Fireship's Code Report replaces paid AI subscriptions with five open source tools: Ollama, 9router, Headroom, Dify, and OpenHands. Here is what each one does, and an honest look at when self-hosting actually saves a small team money.
In this September 2026 episode of The Code Report, Fireship argues you can cancel most paid AI subscriptions and self-host a cheaper AI stack built from five open source tools: Ollama, 9router, Headroom, Dify, and OpenHands. Ollama runs models locally. 9router puts all your model providers behind one endpoint with automatic fallback. Headroom compresses context to cut input tokens. Dify is a visual builder for AI workflows. OpenHands is an autonomous coding agent.
The video tallies the subscriptions behind the "$320 a month" in its title: Cursor, Claude Max, GPT Pro, and Gemini Ultra. It's aimed at developers, and the episode is sponsored. The ideas apply to any small team watching its AI bill, but self-hosting is not automatically cheaper.
Video credit: “5 open source tools that replaced my $320/mo AI stack...” by Fireship, published September 7, 2026 on YouTube (1.26 million views when we wrote this). Watch on YouTube. All rights to the video belong to its creator; we embed it with YouTube's standard player and add our own commentary.
Fireship opens with the math that started it:
Cursor, $20. Claude Max, $100. GPT Pro, another 100. Gemini Ultra, another 100.
| Tool | What it does (per the video) | How it cuts cost | Main catch |
|---|---|---|---|
| Ollama | Command line and API for downloading and running open-weight models on your own machine | Zero per-token cost; prompts stay private | Frontier-size models need serious hardware |
| 9router | Self-hosted proxy that puts dozens of model providers behind one OpenAI-compatible local endpoint | Fallback tiers use subscriptions you already pay for first, then cheap or free providers; tracks usage and compresses tool output | One more service to run and keep updated |
| Headroom | Context compression layer between your app and the model provider | Shrinks tool outputs, logs, and other noise before they become billable input tokens | Only pays off if you send a lot of context |
| Dify | Visual drag-and-drop builder for AI workflows, exposed as an API | Build AI features without a separate paid workflow tool | Still calls a model, so token costs remain |
| OpenHands | Autonomous coding agent you self-host; works from GitHub issues | Runs with OpenAI or Anthropic models, or local models through Ollama | Agents left running can burn tokens fast if pointed at paid models |
Fireship's one-line pitch for Ollama is the easiest way to understand it:
Essentially, it's like Docker for large language models.
The video builds a stack in layers, and it's worth seeing it that way because each tool solves a different cost problem:
One clever design feature of this tool is that it's reversible.
The key idea is that you don't have to choose between local and frontier models. Fireship keeps the option to tap Claude and GPT for hard tasks while routing everything else to cheaper options.
Fireship is honest about the biggest limit:
The big problem though is that you probably don't own the hardware to run anything close to state-of-the-art.
That shapes the whole decision. Here is our rule of thumb for small teams:
| Situation | Self-hosting likely saves money? | Why |
|---|---|---|
| Heavy API usage (agents, long context, big codebases) | Often yes | Routing and compression cut a real, recurring token bill |
| Several overlapping chat subscriptions per person | Yes, partly | Consolidating and routing removes duplicate seats |
| Simple, repetitive tasks (classifying, summarizing, drafting) | Often yes | A small local model is usually good enough |
| Privacy-sensitive data you can't send out | Yes, for compliance | Local models keep prompts on your machine |
| Light, occasional use | Usually no | One subscription costs less than the server and the time |
| Nobody on the team can maintain servers | Usually no | Breakages, updates, and security become unpaid work |
| Tasks that need frontier-level reasoning | No | You still pay a big provider; self-hosting only trims the edges |
A subscription bundles things you don't see. When you self-host, you take them on:
The fair summary: self-hosting turns a monthly bill into a mix of a smaller bill plus your time. That trade is great for developers who enjoy it, and expensive for a business owner who doesn't.
If you want lower AI costs without rebuilding your stack, start small:
For a wider look at which tools are worth paying for, see AI tools for solopreneurs. If you're weighing self-hosted agents generally, what is OpenClaw covers a similar trade-off, and AI agents vs agentic AI explains the terms.
Dooza is an AI-native company that builds AI products and services for small businesses, from the Dooza Workforce app to the Dooza Agents platform. Every product starts with a refundable pilot: 100% refund within 14 days.
Fireship's stack is built for developers who want to run their own infrastructure. Most small business owners we talk to want the result, not the servers. That's the gap Dooza fills.
Dooza Workforce gives you ready-made AI employees for jobs like email, social media, SEO, and lead generation, with no hosting to manage. On Dooza Agents, Dooza engineers build and maintain custom agents for your workflows, so model choice, cost control, and upkeep are handled for you. If your need is narrower, look at workflow automation, an AI receptionist, or AI customer support. Our guide to automating business processes is a good starting point.
Pricing depends on the product; see pricing. Every product starts with a refundable pilot.
Want lower AI costs without running servers? Book a free 30-minute call to scope your pilot. Start with a refundable pilot — 100% refund within 14 days.
Ollama for running models locally, 9router for routing requests across providers with fallback tiers, Headroom for compressing context, Dify for visual AI workflows, and OpenHands for autonomous coding agents.
Fireship adds up Cursor at $20 plus Claude Max, GPT Pro, and Gemini Ultra at $100 each, as quoted in the video, which comes to $320 a month.
Only sometimes. It saves money when API usage is high, tasks suit small local models, and someone can maintain the servers. For light use or teams without technical staff, a subscription is usually cheaper overall.
For simple tasks, often yes. But as Fireship notes, most people do not own hardware to run anything close to state-of-the-art models, so frontier-level work still needs a big provider.
According to the video, Headroom sits between your app and the model provider and compresses tool outputs, logs, and other noise before they become billable input tokens. The compression is reversible because the original is cached locally.
Yes. Dooza Workforce provides ready-made AI employees and Dooza Agents provides custom agents built and maintained by Dooza engineers. Every Dooza product starts with a refundable pilot: 100% refund within 14 days.
Automate your business with AI employees that work 24/7. Start with a refundable pilot: 100% refund within 14 days.
In his viral video, Mrwhosetheboss tests a keyboard PC, a 16 mm mechanical keyboard, a 6.4 mm pair of earbuds, paintable light, and more. Here is every device, its thickness, his verdict, and the lessons for anyone buying gear for work.
Fireship's Code Report covers OpenAI's claim that its agents cracked the Navier-Stokes Millennium Prize Problem, and an NYU mathematician's account of why the story is more complicated. Here is a balanced summary and the lesson for anyone relying on AI output.
Start with a refundable pilot — 100% refund within 14 days. A Dooza engineer scopes it with you on a free 30-minute call. Pricing depends on the product; see pricing.
Refundable pilot · 100% refund within 14 days · No contracts