Three AI CLIs, One Desktop, a $30 Budget

01 / The $30 Receipt Starts With Hardware You Own

My proposed desktop setup starts with a purchasing rule: buy nothing until the existing machine becomes the bottleneck. Keep your current computer, monitor, keyboard and editor. Open three terminal tabs labeled Claude Code, Codex and Gemini CLI. Together means available side by side, not automatically running three agents on every task.

The numbers below are a spending plan, not a tested performance claim or a provider quotation. All amounts are USD. Existing internet service and electricity remain household expenses; they are not magically free, but no incremental charge is assumed here.

• ■ Existing computer and peripherals: $0 in new purchases.
• ■ Additional monitor, GPU and desktop accessories: $0 allocated.
• ■ Claude Code API allowance: $12.
• ■ Codex API allowance: $8.
• ■ Gemini CLI API allowance: $4.
• ■ Shared troubleshooting and retry reserve: $6.
• ■ Total planned incremental spending: $30.

Those allowances belong in your own ledger. They are not subscription prices, guaranteed workloads or automatically enforced limits.

02 / One Relay Key, Three Separate Client Configurations

CheapRelay describes its service as pay-as-you-go access through one key, with no monthly fee or subscription and billing for real tokens, including cache reads and writes. That fits this budget structure: allocate money to work rather than three recurring commitments. Current model tariffs were not independently verified for this article, so none are presented as quoted prices.

Shared credentials do not mean identical configuration. Claude Code supports gateway routing through ANTHROPIC_BASE_URL, but the gateway must expose a supported API format. Use CheapRelay’s Claude Code instructions rather than assuming an OpenAI-compatible endpoint is interchangeable with an Anthropic-format endpoint. ([code.claude.com](https://code.claude.com/docs/en/llm-gateway))

For Codex, define a custom provider with a base_url and an env_key naming the environment variable containing your relay credential. Keep the actual secret out of the provider definition. ([learn.chatgpt.com](https://learn.chatgpt.com/docs/config-file/config-reference?utm_source=openai)) Put provider configuration in the user-level config.toml, not the repository’s project-local configuration. ([learn.chatgpt.com](https://learn.chatgpt.com/docs/config-file/config-advanced?utm_source=openai))

Codex requires Responses API compatibility; a working chat-completions request alone does not establish that compatibility. Verify streaming and a small tool call before assigning repository work. ([learn.chatgpt.com](https://learn.chatgpt.com/docs/enterprise/roll-out-a-gateway?utm_source=openai)) Gemini CLI exposes GOOGLE_GEMINI_BASE_URL for a custom endpoint; follow the relay’s instructions for its native route and avoid duplicating a version suffix. ([geminicli.com](https://geminicli.com/docs/reference/configuration/?utm_source=openai))

03 / Give Each Terminal a Job Before Spending $1

My suggested division is Claude Code for a difficult code change, Codex for a separate implementation or test task, and Gemini CLI for a bounded explanation or document review. These are workflow choices, not rankings or benchmark results.

Keep one agent responsible for editing a given working directory at a time. Ask another for review only when you can name the uncertainty it should resolve. Three independently generated answers are not automatically worth three bills.

For the first experiment, set a combined $1 checkpoint across all three clients. Record the opening balance, make one small request per client, and reconcile the debits. This checks routing and accounting before a long autonomous session.

04 / A Sample Workday Receipt Without Fake Token Prices

Here is an illustrative ledger, not observed usage or a price estimate:

• ■ Claude Code task: assumed debit of $0.35.
• ■ Codex test task: assumed debit of $0.20.
• ■ Gemini CLI review: assumed debit of $0.10.
• ■ One retry: assumed debit of $0.15.
• ■ Example daily total: $0.80.
• ■ Twenty identical days: $16, leaving $14 of the $30 budget.

Replace every assumed debit with actual account charges. At $1.50 per day, twenty working days consume the entire $30. That arithmetic is useful; promising a fixed number of coding sessions without measured usage is not.

05 / Make the Receipt Control Tomorrow’s Work

Track the date, client, task, selected model, reported token categories and final charge. Reconcile against billed usage rather than counting prompts. Keep the $6 reserve separate, and stop to review spending before transferring it into an exhausted allowance.

Sign up for CheapRelay when you are ready to run that first $1 routing check. The frugal result is not three agents constantly talking: it is three available tools, a deliberate task split and a receipt you can explain.