AI Automation Agency Tech Stack: What to Use and Why

An AI automation agency tech stack needs eight layers: a coding agent to build, a code host, a database, hosting, connections to client tools, messaging, monitoring and payments. Pick one tool per layer, learn it well and reuse it on every client. Decide early which accounts the client owns.

This is the stack taught in Stride AI Academy, with the reason for each choice and the alternatives you'll run into. Tool details were checked against each vendor's docs in October 2026.

How a client build moves through the stack

1. 1BuildClaude Code or Codex writes the code
2. 2VersionGitHub holds every change
3. 3StoreSupabase or Postgres
4. 4HostVercel runs it
5. 5ConnectAPIs and MCP to client tools
6. 6WatchLogs and error alerts

Start from the use case

Start from the use case, not the AI tool. A lead follow-up workflow needs messaging and a CRM integration. A multi-step job built on AI agents also needs a coding agent, clean data sources and monitoring. Decide what you'll automate for the client, then add only the layers that job needs.

1. Building: a coding agent

What we use: Claude Code and Codex.

Most agency builds are integrations and small apps: lead follow-up, inbox triage, dashboards, internal tools. A coding agent writes most of that code from a clear description, and you review and test it. Using code rather than a visual workflow builder means the build lives in a repository you control, with no per-task pricing and no ceiling on what it can do.

Both agents read a project instruction file (CLAUDE.md or AGENTS.md) and connect to outside tools through MCP. See Claude Code vs Codex for the differences.

Visual builders are a reasonable alternative for simple jobs. If you use one, learn it well and know its limits before you sell around it.

2. Code: GitHub

Every client project goes in its own private repository. That gives you history (undo any change), branches (try changes safely), and a clean handover. Hosting platforms deploy straight from it.

3. Data: Supabase or Postgres

Most automations need somewhere to keep state: leads, message logs, run history, settings per client. Supabase gives you a Postgres database with an API and auth. Plain Postgres works too.

Supabase publishes an MCP server, so Claude Code can query the database and inspect schemas while you build. Scope it to one project and keep it read-only where you can (Supabase's MCP docs).

4. Hosting: Vercel

Vercel builds and hosts web apps and server functions straight from GitHub: every push deploys, and branches get preview links you can send to a client. See how to deploy a Claude Code app to Vercel.

5. Connections: APIs and MCP

Client work is mostly connecting systems: a form to a CRM, a CRM to a calendar, an inbox to a help desk. Two kinds of connection:

• The client's tool APIs, called from your code, for the automation itself
• MCP servers, which give your coding agent access to tools while you build (GitHub, the database, error tracking, deployment logs)

Tools you'll connect most often

• GitHubCode and issues
• SupabaseData
• VercelHosting and logs
• TwilioSMS and voice
• StripePayments
• SentryErrors

For workflow automation, tools like n8n and Make handle orchestration: routing data between systems in a set order.

6. Messaging: SMS, email and voice

Lead follow-up, reminders and missed-call text-back all need a messaging provider. For SMS and voice that's a provider such as Twilio, plus an email sending service for transactional email.

Two rules from day one: in the US, business texting from regular 10-digit numbers needs A2P 10DLC registration (Twilio's A2P 10DLC overview), and every message flow needs consent and a working opt-out.

7. Operations: the CRM and pipelines

Your client's leads and appointments have to live somewhere the client's staff actually look. Either connect to the CRM they already use, or run their pipeline in a platform you manage. At Stride that platform is StrideOps, the AI operating system with CRM, pipelines, voice agents and AI follow-up that Josh Pocock's companies run on.

8. Monitoring and payments

• Monitoring: log every automation run, and get alerted on failures. An error tracker such as Sentry catches crashes in the code. Your own run log catches the quiet failures, like a message that never sent.
• Payments: if a build takes payments, use the provider's test mode while building. For your own invoicing, any invoicing tool works. Charge a setup fee and a monthly retainer (how to price it).

Who owns each account

Decide this before the first build:

Account	Usually owned by	Why
CRM, email, phone numbers	Client	Their customer data and their numbers
Messaging provider	Client, or you with costs passed through	Usage scales with their volume
Database and hosting	Either	Client-owned makes handover clean. Agency-owned lets you bundle hosting into the retainer
GitHub repository	Either, transferred at handover if the client owns the build	Matches the IP clause in your contract
Your coding agent	You	It's your tool, not part of the deliverable

Write it down in the contract (what to include).

Reuse what works

The stack matters less than what you reuse. Keep a template for each build and a library of tested prompts, so every client starts from a framework instead of a blank page. Give each build one measurable metric, and audit it monthly so the client can see what it did.

Choosing tools: four rules

1. One tool per layer. Switching costs more than the small gains from a "better" tool.
2. Prefer tools with an API and an MCP server. Your coding agent can then work with them directly.
3. Keep it boring. Clients pay for reliability. Popular, well-documented tools fail less and are easier to hand over.
4. Price the stack into the retainer. List which tool costs you cover and which the client pays directly.

Learn the stack

The Stride AI Automation course covers building agents for lead follow-up, inbox triage, client reports and content pipelines on this stack, with weekly live calls for questions. Start with the free guide library for setup walkthroughs.

FAQ

What tech stack do AI automation agencies use?

It varies, but most need the same layers: a way to build (a coding agent or visual builder), a code host, a database, hosting, connections to client tools, a messaging provider, monitoring and payments. Stride AI Academy teaches Claude Code and Codex, GitHub, Supabase, Vercel, MCP and StrideOps.

Do I need to code to run an AI automation agency?

You don't need to write code by hand anymore, since coding agents write most of it. You do need to understand how systems connect, describe builds clearly, read API docs and test the results.

Should I use n8n, Make or code?

Visual builders are fine for simple workflows and quick tests. Code written with a coding agent avoids per-task pricing, lives in a repository you control and has no ceiling on complexity. Pick one approach and get good at it before you sell.

Who should pay for the tools?

Usage-based tools such as SMS, voice and AI model usage scale with the client's volume, so many agencies pass those through or include them up to a cap. Write the arrangement into the retainer so there are no surprises.