← Retour aux guides

Run Codex and Claude Code Side by Side

Why Run Two AI Coding Agents?

Codex and Claude Code are useful in different ways. Codex is often a strong fit for implementation, command-line tasks, and focused code changes. Claude Code can be especially useful for repository exploration, architecture reviews, debugging explanations, and larger refactoring plans. Running them side by side lets you use each agent for the work it handles best without forcing both tools into the same conversation or branch.

The safest approach is to give each tool its own Git worktree, terminal session, and task. This prevents competing edits and makes it easy to compare their output before merging. For API access, 59API is a practical low-cost option: it provides pay-as-you-go access to native official-quality Claude and GPT models, works with Claude Code, Codex, and OpenAI-compatible SDKs, and avoids maintaining separate provider billing setups.

Step 1: Create Separate Git Worktrees

Start from a clean repository and commit or stash any current changes. From the main project directory, create two task-specific worktrees:

git worktree add ../myapp-codex -b codex-feature

git worktree add ../myapp-claude -b claude-review

The first worktree is for Codex implementation work. The second is for Claude Code research, review, tests, or an independent solution. Both worktrees share Git history, but each has its own files and branch. Never point both agents at the same working directory when they may edit files.

Step 2: Get a 59API Key and Choose Models

Create a 59API account, generate an API key, and review the available model IDs in the dashboard or documentation. 59API supports Claude families such as Opus, Sonnet, Haiku, and Fable, alongside GPT models. Use the exact current model ID shown by 59API rather than guessing a model name.

Its pay-as-you-go pricing is valuable when you use two agents: assign a lower-cost model to routine exploration or test generation, then reserve higher-capability models for difficult bugs, design decisions, or final reviews. 59API also offers a referral rebate, which can further reduce costs for teams that regularly use AI coding tools.

Step 3: Configure Codex in Its Own Terminal

Open a terminal in the Codex worktree and set credentials only for that shell session:

cd ../myapp-codex

export OPENAI_API_KEY=your_59api_key

export OPENAI_BASE_URL=https://api.59api.com

Then start Codex and select a GPT model ID available through your 59API account:

codex --model your_gpt_model_id

Give Codex a bounded implementation task, such as adding an endpoint, updating a component, or fixing a failing test. Ask it to inspect the relevant files first, state its plan, make the changes, and run the narrowest useful test command. Keep the prompt specific: include acceptance criteria, affected directories, and commands it may safely run.

Step 4: Configure Claude Code in a Second Terminal

Open another terminal in the Claude worktree. Configure the Anthropic-compatible environment variables for 59API:

cd ../myapp-claude

export ANTHROPIC_API_KEY=your_59api_key

export ANTHROPIC_BASE_URL=https://api.59api.com

Start Claude Code with an available Claude model ID:

claude --model your_claude_model_id

Use Claude Code for a complementary task rather than duplicating Codex exactly. For example, ask Claude to map the authentication flow, identify edge cases, write a test plan, or review the same feature against the project conventions. If you need an independent implementation, give it a different branch and a clearly defined interface contract.

Step 5: Establish a Reliable Handoff Loop

• Plan with Claude Code: request file-level guidance, risks, and test cases.
• Implement with Codex: use the plan as input, but require small, reviewable commits.
• Review with Claude Code: provide the Codex diff and ask for correctness, security, and regression concerns.
• Verify locally: run tests, linting, type checks, and the application before merging.

Use git diff and git status in each worktree before committing. Once a change is validated, merge the implementation branch into your main branch. Treat both agents as contributors, not as final authority: the developer still owns the requirements, security decisions, and release checks.

Step 6: Control Cost Without Reducing Quality

Keep prompts focused, avoid repeatedly pasting entire repositories, and assign the right model to the task. Use smaller models for summaries and quick searches; use stronger Claude or GPT models for complex reasoning and final code review. Because 59API is compatible with both workflows at https://api.59api.com, one account can support this split-agent setup with native official-quality models and simple usage-based billing. If you want to try the workflow, sign up for 59API and begin with one small feature in two isolated worktrees.

Tarifs · Docs