nc://account-pool

Account Pool: several Claude and Codex accounts in one pool -- each one's usage windows, which account to send the next request to, and a park switch that keeps a spent one out of the rotation. A Nodecode rewrite of roqianjas/hermes-account-pool by roqianjas.

One GET per account, to the vendor that owns that account's token: Anthropic's OAuth usage endpoint for a Claude subscription, ChatGPT's usage endpoint for a Codex one. Nothing is shared between accounts and nothing goes anywhere else. A wake -- one tiny request that starts an idle account's 5-hour window -- goes to that same vendor. Park decisions are a plain text file under the home.

Get it

Tell the agent "add the account-pool cell", or run nodecode add nc://account-pool.

Set it up

List one entry per account in the cell's accounts, as label:provider:ENV_VAR, with the provider anthropic or openai-codex. label:ENV_VAR is a Claude account, which is the common case:

{
  "account-pool": {
    "accounts": [
      "claude-main:anthropic:CLAUDE_MAIN_TOKEN",
      "claude-backup:anthropic:CLAUDE_BACKUP_TOKEN",
      "codex-main:openai-codex:CODEX_MAIN_TOKEN"
    ]
  }
}

Each entry names the environment variable holding that account's access token: the token of that account's own sign-in (a Claude Code sign-in's OAuth access token, or a ChatGPT Plus/Pro sign-in's). The config file holds no secret, only the names. An account whose variable is unset still appears in the report, saying which variable to export; with none set the cell refuses at start and names them all.

The rest of the section: cache_seconds (60) bounds how long one account's answer is reused, timeout_seconds (12) one usage read, wake_timeout_seconds (60) one wake, wake_model (claude-haiku-4-5) and codex_wake_model (gpt-5.4-mini) are the models a wake sends to, and state_file (account-pool/parked.txt) is where park decisions are kept. See config.example.jsonc.

Use it

Ask, and the model calls the tool:

By hand: /account-pool prints the report, and /account-pool best, /account-pool park LABEL, /account-pool unpark LABEL, /account-pool wake LABEL and /account-pool refresh do the rest. From eval:

(account-pool:report)                    ; the text /account-pool prints
(account-pool:report :account "claude-main" :refresh t)
(account-pool:best)                      ; most headroom first
(account-pool:windows "codex-main")      ; :windows, :details, :reason as a plist
(account-pool:park "claude-backup" :hours 6)
(account-pool:wake "claude-backup")

An account that answers a refusal, a limit already reached, or no window is one line of the report, not a failure. A missing label reads ERROR: ACCOUNT-POOL-ERROR and lists the ones the pool holds.

What it does not do

The original also adds and verifies sign-ins through Hermes's own credential pool, rotates a running session between accounts, keeps token counts per account and session, schedules weekly parks, runs a five-minute balance timer and an alerting channel, and draws a Hermes Desktop page. Here the accounts are the section's list, the switch is park, and the report is text.

MIT, see LICENSE.