withvibe
Concepts

Models & API keys

withvibe runs on Claude. You can pin a specific model or let it pick per turn, and you can decide whose API key pays for a run — the individual's, the workspace's, or the server's.

The models

Pick a model per env, or set a default for the whole workspace. The menu is the same in both places:

  • Auto — a cheap classifier picks Opus, Sonnet, or Haiku per turn based on how hard the turn is. The default behavior.
  • Opus 4.7 — highest quality, slowest, most expensive.
  • Sonnet 4.6 — the balanced default: strong reasoning at moderate cost.
  • Haiku 4.5 — fast and cheap; best for short, simple turns.
Claude today
withvibe is Claude-only right now. The agent layer is designed to be provider-agnostic, so other models are on the roadmap — but there is no non-Claude engine today.

How Auto picks a model

On a turn set to Auto, withvibe runs a fast, cheap Haiku classifier over your prompt and sorts the turn into one of three tiers. Pinning a concrete model skips this entirely — pin Opus and every turn is Opus.

  • simple → Haiku — short follow-ups, status questions, trivial edits, lookups, formatting, single-file reads.
  • medium → Sonnet — typical feature work, bug investigation, refactors. This tier deliberately also covers surveys, audits, and codebase exploration — reading a lot of files is mechanical work, not a reason to reach for Opus.
  • hard → Opus — reserved for reasoning that no amount of file-reading replaces: architecture from scratch, novel algorithms, multi-system trade-offs, security-correctness review, genuinely hard debugging.

The classifier is robust by design: if there's no API key, or it fails or returns something unparseable, the turn quietly falls back to Sonnet. Repeated identical prompts are cached briefly so you don't pay to classify the same thing twice.

Where the model setting lives

Resolution is layered, most specific wins:

  1. Per env — an env's own model choice, if set.
  2. Per workspace — the workspace default, otherwise.
  3. Auto — if neither is set.

API keys: personal, workspace, server

Runs need an Anthropic API key. withvibe resolves it per turn, based on who sent the prompt:

  1. Your personal key, if you've set one. Your turns then bill to your own Anthropic credits.
  2. The workspace key, as the shared fallback for anyone without a personal key.
  3. The server's ANTHROPIC_API_KEY, as the last resort.

Setting your personal key

Open Account and find the Anthropic API key card. Paste a standard key (sk-ant-api…) or an OAuth token (sk-ant-oat…) and save.

  • The key is stored securely and never shown again — only a …last4 hint is displayed.
  • Remove it any time to fall back to the workspace key.
  • When set, your sessions use it in preference to the workspace key.
Why personal keys
Personal keys let usage and spend land on the person doing the work instead of pooling on one shared account — handy for cost attribution, or for letting teammates use their own Claude plan inside a shared env.