Self Hosted AI

How I'm Rebuilding Cursor with My Own Homelab AI — Zero API Costs

🚫 Why Not Just Pay for Cursor or Copilot?

Because I knew I could do better — for free.

Cursor and GitHub Copilot Chat are incredible tools. They inject context, let you chat about your codebase, and feel like having a pair-programmer built into your editor. But here's the thing:

They require paid API keys (OpenAI, Claude, etc.)
They tie your usage and logs to external systems
They don’t respect your compute — they rent you theirs

I wanted the same dev-experience, but entirely self-hosted, powered by my own GPU server at home. No token limits, no vendor lock-in.

And I’m almost there.

🧱 My Stack (aka The Homelab Dev Agent Engine)

Component	Role
Ollama	Local LLM runtime
DeepSeek-R1	Fast, local, open chat model
Node.js API	My custom wrapper server
Avante.nvim	Neovim AI interface plugin
PostgreSQL	Token, prompt, error tracking
Tailscale VPN	Internal access across machines

I wanted to mimic Cursor — with /slash commands, local file injection, contextual AI completions — using nothing but my own infrastructure.

🧩 How It Works (End to End)

My Node.js API does the heavy lifting:
- Validates bearer/API keys
- Accepts selected_files from Avante
- Reads each file and injects its contents into the prompt (like Copilot Chat)
- Sends the structured request to Ollama, locally
- Streams the AI response back to Neovim
- Logs every prompt, token count, and even simulated OpenAI/Claude pricing into PostgreSQL
The Ollama Server, running DeepSeek-R1 on my GPU workstation, handles completions locally.

Avante.nvim in Neovim is configured to call my API:

require("avante").setup({
  endpoint = "https://api-ai.cargdev.io/api/chat",
  api_key = "local-dev-key",
  model = "deepseek-r1:latest"
})

🪄 Cursor-Like Features I Already Have

Feature	Status	Implementation
File-aware chat	✅	Avante + my wrapper inject file contents
Slash commands	✅	`extensions.avante.make_slash_commands`
Streaming responses	✅	Via Ollama stream + SSE
Chat UI in Neovim	✅	Avante handles buffer formatting
Prompt + error logging	✅	PostgreSQL backend
Token cost simulation	✅	GPT-4, Claude, Gemini savings logged

🧪 Still Missing Full Cursor Behavior — But Getting Close

I don’t yet have true buffer modification support (e.g. code refactor suggestions auto-applied). Cursor and Copilot Chat can modify code intelligently based on full-project context.

But the pieces are almost in place:

My wrapper supports per-file context
Prompt injection works for any file selection
Avante can be extended or patched to support diff/edit mode

And unlike Cursor, my tool:

Runs 100% local
Is fully inspectable
Costs me $0/month

📊 Tracking Everything

Every request logs:

{
  "model": "deepseek-r1:latest",
  "total_tokens": 9992,
  "prompt": "...",
  "selected_files": ["mcphub.lua"],
  "created_at": "2025-05-24"
}

I also record:

GPT-4 equivalent cost (e.g. $0.10)
Claude and Gemini pricing
My actual cost: $0

This makes it easy to say: “This week I saved $7.20 in token fees.”

🙌 Want to Use or Contribute?

Check out the repo:

🔗 https://github.com/CarGDev/apiAi

You’ll find:

Express.js wrapper code
Ollama integration
Token tracking logic
Prompt/error endpoints
Sample setup for Avante.nvim

Feel free to fork it, log an issue, or submit a pull request. There’s plenty of room to grow: edit suggestions, embedded diff views, buffer streaming, agent queues, you name it.

☕ Final Thoughts

This isn’t just about saving money — it’s about owning your tools.

I’m not building a Cursor clone for the hype.
I’m building a private AI dev agent, fully under my control.

If you’re a Neovim user, have a spare GPU, and want to break free from paid APIs, give this a try.

There’s still a lot to add — but the core is solid.

And yeah, I proudly say it:

“I’m running my own Cursor-style AI stack — from my homelab.”

More posts coming soon. Until then, keep breaking and building. 💻🔥

How I'm Rebuilding Cursor with My Own Homelab AI — Zero API Costs

🚫 Why Not Just Pay for Cursor or Copilot?