Models & accounts
Which model serves a turn is four questions, not one: the provider, the account, the agentic loop it runs in, and the model itself. The picker asks all four; this is what each of them changes.
On this page (7 sections)
Every credential here belongs to your sandbox. You connect your own subscription and turns spend your allowance — intentic never resells inference, and the platform never holds the token.
Providers
| Provider | What you connect |
|---|---|
| Claude Code | A Claude subscription. The sandbox owns the credential — you sign in to Anthropic, not to intentic. |
| Codex | A ChatGPT subscription. Runs on your ChatGPT plan through the bundled translator. |
| Grok | A SuperGrok subscription. Signs in with xAI's device-code flow; there is no key to paste. |
| Kimi Code | A Kimi Code subscription. Served under the Claude Code loop. |
| A Google sign-in — no subscription. Labelled for the account rather than the model family: this one channel vends Gemini's models alongside Claude and GPT-OSS. The useful answer for anyone who has connected nothing yet. |
Two more sources are open-ended rather than built in:
| Also | What it is |
|---|---|
| A model endpoint | Point the sandbox at any OpenAI-compatible API. Its catalog is read from the server itself, so the picker offers what that server actually serves. |
| An ACP agent | Any agent speaking the Agent Client Protocol, added as a capability. Adding one needs no update to intentic. |
Connect them on Sandbox → Agent. A subscription connects by signing in; nothing asks you for an API key you'd have to rotate by hand.
Several accounts of one provider
A sandbox can hold more than one account per provider — a work Claude and a personal one, two ChatGPT seats — and a turn can name which serves it. Absent, it takes the provider's first.
Naming one matters most where nobody is watching. An automation, a Doorbell, a nightly workflow: when the first account runs out of headroom, or belongs to an organisation that switched the plan off, every unattended run errors against it until someone reads the row. That is why an automation can pin an account and a chat rarely bothers.
The harness
The harness is the agentic loop that serves the turn, and it is a different axis from the provider — the same model behaves differently under each, which is why the picker asks separately.
- Native runs each provider on its own runtime, with its own subscription sign-in.
- Claude Code runs any provider's model through the Claude Code loop instead.
For most providers there is nothing to choose: Claude is always its own Claude Code loop, and Kimi and Google have no native runtime at all — all three run that loop whatever you pick. Only Codex and Grok have a native runtime to switch away from.
The choice is worth making because the loops genuinely differ in what they support, and the app tells you rather than letting you find out mid-turn: a runtime with no mid-turn steering, no clarifying questions, no per-tool approvals, no MCP tools or no reasoning-effort control has those limitations listed against it, and the composer hides the controls that runtime would ignore.
Per-turn controls
Beyond the model there is reasoning effort, extended thinking, and a fast-speed request — each offered only where the runtime actually honours it. A conversation remembers what its last turn ran with, so opening it again seeds the composer from that conversation rather than from whatever another tab last picked.
Permission modes are the other half of this. A runtime that reports the full axis gets per-tool approval cards; one that doesn't gets propose-then-approve or run, and nothing in between — the container is the isolation boundary there, and every tool call inside it is pre-approved.
Models for work you didn't start
Two settings on Sandbox → Agent decide what runs when nobody picked:
- Agent runs. The model for unattended work — an automation's wake, a chore, a documentation or acceptance run, the fix a failed pre-push check proposes. Pinned rather than derived, because that work has to read a failing suite or a container log and repair something, and a wrong guess is billed in whole sessions.
- Quick helpers. The model behind one-click helpers that aren't a conversation, like autofilling a commit message. Defaults to Auto, which resolves from whatever accounts are connected at the moment it's read — so it can never name a provider you have no credential for, and it improves by itself when you connect one.
What it costs
Sandbox → Usage reports the durable spend ledger: a row is appended when a turn ends, grouped by day, provider, account and model, with cache hit rates alongside. Rows are written daemon-side and never edited, so it is a record of what was actually spent rather than an estimate. Individual agent cards carry their own cost and token counts, and a workflow step can be capped in dollars — the one resource you cannot recover after an unattended fan-out.
When a provider refuses a turn — a spent allowance, a plan switched off — the last refusal per provider is kept and shown beside the account, so a fleet that has quietly stopped working says why.
Next
- Automations & workflows: where pinning an account and a model actually pays off.
- Access & sharing: who else may spend this sandbox's allowance.