Models & Subscriptions#

No Happy plan for model access, no key to paste. Sign in to Claude Code, Codex, or the Grok CLI on the machine and Happy uses those credentials.

Availability is machine-wide, because the daemon owns the catalog. Requests go to the vendor with your credentials; Happy never proxies them.

Switching mid-session#

Change a session's model or reasoning effort at any time, from the app or the phone. The transcript stays. Each provider's opaque data (reasoning, response items) is stored verbatim and handed back unchanged, so a switch does not corrupt history. The new model gets its own vendor's prompts and tools; permission mode, sandbox, and workspace do not change.

Several models in one task#

Subagents each pick their own model and effort; nothing is inherited. One session can plan on one provider, implement on another, review on a third. See Agents That Never Die.

Compaction#

Happy decides when; the provider decides how, using its native compaction or its own compaction prompt. The result is a visible message in the transcript naming what it replaced.

Multiple accounts#

[providers.work_codex]
type = "codex"
enabled = true
auth_file = "/Users/me/.codex-work/auth.json"

Named instances appear in the picker as separate providers. To rotate between same-type accounts automatically, pool them behind a smart provider and hide the members:

[providers.codex]
hidden = true

[providers.work_codex]
type = "codex"
enabled = true
hidden = true
auth_file = "/Users/me/.codex-work/auth.json"

[providers.codex_pool]
type = "smart"
strategy = "round_robin"
providers = ["codex", "work_codex"]
enabled = true

A new agent or subagent starts on a random member and stays there; an auth or quota-exhaustion error moves it to the next one. hidden = true only removes an account from the picker, and a pool still routes to it. See Configuration.

Usage#

Provider quotas are polled about every fifteen minutes and shown in the app. They are observations, not billing records.