← Retour aux guides

Build a Claude Code Agent on a Lean API Budget

Build a Practical Coding Agent with Claude Code

Claude Code is most useful when you treat it as an agent inside a real repository, not as a chat window for isolated snippets. A good coding agent can inspect files, understand local conventions, make a focused change, run tests, and explain what it changed. This tutorial shows how to create that workflow with Claude Code and configure a cost-conscious model connection through 59API.

59API is an AI API relay offering pay-as-you-go access to native official-quality Claude models, including Opus, Sonnet, Haiku, and Fable, along with GPT models. Its Anthropic-compatible access makes it a practical low-cost option for developers who want to use Claude Code frequently without committing to a large prepaid plan.

1. Prepare a Repository for Agent Work

Start with a small, healthy project. Your agent will be more reliable when the repository has a clear test command, a formatter, and a predictable structure. Before opening Claude Code, make sure the project installs and tests successfully on your machine.

• Initialize version control and commit the current working state.
• Add a README with setup, run, and test commands.
• Use a package manager lock file so dependency changes are reproducible.
• Keep secrets in environment variables or ignored local files, never in prompts or source control.

For example, a JavaScript project should have an explicit test script such as npm test. A Python project should document whether it uses pytest, uv, Poetry, or a virtual environment. These details give the agent concrete commands to verify its own work.

2. Configure Claude Code to Use 59API

Install Claude Code using the current installation method documented by Anthropic, then create an API key in your 59API account. Store the key in your shell profile, secret manager, or a local environment file that is excluded from Git.

• Set ANTHROPIC_AUTH_TOKEN to your 59API API key.
• Set ANTHROPIC_BASE_URL to https://api.59api.com.
• Optionally select the Claude model supported by your Claude Code configuration, such as a Sonnet model for everyday implementation work.
• Restart your terminal before launching Claude Code so it receives the new variables.

Use the exact model identifier shown in the 59API dashboard or documentation. Sonnet is usually a strong default for implementation and debugging. Haiku can be useful for quick, low-cost repository questions, while Opus is better reserved for difficult architecture decisions, complex refactors, or investigations with many moving parts. This model routing approach helps control spend without sacrificing quality where it matters.

3. Add a CLAUDE.md Instruction File

Create a CLAUDE.md file in the repository root. Claude Code reads project instructions from this file, making your agent more consistent across sessions. Keep it short, specific, and easy to maintain.

Include the project purpose, supported language version, test command, formatting command, and rules for making changes. State that the agent must inspect relevant files before editing, avoid unrelated refactors, run the narrowest relevant tests first, and summarize modified files and remaining risks. You can also specify boundaries such as “do not change database migrations without asking” or “do not edit generated files.”

A useful instruction is: Before modifying code, describe the implementation plan in two to five bullets and wait for confirmation when the change affects public APIs. This prevents an autonomous workflow from making broad changes based on an ambiguous request.

4. Give the Agent a Bounded First Task

Open the repository in Claude Code and begin with a task that has a measurable result. Avoid requests such as “improve the application.” Instead, describe the behavior, constraints, and acceptance criteria.

• Ask the agent to locate the request validation path.
• Require a new validation rule and a test that demonstrates it.
• Tell it which command to run after the change.
• Ask for a concise final report including files changed and test output.

A strong prompt might be: “Add validation that rejects empty display names. Preserve existing error response formats, add unit tests for valid and invalid input, run the relevant test suite, and do not modify unrelated endpoints.” This gives Claude Code an objective, an area to investigate, and a stopping condition.

5. Use an Inspect-Plan-Edit-Test Review Loop

Reliable coding agents follow a repeatable loop. First, ask Claude Code to inspect the relevant code and identify dependencies. Next, review its plan before it edits files. Then let it implement the smallest viable change. Finally, require it to run tests and show the result.

Do not skip human review. Check the diff, especially around authentication, authorization, data deletion, shell commands, and dependency updates. If the agent proposes a larger refactor, split the work into separate commits. Small, test-backed changes are easier to verify and cheaper to iterate on.

6. Control Cost Without Weakening Your Workflow

Cost management is primarily about reducing unnecessary context and using the right model for each task. Ask focused questions, point to likely directories, and request summaries instead of repeatedly pasting large files. Use a fast model for exploration, then switch to a stronger model for the final implementation or review when needed.

Because 59API provides inexpensive, pay-as-you-go access with no model downgrade, it is well suited to this iterative agent workflow. You can experiment with short planning sessions, tests, and review passes while keeping API usage predictable. If you are ready to connect Claude Code to a lower-cost relay, sign up for 59API and start with a small development budget.

7. Turn Successful Sessions into Reusable Agent Rules

Whenever Claude Code solves a recurring problem correctly, capture the lesson in CLAUDE.md or a project-specific checklist. Add conventions for naming, tests, migrations, API compatibility, and deployment checks. Over time, these local rules turn Claude Code from a general assistant into a coding agent that understands how your team actually ships software.

Tarifs · Docs