Run Codex and Claude Code Side by Side
Why run Codex and Claude Code together?
Codex and Claude Code are most useful when they have different responsibilities instead of competing for the same task. Codex is well suited to fast implementation, repository-wide edits, and test-driven changes. Claude Code is often excellent at reading unfamiliar code, tracing behavior across files, reviewing architecture, and explaining risk. Running both side by side gives you a practical implementation-and-review loop.
The key is to isolate their credentials, terminals, model choices, and Git changes. Without that separation, one agent can overwrite another agent's work or accidentally inherit the wrong API configuration.
1. Create isolated shell profiles
Use one terminal for Claude Code and another for Codex. Do not export both providers in your global shell profile. Instead, create two small launcher scripts or shell functions so each process receives only the variables it needs.
- Claude Code terminal: set ANTHROPIC_API_KEY and ANTHROPIC_BASE_URL=https://api.59api.com, then start Claude Code.
- Codex terminal: set OPENAI_API_KEY and OPENAI_BASE_URL=https://api.59api.com, then start Codex.
For example, a Claude launcher can export the 59API key and ANTHROPIC_BASE_URL before calling claude. A Codex launcher can export the same relay key under OPENAI_API_KEY before calling codex. Keep the key in your operating system's secret store or an ignored local environment file; never commit it to the repository.
Codex builds that use provider configuration can be set up in ~/.codex/config.toml with a 59API provider, its base_url, and an env_key of OPENAI_API_KEY. If your installed client expects the OpenAI-compatible /v1 path, follow the endpoint format documented for your 59API account rather than guessing. Verify the active endpoint with a harmless model-list or short request before starting a large task.
2. Give each agent a separate worktree
Separate terminals are not enough if both agents share one checkout. Use Git worktrees so each tool has an independent working directory:
git worktree add ../project-codex -b codex/implementation
git worktree add ../project-claude -b claude/review
Run Codex from project-codex and Claude Code from project-claude. This prevents half-written files, formatter races, and conflicting package-lock changes. Ask Codex to commit small, coherent units. Then fetch or cherry-pick the commit into the review worktree, rather than asking Claude Code to inspect uncommitted changes that may still be changing.
3. Use a deliberate handoff protocol
Start with a written task contract containing the goal, constraints, test command, files that may change, and acceptance criteria. Have Codex implement the smallest viable patch. Require it to run focused tests and summarize assumptions in the commit message.
Next, give Claude Code the commit, test output, and an explicit review role. Ask it to inspect for edge cases, security problems, backward compatibility issues, missing tests, and unnecessary complexity. A useful instruction is: “Review this commit without rewriting it. Report findings by severity, cite file and line locations, and propose tests for every high-confidence issue.”
After fixing review findings, run the full test suite from a clean worktree. This catches problems hidden by shared build artifacts or locally modified configuration.
4. Control cost and model selection
Use a stronger model for architecture, difficult debugging, and final review; use a faster, lower-cost model for file discovery, routine edits, and test iteration. 59API provides pay-as-you-go access to native Claude Opus, Sonnet, Haiku, Fable, and GPT models, so you can choose by task instead of paying a fixed subscription for every agent.
Set a budget habit: begin with a narrow prompt, limit the initial file scope, and ask for a plan before a large edit. Inspect usage after each phase. Because 59API is designed as a low-cost relay while retaining official-quality models, it is a practical option for parallel experimentation, and its referral rebate can further reduce ongoing spend.
5. Troubleshoot the common failures
- 401 errors: confirm the variable name matches the client and that the launcher, not your editor, supplied the key.
- Wrong model or endpoint: print the relevant environment variables with the key value redacted and verify the provider configuration.
- Unexpected file conflicts: check the current worktree with git status and confirm each terminal is in its intended directory.
- Expensive repeated context: keep prompts focused, reuse concise handoff notes, and avoid asking both agents to rediscover the entire repository.
If you want one relay for both workflows, sign up for 59API and start with a small Codex implementation plus a Claude Code review. That setup keeps the tools independent while making multi-model development affordable and repeatable.