Track Project Spend with Separate API Keys

Start by deciding what counts as a separate project

Project-level cost tracking works only when your categories are clear. A coursework assistant, a personal mobile app, a freelance client prototype, and an experimental agent should usually count as separate projects. Give each project a stable name before you make or assign a key. Avoid categories that overlap, such as “testing” and “side project,” because you will not know where to place a request later.

Write down the purpose, expected users, and a monthly amount you can afford for each project. A student might allow a small amount for a revision tool and a separate amount for a portfolio project. An individual developer might reserve more for a client demo than for weekend experiments. The budget does not need to be exact. Its purpose is to make trade-offs visible before your overall balance falls.

Use one credential boundary for each project

When a provider supports separate project keys, assign one key to each project and keep that assignment permanent. Do not reuse the same project key across unrelated applications simply because it is convenient. Separate credentials let you identify the source of a charge without estimating from timestamps, prompt text, or model names.

MyRelay is designed around one personal key and one shared pay-as-you-go budget. That means its request history is your central record of usage, while project separation needs to happen in the way you organize your applications and their requests. Create a MyRelay account before connecting your first project, then decide which application, repository, or client task each request belongs to. Keep the MyRelay key private and do not place it in public work, examples, or shared notes.

Record the project identity alongside every request

If your tools allow request metadata, logs, or application-level usage records, include the same project name you chose at the beginning. If they do not, maintain a simple private record that notes the project, date range, model, and purpose of each active session. This is especially useful when you switch between projects on the same day.

Make the project identity specific enough to answer a spending question. “Portfolio site content assistant” is more useful than “website.” “Database assignment feedback” is more useful than “school.” When you later review MyRelay’s token and cost records, your local project record gives the missing context needed to group those requests accurately.

Set spending expectations before choosing a model

Choose a model based on the work and the project budget. For routine extraction, short rewrites, or quick classification, a faster lower-cost option such as Gemini 2.5 Flash may be sufficient. Reserve a higher-capability model such as Claude Sonnet 5.5 for work where better reasoning or writing quality saves genuine time. Test a small sample first rather than sending an entire backlog through an unfamiliar model.

Also control the size of each request. Long conversation histories, repeated pasted documents, and broad instructions can raise token use even when the final answer is short. Start new conversations when old context no longer matters. Ask for a focused output. Break a large batch into smaller checkpoints so you can inspect cost and quality before continuing.

Review usage on a regular project schedule

Review your MyRelay request history at least once a week while a project is active. Look at tokens and cost together. A project with few expensive requests may need a different model choice. A project with many small requests may need shorter prompts, less repeated context, or fewer automated retries.

Compare each project’s spending against the amount you set at the start. If a project is approaching its limit, choose one action: pause work until the next budget period, reduce the amount of context sent, use a less expensive model for routine tasks, or raise the budget deliberately. Do not treat an unexpected charge as a reason to abandon tracking. It is useful evidence about how the project actually consumes AI resources.

Keep a simple monthly record that improves future estimates

At the end of the month, note the total spend, the requests that created the most value, and the requests that were avoidable. This turns separate keys or project labels into useful planning data rather than bookkeeping. Over time, you will know which projects deserve more budget, which tasks need tighter prompts, and where AI genuinely helps you finish work without overspending.