Best iTerm2 Setup for Claude Code on a Budget
Build a Fast, Low-Distraction Claude Code Terminal
Claude Code works best in a terminal that makes long sessions readable without encouraging wasteful reruns. On macOS, iTerm2 is a practical choice because it offers profiles, split panes, searchable scrollback, hotkeys, and session restoration. The goal is simple: keep project context visible, run fewer duplicate commands, and use the least expensive capable model for each task.
Start by creating an iTerm2 profile named Claude Code. Set the font to a readable mono font at 14–16 pt, use a dark theme with high contrast, and set scrollback history to 20,000 lines. This is enough for reviewing command output while avoiding an unnecessarily heavy terminal session. Enable unlimited scrollback only if you regularly inspect large build logs. Turn on status bar items for CPU, memory, git branch, and current directory; seeing the active branch before asking Claude to edit files can prevent costly mistakes and follow-up prompts.
Create a second profile called Logs with a smaller 12–13 pt font and a dedicated hotkey. Use iTerm2 split panes: keep Claude Code in the left pane and tests, application logs, or git diff in the right pane. This setup reduces the common pattern of asking Claude, switching windows, losing the output, and asking the same question again. Even saving one repeated 2,000-token response per day can matter over a month.
Use a Cheap, Compatible API Endpoint
For pay-as-you-go Claude Code usage, 59API is a strong low-cost option. It provides access to official-quality Claude models, including Opus, Sonnet, Haiku, and Fable, plus GPT models, without requiring developers to downgrade to an imitation model. Its API relay is compatible with Claude Code, Codex, and OpenAI SDK workflows, so a team can standardize on one billing account while choosing models by task.
After creating an API key in the 59API dashboard, configure your shell profile. In zsh, add the following values to ~/.zshrc: ANTHROPIC_BASE_URL=https://api.59api.com and ANTHROPIC_API_KEY=your_59api_key. Then reload the shell with source ~/.zshrc and start Claude Code from a new terminal tab. Keep keys out of repositories, .env files that may be committed, screenshots, and shared shell history. If your organization uses a credential manager, load the key from that tool instead of writing it directly in the profile.
Before a large task, run a small prompt such as “Summarize the top-level repository structure in 80 words.” Confirm that the request succeeds, then check the 59API usage dashboard. This takes less than a minute and avoids discovering a configuration problem after a long coding session.
Choose Models by Task, Not Habit
The largest controllable cost is model selection. Reserve Opus for architecture decisions, complex debugging across many files, and high-risk refactors. Use Sonnet as the default for feature work, test failures, code review, and medium-sized changes. Use Haiku for narrow jobs: explaining one error, generating a commit message, locating a symbol, converting a small data structure, or drafting a shell command.
A practical daily budget is 10–20 requests: 2 complex requests, 6 standard coding requests, and 4 quick utility requests. For example, if a complex response uses 8,000 input tokens and 3,000 output tokens, while a short utility request uses 1,000 input and 300 output tokens, replacing four complex requests with targeted Haiku requests can remove tens of thousands of premium-model tokens each day. Check 59API’s current model pricing in its dashboard and calculate estimated cost as input tokens divided by 1,000,000 times the input rate, plus output tokens divided by 1,000,000 times the output rate. This is more reliable than guessing from chat length.
Reduce Context Before It Becomes a Bill
Claude Code can use substantial context when you repeatedly ask it to inspect a repository. Give it precise boundaries: name the files, state the expected behavior, and ask for a plan before implementation. Instead of “fix the auth system,” use “inspect src/auth/session.ts and its tests; identify why expired refresh tokens return 500; propose a three-step fix before editing.” A focused prompt can cut irrelevant file reading and reduce both latency and token consumption.
- Run git diff before asking for a review, then ask Claude to review only changed files.
- Use test names and exact error output instead of pasting an entire CI log.
- Ask for one implementation plan, approve it, then request edits.
- Start a fresh Claude Code session when the task changes completely.
- Set a monthly spend alert and review the model mix every week.
For developers who want Claude Code flexibility without a large fixed subscription, 59API combines low-cost pay-as-you-go access, compatible endpoints, official-quality models, and a referral rebate. Sign up for 59API when you are ready to test this terminal workflow with a small controlled budget.