opencode2-chatgpt-pool

Pool multiple ChatGPT Pro/Plus OAuth accounts in OpenCode V2 and move to another account when one is rate limited.

The plugin deliberately reuses OpenCode's built-in OpenAI login. It does not copy the OAuth flow or replace the provider implementation:

1. OpenCode logs in and stores its normal active OpenAI credential.
2. The plugin captures each ChatGPT credential into a local pool before the next login replaces it in OpenCode's single-credential store.
3. Immediately before a Codex request, the plugin supplies a pooled bearer token and the matching chatgpt-account-id header.
4. A 429 cools that account. When another account is ready, the plugin removes the provider's long Retry-After; OpenCode's own retry reissues the request through the next account after its normal short backoff.

The default strategy is sticky: an account stays selected until it is limited. That preserves session affinity and prompt-cache locality better than rotating every request.

Install

The V2 plugin API is beta and host-coupled. This version is built and verified against @opencode-ai/plugin@0.0.0-beta-17595 (opencode2 --version → v0.0.0-beta-17595).

Always install with the explicit @beta tag. npm assigned latest during the one-time package bootstrap, but the release workflow never publishes or moves that tag while OpenCode V2 remains beta.

Install the beta-pinned release in both config files. The server plugin routes requests and owns the pool; the TUI plugin provides /chatgpt-pool and switch toasts.

// ~/.config/opencode/opencode.jsonc
{
  "plugins": ["opencode2-chatgpt-pool@beta"],
}

// ~/.config/opencode/cli.json
// Use the package name only; the TUI resolves its ./tui entrypoint.
{
  "plugins": ["opencode2-chatgpt-pool@beta"],
}

OpenCode installs package dependencies into its plugin cache. Restart the background service after changing plugins:

opencode2 service restart

For development from a local checkout, install its dependencies and point both OpenCode halves at their source entrypoints:

cd /path/to/opencode2-chatgpt-pool
bun install

// ~/.config/opencode/opencode.jsonc
{
  "plugins": ["/absolute/path/to/opencode2-chatgpt-pool/src/index.ts"],
}

// ~/.config/opencode/cli.json
{
  "plugins": ["/absolute/path/to/opencode2-chatgpt-pool/src/tui.ts"],
}

Add accounts

With the plugin running:

1. Use /connect, choose OpenAI, then ChatGPT Pro/Plus (browser) and log into the first account. An already-connected account is captured when the plugin starts, so this step can be skipped for it.
2. Run /connect again and log into another account.
3. Repeat for any additional accounts.

OpenCode still shows only its latest OpenAI connection. The plugin retains the earlier accounts in its pool.

Run /chatgpt-pool to list or remove accounts. The server tools also let other OpenCode clients ask an agent to manage them. Both interfaces expose only account IDs, emails, and availability; OAuth tokens never appear. OpenAI must be disconnected through /connect before an account can be removed, otherwise OpenCode could capture the active credential into the pool again.

When routing changes accounts, the TUI plugin shows a toast without adding anything to the session or model context.

Accounts are stored at $XDG_DATA_HOME/opencode2-chatgpt-pool/accounts.json, defaulting to ~/.local/share/opencode2-chatgpt-pool/accounts.json. The directory and file are forced to modes 0700 and 0600. Tokens remain plaintext, like OpenCode's own local credential store, so protect your user account and backups.

Options

{
  "plugins": [
    {
      "package": "/absolute/path/to/opencode2-chatgpt-pool/src/index.ts",
      "options": {
        "strategy": "sticky",
        "rateLimitCooldownMs": 60000,
        "authFailureCooldownMs": 300000,
        "refreshBufferMs": 300000,
        "tools": true,
      },
    },
  ],
}

Option	Default	Behaviour
strategy	"sticky"	"sticky" changes account only after a failure; "round-robin" advances every request.
rateLimitCooldownMs	60000	Fallback cooldown when a 429 has no usable retry header.
authFailureCooldownMs	300000	Cooldown after refresh failure, 401 or 403.
refreshBufferMs	300000	Refresh a token this long before expiry.
tools	true	Expose chatgpt_pool_list and chatgpt_pool_remove. Disable to save request schema tokens.

Provider retry headers override the configured rate-limit fallback. If every account is cooling down, the plugin preserves the provider's original retry delay and lets OpenCode handle it normally.

Development

Requires Bun:

bun install
bun run check

Terms

Use only accounts you own and follow OpenAI's applicable terms. The plugin is not intended to bypass account restrictions or share subscriptions between people.

License

MIT