vibe.vim

Overview

AI coding assistant for Vim and Neovim. Chat with your model about code and perform inline edits — works in classic Vim 8.2 and Neovim, with local or hosted models.

• Chat Chat works end-to-end with Ollama, Anthropic, OpenAI, and any OpenAI-compatible endpoint.
• ACP ACP subprocess providers are supported. Inline diff accept/reject UI is in progress.
• Vim 8.2 and Neovim — not Neovim-only
• No npm. One Go binary, one Vimscript file, nothing else to install
• Local models first — Ollama works out of the box, no config required
• ACP mode — front-ends any AI CLI tool (Claude Code, Codex CLI, custom scripts) as a backend directly, so you get the full agent loop without leaving Vim
• Provider choice — Anthropic, OpenAI, Cerebras, Groq, or bring your own via the inherits config key

Requirements

• Vim 8.2+ or Neovim
• Go 1.22+ (to build the binary)
• A running provider — Ollama by default, no config needed

Install

1. Build the binary

git clone https://github.com/zblauser/vibe
cd vibe
go build -o vibe ./cmd/vibe
mv vibe ~/.local/bin/   # or anywhere on your PATH

2. Install the plugin

With vim-plug:

Plug 'zblauser/vibe', {'rtp': 'plugin'}

Or manually — copy plugin/vibe.vim into ~/.vim/plugin/ (Vim) or ~/.config/nvim/plugin/ (Neovim).

3. Configure a provider (optional)

Without a config file, vibe defaults to Ollama at localhost:11434. To use other providers, create ~/.config/vibe/config.toml. See Provider setup below.

Usage

Command	Description
:Vibe [prompt]	Open chat panel, optionally send an initial prompt
:VibeChat	Toggle the chat panel
:'<,'>VibeEdit [prompt]	Send a visual selection for inline edit
:VibeInline [prompt]	Send the current line for inline edit
:VibeCancel	Cancel the current stream
:VibeProvider <name>	Switch provider for this session
:VibeAddFile <path>	Pin a file into every subsequent request
:VibeStatus	Show current provider, model, session, pinned files
:VibeDebug	Show internal state

Chat panel keys

The panel is two panes: a scrollable history (top) and an input buffer (bottom).

Input pane:

Key	Action
<CR> (normal mode)	Send
<C-s> (insert mode)	Send
<Tab>	Move focus to history
<C-c>	Cancel the current stream
q	Close the panel

History pane:

Key	Action
<Tab>	Move focus to input
<C-c>	Cancel the current stream
q	Close the panel

Smoke test

vibe test-provider ollama
# → should print the model's response and token usage

---

Provider setup

Ollama (default — no config needed)

ollama pull llama3.2

# ~/.config/vibe/config.toml  (optional — these are the defaults)
default_provider = "ollama"

[providers.ollama]
endpoint = "http://localhost:11434"
model    = "llama3.2"

Anthropic

Set the ANTHROPIC_API_KEY environment variable, then:

:VibeProvider anthropic

Or configure explicitly:

default_provider = "anthropic"

[providers.anthropic]
api_key_env = "ANTHROPIC_API_KEY"
model       = "claude-sonnet-4-6"

OpenAI

[providers.openai]
api_key_env = "OPENAI_API_KEY"
model       = "gpt-4o"

Custom OpenAI-compatible endpoint

The inherits = "openai" key routes through the OpenAI-compatible provider, so any endpoint that speaks /v1/chat/completions works:

[providers.cerebras]
inherits    = "openai"
endpoint    = "https://api.cerebras.ai/v1"
api_key_env = "CEREBRAS_API_KEY"
model       = "llama3.1-8b"

[providers.groq]
inherits    = "openai"
endpoint    = "https://api.groq.com/openai"
api_key_env = "GROQ_API_KEY"
model       = "llama-3.3-70b-versatile"

ACP — subprocess providers

type = "acp" launches a local command for each request. The binary sends a single JSON line on stdin and reads NDJSON events from stdout. Any script in any language can implement this:

[providers.myscript]
type    = "acp"
command = "python3 ~/scripts/my_provider.py"
env     = {MODEL = "custom-model", LOG = "1"}

ACP wire protocol:

Input (one JSON line on stdin):

{"messages":[{"role":"user","content":"..."}],"system":"..."}

Output (one JSON object per line on stdout):

{"type":"text","text":"..."}
{"type":"thinking","text":"..."}
{"type":"end","usage":{"input":100,"output":50}}
{"type":"error","message":"something went wrong"}

The subprocess inherits the parent environment; env keys override or add to it.

Change Log

v0.2.0 (Latest)

Provider update

Features

• :VibeChat opens a two-pane split; type in the bottom pane and send with <CR>
• Current buffer is automatically sent as context with every message
• :'<,'>VibeEdit and :VibeAddFile send selections and pinned files
• Streaming responses render chunk-by-chunk with thinking block support
• Session history is kept — the model sees the full conversation
• :VibeProvider switches providers mid-session
• Ollama, Anthropic, OpenAI, and ACP subprocess providers all work

Previous Versions

Previous Changes

v0.1.0

• Initial commit
• Ollama local chat

Roadmap

• Inline diff UI — accept/reject edit proposals in a proper diff view
• Tool approval UX — approve/reject/always-allow tool calls from agentic backends

Contributing

If you share the belief that simplicity empowers creativity, feel free to contribute.

Contribution is welcome in the form of:

• Forking this repo
• Submitting a Pull Request
• Bug reports and feature requests

Please ensure your code follows the existing style.

Project layout

cmd/vibe/           binary entry point + protocol loop
internal/protocol/  NDJSON message types (binary ↔ plugin)
internal/provider/  provider interface + Ollama, Anthropic, OpenAI, ACP
internal/session/   conversation state + streaming
internal/config/    config.toml loader
plugin/vibe.vim     the entire Vim plugin (one file)
doc/vibe.txt        :help vibe

Thank you for your attention.

This project started out of curiosity and a simple C text editor tutorial. If you hit any issues, feel free to open an issue on GitHub. Pull requests, suggestions, or even thoughtful discussions are welcome.