

Be particularly careful with Activepieces CSV imports: its docs warn that a table holds 10,000 records by default and rows beyond available capacity can be dropped without an error. A completed import is not a reconciled import. Compare counts and resolve capacity before switching.

Reconnect credentials and stage the cutover

Create fresh destination connections and switch one active writer at a time. Relay says its stored app credentials/tokens will be deleted with the account; that does not transfer credentials into another vendor or revoke every separately created key in your business.

Use organization-owned identities where appropriate, least-privilege OAuth scopes and the destination’s protected connection settings. Reconnect in the provider’s authorization flow. Rotate Relay-specific API keys after dependencies are identified; do not blindly rotate a shared production key before other consumers are updated. Revoke obsolete provider grants and replace old webhook/MCP endpoints after cutover. Do not put secrets in exported prompts, workflow JSON or shared spreadsheets.

AI-generated cutover diagram. Relay exports are evidence, not a running fallback.

Stage the switch:

1. Freeze and snapshot. Inventory enabled workflows, pending approvals, Sequence positions, scheduled jobs and last processed IDs. Assign an owner and manual fallback for each business-critical process.
2. Rebuild disabled. Import data into test destinations, reconnect accounts and run the connector acceptance cases. Preserve business IDs rather than relying on a new platform’s internal run IDs.
3. Shadow safely. Compare outputs without sending messages, charging money or duplicating CRM writes. Shadow mode means disabled side effects, not two production systems both writing.
4. Cut over a narrow slice. Stop the old writer, record the event watermark, redirect the trigger and enable the new writer. Maintain an idempotency ledger so replayed events cannot create duplicate invoices, emails or records.
5. Reconcile and expand. Compare incoming events, completed outcomes, table counts, pending decisions and failed items. Set alerts and spending limits. Expand only after the owner accepts the observed results.
6. Rollback by stopping writes first. If acceptance fails, pause the replacement, preserve its logs and reconcile partial effects before replay. Before shutdown, Relay is an option only if access still works and the old path is safe. After shutdown, rollback must mean a tested alternate path or manual processing—not reactivating Relay.

On deadline day, do not compress a multi-day validation program into an untested production switch. Preserve evidence, move critical work to a documented manual queue and rebuild safely from there. Ask Relay support about unfinished exports while access remains; no recovery after deletion is promised here.

For a broader shortlist, see our small-business automation guide. For a tighter decision, compare Make vs n8n, n8n vs Zapier, or Activepieces pricing. This guide’s recommendation remains: export now, prove the hardest workflow, then commit.

Relay.app alternatives FAQs

Can Make, Zapier or n8n import my Relay JSON directly?

Relay’s documentation says its JSON does not work through other builders’ native importers. Use it and the build prompt as reconstruction evidence. Validate any destination-specific converter separately; no converter was tested for this guide.

Which alternative is closest to Relay’s human approvals?

Zapier’s Human in the Loop is a strong managed-workflow candidate for approval and data collection. n8n offers selective human review of AI tools. Activepieces lists a Legacy Approval piece. The closest fit depends on reviewer access, editing, rejection, timeout and audit requirements—not a universal ranking.

Are Tables and run history automatically restored?

No. Table CSVs need schema and row-count reconciliation. Run-history files preserve evidence, not a live workflow’s pending execution state. Record outstanding Sequence work and approvals separately.

Is the cheapest replacement the one charging per run?

Not necessarily. Long workflows may benefit from run-based billing, but AI charges, plan features, runtime limits and operator time can dominate. Compare the same business workload with each vendor’s own ledger.

What if I missed the export deadline?

Locate previously downloaded ZIPs, prompts, screenshots and data in connected business apps. Contact Relay support, but do not assume deleted account data can be restored. Reconstruct critical operations from records you still control.

How this guide was evaluated

This is vendor-document research and editorial analysis, not hands-on migration testing. Relay’s deadline and export contract were checked on September 14, 2026, alongside the first-party pricing and documentation linked above. Recommendations prioritize workflow fit, operating burden and recoverability; workload figures are explicitly illustrative.

Vendor migration/alternatives pages were treated as interested-party search signals, not independent comparative evidence. Make’s base-price page was blocked, n8n’s extracted pricing page omitted numeric prices, and Zapier’s linked variable-AI rate card was unavailable. Those details are not presented as verified. Confirm live quotes and plan entitlements before purchasing. Generated visuals explain procedures; they are not product screenshots or proof of successful tests.