I Built an AI Chatbot for $0. Here’s the Exact Stack.
You don’t need money to ship a real, production-grade AI chatbot. You need to know which free tiers to stack. Here’s the exact toolchain.
Opinions expressed by Entrepreneur contributors are their own.
Key Takeaways
- Every layer of a modern chatbot, from the logo to the database to the voice output, now has a genuine free tier, not just a trial.
- Almost all of these tools cap usage rather than time. That means the limit you’ll hit first is volume of users or requests, not a countdown clock.
- The one layer worth watching closely is the coding assistant itself. Claude Code’s free chat plan doesn’t include it, so sustained use eventually means either a Guest Pass, or leaning more on Codex’s own free plan.
I’m an AI engineer and I’m currently building ChatMakr, an AI-powered platform that creates multi-agent chatbots in seconds. I also built Rune Score, a sports analytics chatbot, and it cost me $0.
From the outside, it looks like a simple chatbot. Under the hood, it’s a distributed system that has to handle the freshness of sports data, user rephrasing, extraction failures, LLM call costs, streaming, memory, fallbacks and security. Here’s the stack I actually used to build it layer by layer, what I did with each tool and what’s free about it.
Branding: Figma
Before writing a line of code, a chatbot needs a face. I used Figma to generate the chatbot’s logo straight from a text prompt describing the style and mood I wanted, then iterated on it a few times inside Figma itself rather than handing it to a designer. Figma’s own free plan covers individual use with unlimited file storage.
Code hosting: GitHub
The whole project lives in a GitHub repository, private while the chatbot was still rough, and it’s the same repo both Claude and Codex connect to. GitHub’s free plan gives unlimited private repositories with a limited feature set and 2,000 minutes a month of Actions included on GitHub Free for standard runners on private repos, enough to run tests and basic automation without hitting a wall. Standard runners on public repositories stay free no matter how much you use them.
Development: Claude Code, with Codex as the judge
Claude Code did the actual building, the chatbot’s backend and the AI agents running inside it. I kept it connected directly to the GitHub repository and never let it touch the main branch directly: every task ran on its own branch, and nothing merged until I’d reviewed the diff myself and approved the pull request. The genuine free route in is a Guest Pass: if someone you know already pays for Claude Max, they can send you one, and it unlocks a full week of free Pro-level access.
I gave Codex a different job, connected to the same repository but never writing production code itself. It reviewed what Claude had actually shipped, acting as an independent judge on pull requests, and I also used it to draft the next development prompt I’d hand back to Claude, effectively having one model plan the work and another execute it. Codex has a genuine free plan at $0 a month. Free accounts run on GPT-5.6 Terra, OpenAI’s cost-efficient model for that tier. The free plan is capped at a usage limit rather than a time limit, which is exactly the size of job a reviewer role needs.
Domain and security: Cloudflare
I bought the chatbot’s domain name through Cloudflare, where registration is usually cheap since Cloudflare sells at cost with no markup, and kept the DNS in the same place rather than splitting it across providers. From there I used Cloudflare’s free security layer to protect the domain, including rate limiting rules to stop any single source from hammering the chatbot’s endpoints. Cloudflare’s WAF documentation covers exactly this kind of rule, and the core protections, DNS, DDoS protection, and basic security features, are included on the free plan.
Deployment: Railway
I used Railway for both the backend and the frontend, with a separate test environment so I could ship changes quickly without risking the live version. New accounts get a one-time free credit grant before any card is required, and unlike some free hosts, Railway doesn’t put your service to sleep by default the way its own documentation describes for its optional Serverless feature, so there’s no cold start tax on the free tier unless you turn that behavior on yourself.
Authentication: Clerk
Users sign up and log in through Clerk, and I built the whole flow using Clerk’s own test environment first, so I could try edge cases without creating real accounts or sending real emails. The free plan covers up to 50,000 monthly active users at no cost, which is more than enough room to move from testing into a real launch on the same setup.
Database: Neon
Conversations, users and application data all sit in Neon, a serverless Postgres database that scales down to zero and stops costing anything the moment nobody’s using it. The free tier includes enough storage and monthly compute hours for a real early product, with no credit card required to start.
Observability: Langfuse
Once the chatbot’s AI agents were actually running, I needed to see what they were doing under the hood: which prompts fired, where they failed, and what they cost. Langfuse traces all of that, and its free Hobby plan includes a healthy monthly allowance of tracing events, enough to debug and monitor a real chatbot without paying anything.
Web research: Exa and Firecrawl
I wired Exa and Firecrawl directly into the AI agents responsible for research inside the chatbot’s runtime, so the chatbot itself can decide when it needs to look something up. Exa handles the search, finding relevant pages for a given question, and its free tier hands new accounts starting credit plus a recurring monthly allowance with no payment method required. It also ships a playground where you can test a search query directly in the browser before wiring it into your agent. Firecrawl handles what comes next, turning a page Exa found into clean text the agent can actually read, and its own free plan includes a monthly credit allowance with no card needed either.
Testing AI models: NVIDIA
Before committing to a specific model for the chatbot’s agents, I used NVIDIA’s API Catalog to try several without paying per token. NVIDIA gives any developer free credits to test a wide range of hosted models through its own endpoints, which made it easy to compare options before locking in whichever one actually powers the chatbot in production.
Voice: Azure AI Speech
The chatbot talks back, not just types, so I used Azure AI Speech for text to speech, letting the chatbot switch between a female or male neural voice depending on the setup. The free tier includes a real monthly allowance of neural voice characters at no cost, enough for genuine daily conversations before you’d ever see a bill.
Key Takeaways
- Every layer of a modern chatbot, from the logo to the database to the voice output, now has a genuine free tier, not just a trial.
- Almost all of these tools cap usage rather than time. That means the limit you’ll hit first is volume of users or requests, not a countdown clock.
- The one layer worth watching closely is the coding assistant itself. Claude Code’s free chat plan doesn’t include it, so sustained use eventually means either a Guest Pass, or leaning more on Codex’s own free plan.
I’m an AI engineer and I’m currently building ChatMakr, an AI-powered platform that creates multi-agent chatbots in seconds. I also built Rune Score, a sports analytics chatbot, and it cost me $0.
From the outside, it looks like a simple chatbot. Under the hood, it’s a distributed system that has to handle the freshness of sports data, user rephrasing, extraction failures, LLM call costs, streaming, memory, fallbacks and security. Here’s the stack I actually used to build it layer by layer, what I did with each tool and what’s free about it.
Branding: Figma
Before writing a line of code, a chatbot needs a face. I used Figma to generate the chatbot’s logo straight from a text prompt describing the style and mood I wanted, then iterated on it a few times inside Figma itself rather than handing it to a designer. Figma’s own free plan covers individual use with unlimited file storage.