Every model run, with a receipt.

Solo is free with every capability for one profile. Bring a provider key, a local model, or both; a provider's own bill stays yours.

Paid plans add profiles to an account. Bandicot does not resell model tokens.

Bandicot gives every agent a project, tools, a session that survives restarts, approvals, and a clear place to stop.

The native host owns execution facts. The model proposes work; it does not certify a write, spend, or success.

Keep files, context, plans, and decisions with the project instead of losing them in a disposable chat.

A session is anchored to a project boundary so a pivot starts from an intentional context, not accidental history.

Close a client or reconnect after a pause without turning the work into a brand-new conversation.

The host records session state and clients reattach to it. A browser view does not become the source of truth.

When work can change files or systems, Bandicot makes the decision visible before the effect lands.

The approval boundary fails closed when an action is not allowed. The model cannot grant itself permission.

Use a local endpoint, a private runtime, or a remote provider without rebuilding the project around one model lab.

Provider adapters keep model choice separate from project state. Remote providers receive the prompts sent to them.

See the tokens, cache reads and cost of every turn, per session, profile and model, and the habits behind them.

Bandicot reports what the provider or local runtime exposes. A provider that hides usage shows a dash, never a guess.

Keep a legible trail of the turn and its effects so a pivot can be reviewed instead of guessed at.

The host treats state changes as recorded events. Reverting is itself part of the history.

Pick a small local model from a pinned catalog and install it in one click, checksum verified.

Eight chat models from about 1.25 to 4 GB of memory, on Apple Silicon. Small models suit private, light work; they are not frontier models.

Before a local model downloads, Bandicot checks it against your Mac's GPU memory budget and free disk.

The fit check reads the Metal working-set budget, not physical RAM, and counts weights, context cache and a stated reserve.