# ADR 043: Codex

- HTML version: https://robbiepalmer.me/projects/personal-site/adrs/043-codex
- Project: Personal Site (https://robbiepalmer.me/projects/personal-site.md)
- Status: Accepted
- Date: 2026-06-21

# Context

[ADR 012: Claude Code](/projects/personal-site/adrs/012-claude-code) records why this project uses an agentic coding assistant: deep reasoning, repository-wide changes, and autonomous iteration are central to maintaining velocity as a solo developer. This ADR accepts that basis without repeating it.

Claude Code remains effective, but I regularly exhaust the usage included with the $20/month Claude Pro plan. When that happens, implementation work stops until the allowance resets or moves to usage-based billing. Anthropic's direct subscription solution is the **[Max 5x plan](https://support.claude.com/en/articles/11145838-using-claude-code-with-your-pro-or-max-plan)** at $100/month.

Usage-based billing has not been a cost-effective fallback. In my experience, $20 of additional Claude usage can be consumed within a couple of working cycles. Reports that fully using a $200 subscription would correspond to thousands of dollars of API inference are difficult to verify precisely, but they are directionally consistent with my own experience: subscription allowances are heavily subsidized compared with marginal API usage.

I also experienced severe Claude availability problems earlier in 2026. I have repeatedly perceived both availability and output quality declining around major model releases. The timing does not establish that releases cause the degradation, but the operational effect is the same: relying on Claude alone can reduce velocity when demand is highest.

At the same time, experienced practitioners whose workflows I respect report that Codex and OpenAI's coding models can be at least as effective as Claude when used according to their different interaction model. That is not a claim that one model is universally better. It is sufficient evidence to evaluate Codex as a complementary tool rather than treating Claude Code as the only viable agent.

**[Claude Code on the web](https://code.claude.com/docs/en/claude-code-on-the-web)** supports mobile development in isolated Anthropic-managed environments. This provides a useful security boundary and does not require a personal computer to remain available, but fresh environments can require dependencies, tools, and project context to be prepared again.

**[Codex Remote Connections](https://developers.openai.com/codex/remote-connections)** allow the ChatGPT mobile app to control Codex on a connected host. The host supplies its existing projects, files, credentials, plugins, tools, and shell. The existing always-on Mac mini can provide that persistent development environment, avoiding repeated environment bootstrapping and its associated command output and context consumption. However, this workflow requires the Mac mini to remain available.

This is not a unique Codex capability. **[Claude Remote Control](https://code.claude.com/docs/en/remote-control)** also runs Claude Code on a personal machine while exposing it through web and mobile interfaces. Its server mode can start new sessions remotely, run multiple sessions concurrently, and place sessions in separate worktrees. Claude therefore provides two mobile development options: hosted cloud environments that do not depend on a personal host and Remote Control for working in an established local environment.

Both connected-host implementations are previews: Anthropic labels Remote Control a research preview, while OpenAI introduced Codex mobile remote access as a preview. Codex's documented mobile remote workflow requires a connected host, making its mobile development options a subset of Claude Code's. Codex is being adopted despite this limitation because its independent usage allowance, models, and provider improve capacity and resilience.

# Decision

I will use **[Codex](https://developers.openai.com/codex/)** alongside Claude Code as an agentic coding assistant, including using the Mac mini as a connected host for mobile development.

Codex is included with the **[$20/month ChatGPT Plus plan](https://developers.openai.com/codex/pricing)** across its app, CLI, IDE, and web surfaces. Current limits use a five-hour window for local messages and cloud tasks, with additional weekly limits possible. In my observed use, its allowance appears more generous than Claude Pro, and its reset cadence provides useful capacity when Claude Code is unavailable. This is an operational observation rather than a guaranteed entitlement; both providers can change their limits.

This is a complementary adoption, not a replacement for Claude Code. Using both tools preserves access to their different models, interfaces, and workflows while reducing dependence on one provider's availability, pricing, and usage policy.

