Your Terminal Agent Just Got a Background Reviewer (and It Won't Block Your Flow)
You know that feeling when you kick off a long coding task in your terminal agent, walk away for a coffee, and come back to find it's been stuck waiting on a review step that should've been background work? If you've used AI coding agents for any serious session, you've probably hit this wall. Every Code — a community-driven fork of openai/codex — is built to solve exactly that pain, with a focus on real developer ergonomics rather than just raw model capability.
What It Does
Every Code is a fast, local coding agent that runs entirely in your terminal. It's not a wrapper or a plugin—it's a full agent that can plan, write code, run commands, and even browse the web. The project positions itself as compatible with upstream openai/codex while adding a layer of features that make long automation sessions actually usable.
At its core, Every Code is about orchestration. It runs multi-agent workflows through commands like /plan, /code, and /solve, coordinates multiple CLI agents to tackle different parts of a task, and manages a unified settings hub for theme, approvals, and provider configuration. It also supports MCP (Model Context Protocol), so you can plug in filesystem access, databases, APIs, or your own custom tools.
The browser integration is notable too—it uses CDP (Chrome DevTools Protocol) for headless browsing, captures screenshots inline, and pairs with something called Code Bridge, a local bridge that streams errors, console output, and control from running applications back into the agent. Code Bridge ships with an MCP server, and you can install it by asking Code to pull it from a GitHub URL.
The latest cycle focuses heavily on what the maintainers call "long-session stability." The headline change is that Auto Drive and Auto Review are now decoupled. Previously, background reviews could block the main command flow. Now, pressing Esc returns control immediately, and you can keep typing while review finalization continues in the background. The project also introduced bounded queues and caches across the coordinator, TUI, and core state maps, so memory doesn't balloon during extended sessions.
Why It's Cool
The most interesting feature here is Auto Review—a background ghost-commit watcher that runs code reviews in a separate worktree whenever a turn changes code. It uses codex-5.1-mini-high to analyze changes and reports issues plus ready-to-apply fixes without ever blocking your main thread. That's a genuinely clever design decision: instead of forcing you to stop and review after every step, it treats review as a parallel concern that runs alongside your main automation.
This pairs well with Auto Drive, the multi-agent orchestration system. The README describes it as "self-healing" and capable of shipping complete tasks. When you combine Auto Drive with Auto Review, you get a workflow where the agent keeps making progress on the main task while quality checks happen in the background. The project's stated shift in focus is telling: from "can the model write this file" to "did we verify it works." That's the right priority for anyone who's spent hours debugging code that an agent confidently generated.
The model support is worth noting too. Auto Drive defaults to gpt-5.5 for planning and gpt-5.4-mini for fast coding loops, with reasoning controls at medium, high, and xhigh levels. It also handles newer model aliases like code-gpt-5.5, code-gpt-5.4, claude-opus-4.8, and claude-sonnet-4.6, with compatibility fallbacks for older names. That flexibility matters if you're switching between providers or want to experiment with different models per task type.
The theme system is a nice touch for terminal aesthetics—accessible presets, customizable accents, and live preview via /themes. It's not a headline feature, but it signals that the project cares about the developer experience beyond just raw functionality. Same with the safety modes: read-only, approvals, and workspace sandboxing are all available, which is essential if you're letting an agent run autonomously.
How to Try It
Getting started is straightforward. Clone the repository and check the README for the full setup details:
git clone https://github.com/just-every/code.git
cd code
From there, you'll want to review the release notes in docs/release-notes/RELEASE_NOTES.md for the full feature breakdown and any model configuration details. The key commands to explore are /plan, /code, /solve, and /settings to configure your provider and reasoning levels. If you want to try the browser integration, you'll need to set up Code Bridge by asking the agent to pull https://github.com/just-every/code-bridge.
For a quick taste of the Auto Review workflow, start a session, make a code change, and watch how the background review reports issues without interrupting your typing. The README mentions stress tests covering heavy agent churn plus concurrent Auto Review with Esc/typing responsiveness, so the stability work is backed by actual test coverage.
Final Thoughts
Every Code is best for developers who already use terminal-based AI agents and have hit the limits of single-threaded automation. The decoupling of review from the main command flow is a meaningful quality-of-life improvement—it's the kind of fix that seems obvious once you see it, but most agents don't bother with. The bounded queues and memory caps suggest the maintainers are thinking about long sessions as a first-class concern, not an afterthought.
Is it for everyone? If you're happy with a simple chat-based coding assistant, this might feel like overkill. But if you're running multi-step automation, coordinating multiple agents, or just want your terminal agent to verify its own work without babysitting, it's worth a look. The project is moving fast, and the focus on verified results over raw generation is the right direction for the space.
Follow @githubprojects for more developer tools and open source projects.