Skip to content

Best use cases

mintbot works out of the box. Agents bought with bundled credit start chatting on day one — the default mintbot proxy routes every turn through our pooled LLM credit and deducts the cost from the agent's balance. No API key, no subscription, no setup.

But that convenience has a price: the proxy bills retail per token, which makes it the most expensive way to run an agent long-term. The cheapest way is to plug in an LLM subscription you (probably) already pay for — and you don't need the priciest plan to do it. A $20 / month ChatGPT Plus (or Anthropic Claude Pro) subscription, wired in through the bundled Codex or Claude Code CLI, is the sweet spot for almost every agent.

This page covers when to use which provider, why a $20 subscription is the best-value way to run an agent, when it's worth upgrading, and how to wire it up in 60 seconds.

The four ways an agent talks to an LLM

A mintbot agent picks the active LLM provider from the LLM providers modal in the web panel. Four routes exist:

Route Who pays When to pick it
mintbot proxy (default) Mintbot meters per token; the credit balance on your agent goes down as it answers. Trying it out. Short tasks. No subscription to plug in.
Claude Code OAuth Your Anthropic Pro / Max subscription. Unmetered within Anthropic's per-account quotas. You already pay Anthropic. Long tasks heavy on Claude.
OpenAI Codex OAuth Your ChatGPT Plus / Pro / Team / Enterprise subscription. Unmetered within OpenAI's per-account quotas. You already pay OpenAI — or start fresh with a $20 Plus plan. The most cost-efficient way to run an agent.
Bring your own API key (BYOK) Your own paid API account at any provider (Anthropic, OpenAI, OpenRouter, …). Metered, you pay the API price. You want a specific model that mintbot doesn't ship in the default catalog.

The mintbot proxy is convenient but it's the most expensive option per token: you pay mintbot's retail price on each call. If you're going to use the agent for any real work — drafting, coding, research — wire in your own subscription on day one.

Start here: ChatGPT Plus ($20 / mo) is the sweet spot

The best first move is the entry-level ChatGPT Plus subscription at $20 / month. It gives you:

  • Codex CLI access with quotas that comfortably cover a single agent doing light-to-medium work — drafting, coding, research, the day-to-day.
  • The same flagship models (currently GPT-5), tools, web search and file uploads as the chat product.

Wired into your agent, that $20 replaces the metered mintbot proxy for the vast majority of turns. Compared with spending the same $20 on raw API tokens, Plus buys you far more usage — and for a starter agent it's usually all the capacity you'll ever need. Start here, and only move up if you actually hit the ceiling.

Anthropic works the same way: if you'd rather stay on Claude, the Claude Pro plan (also ~$20 / mo) plugs into Claude Code exactly the same way. Pick whichever subscription you already hold — the Codex route is a recommendation for cost, not a religious preference.

Upgrade to ChatGPT Pro ($200 / mo) once you outgrow Plus

Once the agent genuinely has a lot on its plate — running all day, every day, heavy long-context coding, effectively doing the work of a small IT team — ChatGPT Pro at $200 / month is the natural next step:

  • Extended access to OpenAI's flagship models (currently GPT-5)
  • Much larger Codex quotas — so the effective price per unit of LLM work drops even further
  • Everything Plus has, with far more headroom

Pro is the most LLM you can buy for the dollar at the top end. But that value only shows up once your usage is genuinely high — for a fresh agent, the $20 Plus plan wins on value. Treat Pro as the upgrade you reach for when Plus starts running out of quota, not the place to start.

Higher tiers work the same way

ChatGPT Team and Enterprise (the paid org plans) also expose Codex with their own quotas. The login flow described below is identical — the panel doesn't care which tier you signed up for, as long as the OAuth grant includes Codex access. The same goes for Anthropic Max if you're on the Claude side.

Wire up Codex in 60 seconds

  1. Open your agent's web panel (from the Telegram bot's 🌐 Open agent web panel menu).

  2. In the top bar, click LLM providers & credit.

    Agent panel top bar with the “LLM providers & credit” pill highlighted
    Click the LLM providers & credit pill in the top bar.

  3. Click Connect ChatGPT account.

  4. The panel pops a URL. Open it in another tab, sign in to your ChatGPT account, confirm the device grant.
  5. The panel detects the completion, shows Codex as logged in (for example, your Plus / Pro / Team / Enterprise account), and selects it as the active provider automatically.

That's it. The next message you send the agent goes through your ChatGPT subscription's quota.

What happens under the hood

The bundled codex login --device-auth flow lands an OAuth token in ~/.codex/auth.json on the agent VPS. Hermes (the runtime) talks directly to chatgpt.com using that token. There is no proxy in the middle; mintbot never sees your prompts or your OpenAI account state.

When to keep using mintbot proxy

In some cases it still makes sense to use mintbot credit:

  • Trial / demo — no subscription, just kick the tires.
  • Managed fallback — when a connected subscription or BYOK account is out of quota, paused, or temporarily broken, you can switch back to mintbot credit and keep working without touching third-party accounts.

For your own personal driver agent — the one you DM from Telegram all day for whatever work comes up — the Codex (or Claude Code) route is the right answer.

Quick-start cheat sheet

1. Start with ChatGPT Plus at https://chatgpt.com/pricing  ($20/mo)
2. Open your agent's panel  (Telegram bot → 🌐 Open agent web panel)
3. LLM providers & credit → Connect ChatGPT account
4. Complete the OAuth flow in the browser tab
5. Send the agent a message — it now answers through your ChatGPT Plus quota
6. Running the agent hard? Upgrade to ChatGPT Pro ($200/mo) for much larger quotas — same wiring.