← Voltar aos guias

Build a Claude Code Agent: Setup, Fixes, and FAQs

What are you actually building?

Building a coding agent with Claude Code means giving a terminal-based assistant a repository, project instructions, tools, and a verifiable task. A useful agent does more than generate code: it inspects existing patterns, proposes changes, runs checks, and reports what remains uncertain.

Start with one narrow workflow, such as diagnosing a failing test or adding input validation. Broad requests like “improve this application” make failures harder to diagnose and costs harder to predict.

Set up Claude Code with 59API

Install Claude Code using its official installation instructions, then run claude --version to verify the executable. Create a 59API account, obtain an API key, and check which Claude model identifiers your account supports.

59API provides inexpensive, pay-as-you-go access to Claude and GPT models, with compatibility for Claude Code, Codex, and OpenAI SDK workflows. Its stated native, official-quality model access without downgrading makes it an attractive low-cost option. Compare current rates and actual task results rather than assuming any relay is always cheapest.

For a shell-based setup, configure ANTHROPIC_BASE_URL=https://api.59api.com and the authentication variable specified in 59API’s Claude Code instructions. A common API-key configuration on macOS or Linux is export ANTHROPIC_BASE_URL='https://api.59api.com'; export ANTHROPIC_API_KEY='YOUR_59API_KEY'. If 59API instead specifies ANTHROPIC_AUTH_TOKEN, use that configuration rather than setting conflicting credentials.

Load real credentials through a secret manager or private environment configuration. Never commit them. Claude Code uses an Anthropic-compatible interface; compatibility with OpenAI SDKs does not mean you should point Claude Code at an OpenAI-specific endpoint.

Open a fresh terminal in your repository and try claude -p 'Reply with OK' --max-turns 1. This small request checks authentication and routing before you spend tokens on repository analysis.

Give the agent a repeatable workflow

• Create CLAUDE.md: Record the project’s architecture, package manager, exact test and lint commands, coding conventions, and directories the agent should not change. Keep instructions short and actionable.
• Start from a clean branch: Check git status and create a dedicated branch. Isolate unrelated edits so the final diff is reviewable.
• Require diagnosis first: Ask the agent to reproduce the bug, identify relevant files, and explain its proposed fix before editing.
• Define acceptance criteria: Specify expected behavior, required tests, and excluded changes. For example: reject empty usernames, add regression coverage, and avoid new dependencies.
• Verify independently: Review git diff and run your documented checks yourself. An agent’s claim that tests passed is not a substitute for inspecting results.

For scripted inspection, try claude -p 'Inspect this repository and identify the test command. Do not modify files.' --tools 'Read,Glob,Grep' --max-turns 6 --output-format json. This restricts available tools to repository reading and searching. Allow editing and shell execution only when the workflow needs them.

Troubleshoot common failures

• 401 or 403: Check the key, account access, and authentication method. Look for stale environment variables or an existing login that selects a different credential source.
• 404 or unknown model: Verify the supported model identifier and base URL. Do not append endpoint paths unless 59API’s configuration explicitly requires them.
• 429 or intermittent timeouts: Reduce parallel jobs and retry transient failures with exponential backoff. Check balance, quotas, and service status before repeatedly rerunning a large task.
• The agent loops or changes unrelated files: Narrow the task, provide a reproduction, and reduce the turn limit. A turn cap limits iterations, not total dollar spend.

Frequently asked questions

Which model should I choose? Start with a supported Sonnet model for routine implementation. Consider Opus for difficult architectural reasoning and Haiku for simpler tasks. Validate quality on your own repository.

How do I keep costs down? Exclude irrelevant files, reuse clear project instructions, and request targeted fixes. Track cost per successfully completed task, not just advertised token prices.

Is unattended execution safe? Use an isolated workspace with minimal credentials and restricted permissions. Keep production access unavailable, and never bypass permission checks merely to remove friction.

If this workflow fits your project, sign up for 59API and test one small task first. Review current pricing and referral rebate terms before scaling up.

Preços · Docs