100% local · your data never leaves the device

A real AI agent that runs entirely on your machine. Pair with Ollama, hand it tools and MCP servers, and let it touch the work you can’t send to the cloud — code, contracts, patient notes, financials, anything.

Apple Silicon native
Bring your own model
running locally
vibn·Refactor agent loop
Refactor agent loop~/projects/vibn
14k / 32k
Ask vibn anything…
Features

A real agent. On your machine.

Built in Rust. Powered by Ollama. Designed so you can hand it serious work — and trust where the data goes.

Runs entirely on your machine

Vibn pairs with Ollama on your own hardware. No accounts, no cloud round-trips, no quotas — every prompt, file, and answer stays on the device.

Bring your own model

Qwen, DeepSeek, Llama, Mistral, GPT-OSS — any model Ollama can run. Pick a coder for code, a generalist for prose, a tiny one for fast triage. Switch in one click.

Real tools, not just chat

Read files, edit, search, run commands, generate images, browse — a built-in toolset that touches your actual filesystem, with explicit per-call permission checks.

MCP protocol native

Speak the Model Context Protocol fluently. Wire up Linear, GitHub, your firm's internal MCP server, or any compatible one — and only the ones you trust.

Project-aware memory

Per-project observations build up over time. Vibn remembers conventions, sensitivities, and decisions per workspace so it doesn't drag stale context across matters.

Permission-checked tool calls

Every shell command, every edit, every external call passes through configurable allow / ask / deny rules. Audit what touched what, after the fact.

Hooks for everything

Fire shell commands on session start, before / after edits, on compaction, on commands. Wire Vibn into your existing controls — DLP, backup, audit logging.

Transcripts you own

Sessions land as JSONL in ~/.vibn/transcripts/. Keep them, ship them through your retention policy, or throw them away. There is no cloud copy.

For regulated work

When the data can’t leave.

Local execution turns a hard compliance problem into a non-problem. Vibn runs the model, the tools, and the transcripts on the same machine you already trust with the file. Nothing goes to a SaaS inference provider, because there isn’t one.

Legal

Draft, redline, and search privileged documents without a cloud processor in the loop. Attorney-client material never leaves the device.

  • NDAs and LOIs
  • Discovery search
  • Contract redlines

Healthcare

Summarize charts, draft notes, and pattern-match across PHI on a workstation you already control. No third-party processor, no BAA gymnastics.

  • Chart summaries
  • Intake drafting
  • Coding assistance

Finance & compliance

Reconcile statements, parse filings, and prep audit packets against material that can’t cross a network boundary. Logs stay in your transcripts directory.

  • Audit prep
  • Filing review
  • Internal reconciliations

Vibn doesn’t certify HIPAA, SOC 2, or any regulation on your behalf. It removes the part of the stack that usually breaks them: remote inference. Your IT and compliance teams own the device; Vibn inherits whatever controls already protect it.

Models

Pick a model. Switch any time.

Pick a coder model for code, a generalist for prose, a tiny one for fast triage. Any Ollama model works out of the box, and your hardware decides the ceiling.

qwen2.5-coder:7b
Default coder · 4.7 GB
fast
qwen2.5-coder:14b
Bigger reasoner · 9 GB
balanced
deepseek-coder-v2:16b
Heavy lifting · 9 GB
smart
llama3.2:3b
Ultra-fast · 2 GB
tiny
gpt-oss:20b
Open source GPT · 12 GB
premium
mistral-small:24b
Long context · 14 GB
context
Tools & MCP

Hand it the keys. Carefully.

Vibn ships with a battle-tested built-in toolset and speaks the Model Context Protocol fluently. Every tool call runs on your machine against your files — nothing crosses the network unless you wire an MCP server that explicitly does. Permission checks gate every call.

~/.vibn/config.json
{
  "permissions": {
    "shell": "ask",
    "edit": "allow",
    "fetch": "deny"
  },
  "mcp_servers": {
    "linear": { "command": "npx",
                "args": ["-y", "@linear/mcp"] }
  }
}
Built-in tools
read_filewrite_fileeditgrepgloblist_dirshellplantaskfetchgenerate_imageread_image
MCP servers (any compatible one works)
Linear
GitHub
Stripe
Sentry
Vercel
Chrome DevTools
Playwright
Plaid

Your work. Your hardware. Your data.

One download. Your machine runs the model, holds the files, and keeps the transcripts. Free, open source, and built to stay that way — whether the work is code, contracts, or charts.

Install on macOS (Apple Silicon)
$ curl -fsSL https://vibn.dev/install.sh | sh

One paste. Downloads the app, drops it in /Applications, opens it.

If you grab the .dmg through the browser, macOS marks it as quarantined and refuses to open it. The install script avoids that. Linux & Windows builds coming soon.