Building a Chatbot on Free Tiers: Where the Limits Actually Bite
Most free-tier math for AI products assumes a trial that expires. The stack behind Rune Score, a sports analytics chatbot built at a total cost of $0, points somewhere else. Nearly every layer offers a standing free plan, and the limits that matter are usage ceilings, not countdown clocks. The first wall you hit is request or user volume.
From the outside, a chatbot like this looks simple. Underneath, it has to handle fresh sports data, users who rephrase questions, extraction failures, LLM call costs, streaming responses, conversational memory, fallback logic, and security. Each of those pressures lands on a different tool, and each tool has its own free allowance. Here's how the pieces fit, and where the free tiers get thin.
Branding and code hosting
A chatbot needs a face before it needs a backend. The logo was generated in Figma from a text prompt describing the style and mood, then refined over several rounds without a designer involved. Figma's free plan covers individual use with unlimited file storage, which is more than enough for a logo and a few interface mockups.
The codebase lives in a GitHub repository, kept private while the bot was rough. That same repository is the connection point for both Claude and Codex, so it's the single source of truth for every tool in the chain. GitHub's free plan includes unlimited private repositories with a reduced feature set, plus 2,000 Actions minutes a month on standard runners for private repos. That's enough for tests and basic automation. Standard runners on public repositories are free regardless of usage.
Development: one model builds, another reviews
Claude Code did the building, covering the backend and the AI agents that run inside the chatbot. The workflow was strict. Claude never touched the main branch. Every task got its own branch, and nothing merged until the diff had been reviewed and the pull request approved by hand. That discipline matters more when an AI writes most of the code.
Access is the sticking point. The genuine free route is a Guest Pass. If someone you know already pays for Claude Max, they can send you one, which unlocks a full week of Pro-level access. That's a window, not a permanent plan, and it's the layer where sustained free building gets awkward.
Codex took a different role. It stayed connected to the same repository but never wrote production code. Instead, it reviewed what Claude shipped, acting as an independent judge on pull requests. It also drafted the next development prompt to hand back to Claude. The effect was one model planning the work and another executing it, with a human approving the output.
Codex has a free plan at $0 a month. According to the setup described, free accounts run on GPT-5.6 Terra, OpenAI's cost-efficient model for that tier. The cap is on usage rather than time. For a reviewer role, that's a reasonable fit, since review work arrives in bursts tied to pull requests, not in a constant stream.
Here's the thing: the coding assistant is the one layer where the free story depends on what you're doing. Claude Code's free chat plan doesn't include it, so heavy building means either waiting on a Guest Pass or leaning more on Codex's free tier. Anyone planning a sustained build should price that layer first.
Domain, security, and deployment
The domain was registered through Cloudflare, which sells registrations at cost without markup, and DNS stayed in the same place rather than being split across providers. On top of that, Cloudflare's free security layer protects the domain. Rate limiting rules stop any single source from hammering the chatbot's endpoints. Cloudflare's WAF documentation covers this kind of rule. DNS, DDoS protection, and basic security features are all included on the free plan.
Hosting ran on Railway for both backend and frontend, with a separate test environment so changes could ship quickly without touching the live version. New accounts receive a one-time free credit grant before a card is required. Railway doesn't put services to sleep by default. Sleep behavior is tied to its optional Serverless feature, as described in Railway's documentation. Unless you switch that on, there's no cold-start penalty on the free tier, which matters for a chatbot where a slow first reply reads as a broken product.
Authentication, data, and observability
User sign-up and login run through Clerk. The whole flow was built in Clerk's test environment first, which allowed edge cases to be tried without creating real accounts or sending real emails. The free plan covers up to 50,000 monthly active users at no cost, enough room to go from testing to launch on the same setup.
Conversations, user records, and application data sit in Neon, a serverless Postgres database that scales to zero when idle. That means it costs nothing while nobody is using the product. The free tier includes storage and monthly compute hours sufficient for an early product, and no credit card is needed to start.
Once the agents were live, the next question was what they were doing. Langfuse traces prompts, failures, and costs across the agent runs. Its free Hobby plan includes a monthly allowance of tracing events, enough to debug and monitor a real chatbot. For a system with LLM call costs as a moving variable, that visibility is the difference between guessing and knowing where the money goes.
Research, model selection, and voice
Exa and Firecrawl were wired directly into the research agents inside the chatbot's runtime. That lets the bot decide on its own when it needs to look something up. Exa handles search, finding relevant pages for a question. Its free tier gives new accounts starting credit and a recurring monthly allowance with no payment method required. Its browser playground lets you test queries before wiring them into an agent. Firecrawl then converts a page into clean text the agent can read, with its own free monthly credit allowance and no card needed.
Choosing a model came before committing to one. NVIDIA's API Catalog offers free credits for testing a range of hosted models through its endpoints. That made it possible to compare options without per-token charges, then lock in the one that powers the chatbot in production.
The bot also talks back. Text-to-speech runs on Azure AI Speech, which lets the chatbot switch between female and male neural voices depending on the setup. The free tier includes a monthly allowance of neural voice characters, enough for everyday conversations before a bill would appear.
What the free tiers actually cap
Across this stack, the pattern is consistent. Almost every free plan limits volume: users, requests, tracing events, credits, or characters. Very few limit by calendar. For a builder, that changes the planning question from "how long can I run this?" to "how much traffic can this survive before it costs something?"
That said, the caps are generous for an early product. A chatbot with a real user base can run on these tiers long enough to prove demand, and the upgrade path is clear once it does. The exception is the coding assistant. Review and build work scale with development effort, and the free routes there depend on someone else's subscription or on a more limited free plan. If you're budgeting a sustained build, that's the line item that will need a decision first.