VibeBrain · Intelligence

One mind.
Any model behind it.

Vibe — your AI teammate — isn't welded to one lab. Point his brain at Anthropic, OpenAI, xAI, Kimi, MiniMax, OpenRouter or any OpenAI-compatible endpoint. Swap the model; Vibe keeps his memory, his tools and his manners.

Get the app See the providers ↓

11 provider presets · 2 API dialects · keys in your Keychain — never in a config file

Providers

Eleven presets, zero endpoint archaeology

Every preset ships with the right base URL, a sensible default model and a hint for where the key comes from. The tricky ones — Kimi and MiniMax coding plans, whose endpoints speak the Anthropic dialect on an OpenAI-style provider — are already wired correctly, because we got them wrong once so you don't have to.

Anthropic

Claude models via the Messages API — including extended thinking when the model supports it.

OpenAI

GPT models over Chat Completions, the dialect half the industry copied.

xAI (Grok)

Grok 4 via console.x.ai, pay-per-token.

Kimi — three flavors

Moonshot pay-per-token API, the Moonshot coding plan, and kimi.com's consumer coding plan — three different key systems, each preset labeled so you never paste a valid key into the wrong door.

MiniMax — two flavors

The international API and the M2 coding plan, whose Anthropic-dialect endpoint is auto-detected.

OpenRouter & beyond

One OpenRouter key unlocks hundreds of models. OpenCode Zen and a fully custom OpenAI-compatible preset cover everything else — including local servers.

The dialect problem, solved quietly. Providers speak one of two API formats — Anthropic's Messages or OpenAI's Chat Completions. VibeBrain speaks both, picks the right one per endpoint (even when a provider's coding plan uses the /anthropic path on an otherwise-OpenAI host), and fixes half-pasted base URLs that would otherwise 404 and masquerade as auth failures.

Memory

He remembers you across every project

Vibe keeps a persistent memory — things you've told him, preferences he's learned — and recalls it in every future conversation, in every workspace. Tell him once that you hate emoji in commit messages; he knows it next month, in a different repo.

▸Never in your repo. Memory lives in Application Support — it can't pollute a project folder or show up in git status.
▸Yours to edit. A memory panel shows every learned fact; delete any of them, any time.
▸Model-independent. Switch from Claude to Kimi tomorrow — the memories come along, because they belong to Vibe, not to a vendor.

Vibe's memory

Leo prefers Tailwind over styled-components learned in acme-web
Always run pnpm, never npm, in this team's repos learned in api-server
Deploys happen Fridays only after the standup learned in acme-web

No API key Borrowed brain

you: what does the checkout flow do?
vibe: borrowing the Claude Code session in Terminal 1…
vibe: Checkout has 3 steps: plan select → Stripe session → webhook fulfilment. The webhook is signature-verified in StripeWebhookController.

Borrowed mode is read-only Q&A — Vibe won't edit files or run commands through someone else's session.

Borrowed brain

No key? Vibe borrows one.

Haven't set up an API key yet? Vibe can still answer questions about your project by borrowing whatever AI CLI is already running in one of your terminals — Claude Code, Codex, whichever. Your existing subscription does the thinking; Vibe does the asking.

It's deliberately conservative: borrowed mode answers questions but never edits files, runs commands or saves memories. For the full agent, add a key — any key.

Trust, verified

Keys in the Keychain. Connections actually tested.

Storage

macOS Keychain, per provider

Every provider key is stored in the system Keychain — one entry each — never in a dotfile, a plist or a config you might accidentally commit.

Verification

A real request, not a regex

The connection tester fires a live request and reports honestly: working, credentials rejected, or “reached something, but not a chat endpoint” — so a wrong base URL never wears an auth error's costume.

Thinking