# Alternatives

## Upgrade Claude Pro to Max 5x

* **Pros**: Increases Claude usage fivefold without introducing another coding tool or model family.
* **Cons**: Raises the Claude subscription from $20 to $100 per month while retaining a single provider dependency.
* **Decision**: Rejected. Keeping Claude Pro and adding ChatGPT Plus costs $40 per month in total, which is $60 less than Claude Max 5x, while also introducing provider and tooling diversity.

## Continue with Claude Pro Only

* **Pros**: No additional subscription or tool configuration.
* **Cons**: Regularly reaching the usage limit interrupts delivery and leaves no equivalent agentic fallback.
* **Decision**: Rejected. The existing limit is already constraining development velocity.

## Claude Usage-Based Billing

* **Pros**: Extends Claude Code usage without waiting for a reset or changing the base subscription.
* **Cons**: Variable API expenditure is harder to budget, remains concentrated on Anthropic, and has already consumed $20 within a couple of working cycles.
* **Decision**: Rejected as the default response to limits. It remains available for exceptional bursts of work.

## Claude Remote Control

* **Pros**: Runs Claude Code on a personal machine while being controlled from its web or mobile interface. Server mode can initiate multiple sessions and isolate them in worktrees, providing substantially the same core workflow and setup-efficiency benefit as Codex Remote Connections. Claude Code also offers hosted mobile sessions that do not require the personal machine to remain available.
* **Cons**: It still consumes Claude usage and depends on Claude availability, so it does not address the primary cost and provider-resilience drivers.
* **Decision**: Rejected as the sole solution. Codex remains justified by independent usage and provider redundancy rather than exclusive or broader remote-control features.

# Consequences

## Positive

* **Maintained Velocity**: Work can continue in Codex when the Claude allowance is exhausted.
* **Lower Cost than Claude Max**: Claude Pro plus ChatGPT Plus costs $40 per month rather than $100 per month for Claude Max 5x.
* **Provider Redundancy**: An outage, limit, or pricing change from one provider does not remove the agentic coding workflow entirely.
* **Different Strengths**: OpenAI and Anthropic models reason and interact differently. Learning both creates opportunities to route work to the tool that handles it best.
* **Complementary Mobile Capacity**: Mobile work can move between Claude's subscription allowance and Codex's independent allowance.
* **Persistent Mobile Environment**: Using either connected-host implementation, the Mac mini already has the repository, dependencies, tools, and credentials, reducing repeated setup work in fresh cloud environments.
* **Existing Infrastructure**: An always-on home server gains another useful workload without buying a separate development host.

## Negative

* **Additional Subscription**: The decision adds $20 per month rather than remaining on Claude Pro alone.
* **Narrower Mobile Feature Set**: Codex's mobile remote workflow requires the Mac mini, while Claude Code supports both hosted cloud environments and connected-host sessions. This adoption is justified by capacity and resilience, not feature superiority.
* **Workflow Adaptation**: Codex needs to be used differently from Claude Code. Applying the same prompting and supervision patterns mechanically may produce worse results.
* **Configuration Duplication**: Both tools require installation, authentication, and occasional tool-specific configuration, even though `AGENTS.md` provides shared project guidance.
* **Two Sets of Limits**: Redundancy reduces interruption risk but does not remove usage limits. Codex limits vary with task size, model, and execution surface, and additional weekly limits may apply.
* **Larger Local Blast Radius**: Remote Codex sessions use the host's filesystem, credentials, plugins, browser state, and tools. The Mac mini also handles personal photos and an external backup drive, so an overly broad permission profile could expose unrelated personal data.
* **Host Availability**: Mobile Codex work depends on the Mac mini remaining awake, online, signed in, and running the Codex app.
* **Security Hardening**: The connected host needs least-privilege filesystem rules, sandboxing, action approvals, and ideally a dedicated development account or explicit denial of the photo library and backup drive. The secure relay avoids a public inbound service but does not remove local access risk.
* **Verification Remains Required**: Practitioner recommendations and model reputation do not guarantee better results on this repository. Codex output requires the same testing and review discipline as Claude output.

---

Markdown index of this site: https://robbiepalmer.me/llms.txt