Fireship’s five-tool AI stack, explained

Can running your own AI tools cut subscription costs?

AI coding subscriptions can add up. In a video published on September 7, 2026, developer channel Fireship introduces five tools he says replaced an AI subscription mix costing him $320 a month: Ollama, 9Router, Headroom, Dify, and OpenHands.

Each does a different job. They run models, connect AI services, shorten the material sent to a model, build app workflows, or assign coding tasks to an agent that writes and changes code. You can install and operate the software on your own computer or server—a practice called self-hosting. You can also connect paid services such as Claude or GPT when you need them.

The $320 is Fireship’s account of his own spending: $20 for Cursor and $100 each for Claude Max, GPT Pro, and Gemini Ultra. We have not verified those figures as current list prices or measured the savings. The useful question is what each tool does, and what it would cost you to run.

• Fireship, The Code Report (YouTube)

Ollama: run an AI model on your computer

Ollama downloads and runs models on your own computer. Instead of sending a question to an external AI service, you ask the model installed on your machine. Models whose files you can download and run yourself are known as open-weight models.

When you use only a local model, your prompts do not need to go to an external model provider, and that provider does not charge you for replies. You still pay for the computer and its electricity. Larger models also need more capable hardware.

Fireship makes that tradeoff explicit: small models can run on ordinary machines, while large, advanced coding models need much more hardware. Whether a local model can provide the coding help you currently get from Claude or GPT is a separate question to check.

9Router: connect several AI services in one place

9Router passes requests between coding apps and model providers. Your app sends a request to 9Router, which selects an AI service according to your settings. That gives you one place to manage several service connections.

For example, it can use an existing subscription first, switch to a cheaper pay-as-you-go model when the allowance runs out, then try a free or trial service. Usage-based model fees commonly depend on tokens: small units of text a model reads or writes. Fireship names an OpenAI model as his cheaper option; the corresponding examples in the 9Router README are GLM and MiniMax.

9Router itself does not invoice users. The providers it connects to can still charge you, and its cost dashboard shows estimates rather than actual bills. It also includes RTK, a feature enabled by default that shortens tool results. If compression fails or makes a result longer, it sends the original.

Free services have usage and time limits too. The table records conditions in the 9Router README; check a service’s latest terms before relying on it.

• 9Router GitHub repository

Headroom: send less material to the model

Coding agents read logs, search results, and tool output as they work. Long results increase the amount sent to the model. Headroom shortens that material before it reaches the model, reducing the input. Short or already information-dense text offers less room for compression.

Headroom keeps the original text on your machine so the model can retrieve the full version when it needs more detail. Its README calls this mechanism CCR.

It can work with 9Router: the documented integration lets 9Router ask Headroom to compress a request before forwarding it to a model provider. If Headroom does not respond, 9Router sends the original. Alongside self-hosting, Headroom offers paid support and managed team services.

• Headroom GitHub repository

Dify: arrange the steps of an AI app

Dify lets you connect an app’s processing steps visually. An app might receive information, look up records in a database, and ask AI to write a response. You can expose that workflow through an API—a connection that lets other apps request the result.

Fireship explains the idea with a joke app that matches horses. It takes a horse’s profile, finds a match in a database, then asks AI to explain the pairing. The example is playful, but the sequence is useful: input, information lookup, then an AI explanation.

Dify’s documentation describes Community Edition, which you install yourself, and Dify Cloud, which the company operates for you. Cloud includes a free Sandbox plan. The choice depends in part on whether you want to manage the server yourself.

• Dify documentation

OpenHands: give coding tasks to an AI agent

OpenHands lets you assign and manage work by an AI coding agent. An agent goes beyond replying in a chat: it is an AI program that carries out a task.

The documentation separates the browser interface for conversations and work, Agent Canvas, from the component that runs agents, Agent Server. Developers can also build agents with the Software Agent SDK, a Python toolkit.

Fireship describes keeping agents running on a server and connecting them to models from OpenAI or Anthropic, or to models run through Ollama. The model you choose affects both fees and hardware needs. Alongside running it yourself, there is a managed OpenHands Cloud service and an Enterprise offering for organizations.

• OpenHands introduction

Start with the job you need done

The useful idea in this video is to separate AI work into distinct jobs. Running a model locally, connecting several services, and building an app are different needs. You can choose a starting point without installing all five tools at once.

If saving money is the goal, count model fees, computer or server costs, and maintenance time alongside subscriptions. If your current service’s coding quality matters, try the alternative on the work you need it to do before switching.