✧  ai.nvim  ✧

Query LLMs following OpenAI API specification

---

💡 Idea

LLM providers offer libraries for the most popular programming languages, so you can build code that interacts with their API. Generally, those are wrappers around HTTPS requests with a mechanism to handle API responses (e.g., using callbacks).

To the best of my knowledge, if you want to build a plugin for Neovim that uses LLM, you have to explicitly make requests using a library like curl and take care of requests and responses parsing yourself. This results in a lot of boilerplate code that can be abstracted away.

ai.nvim is an experimental library that can be used to build Neovim plugins that interact with LLM providers: it crafts requests, parses responses, invokes callbacks, and handles errors.

⚡️ Requirements

• Neovim ≥ 0.9
• Curl
• Access to LLM Provider

🚀 Usage

Read the documentation with :help ai-nvim

Plugins built with ai.nvim:

• S1M0N38/dante.nvim ✎ A basic writing tool powered by LLM
• S1M0N38/chatml.nvim ⇋ Convert markdown to LLM JSON requests
• PR your plugin here ...

✨ LLM Providers

There are many providers that offer LLM models exposing OpenAI-compatible API. The following is an incomplete list of providers that I have experimented with:

• Local models (LM Studio, Ollama, llama-cpp, vLLM, ...)

  • base url: http://localhost:[PORT]
  • models: every model supported by the local provider
  • note: free and private.

• Copilot

  • base url: https://api.githubcopilot.com
  • models: gpt-4o-mini, gpt-4o, o1, o3-mini, gemini-2.0-flash-001, claude-3.5-sonnet, claude-3.7-sonnet, claude-3.7-sonnet-thought
  • note: access to SOTA models with GitHub Copilot subscription (free for students)

• OpenAI

  • base url: https://api.openai.com/v1
  • models: gpt-4o, gpt-4o-mini, o1, o1-mini
  • note: access to SOTA models (no free tier)

• Groq

  • base url: https://api.groq.com/openai/v1
  • models: gemma2-9b-it, llama-3.3-70b-versatile, llama-3.1-8b-instant, mixtral-8x7b-32768, qwen-2.5-coder-32b, qwen-2.5-32b, deepseek-r1-distill-qwen-32b, deepseek-r1-distill-llama-70b-specdec, deepseek-r1-distill-llama-70b, llama-3.3-70b-specdec
  • note: crazy fast inference for open source models (free tier)

• Mistral

  • base url: https://api.mistral.ai/v1
  • models: mistral-large-latest, ministral-3b-latest, ministral-8b-latest, mistral-small-latest, open-mistral-nemo, mistral-saba-latest, codestral-latest
  • note: access to Mistral models (free tier)

• Codestral (Mistral)

  • base url: https://codestral.mistral.ai/v1
  • models: codestral-latest
  • note: access to Mistral code models (free for now)

If you want to use other providers that do not expose OpenAI-compatible API (e.g., Anthropic, Cohere, ...), you can try liteLLM proxy service.

There is no future plan to support other API standards besides OpenAI-compatible API.

🙏 Acknowledgments

• base.nvim template.
• mrcjkb's blog posts about Neovim, Luarocks, and Busted.
• mrcjkb and vhyrro repos' for GitHub Actions workflows.
• codecompanion.nvim for Copilot token validation logic.