Qodo plugs organizational context into OpenAI Codex for AI code review

Qodo has built a plugin for OpenAI's Codex marketplace that injects organizational context, codebase rules, and independent review into developers' coding sessions, aiming to catch P0 bugs that span 10,000 or more repositories. The startup, which makes AI tools for code review and software quality, positions itself as the knowledge layer that coding agents need before they touch code they don't fully understand.

Co-founder and CEO Itamar Friedman calls Qodo "the coding agent's best friend." It's a short pitch, but it names a real gap in the market. Coding agents can write functions efficiently. What they can't do natively is know that a seemingly correct local change will break a downstream service hosted by a different Git provider or connected through Kafka.

That's the stakes here. As agents write more of the world's software, the failure mode shifts from bad syntax to bad context — and enterprises with sprawling repository estates have the most to lose.

What does the Codex plugin actually do?

The workflow runs in two phases, both inside the same session. Before Codex starts writing code, the Qodo plugin retrieves the relevant codebase rules, plus pull request history and patterns. After Codex makes its changes, Qodo reviews them and returns findings directly to Codex — before the developer ever opens a pull request.

"Codex is really good at taking a local task and completing it in an efficient way," Itamar says. Qodo's contribution is everything around that local task: codebase knowledge, organizational rules, and independent review delivered through the company's Agentic Toolbox.

The cross-repository piece matters most. Itamar describes customer environments that can span 10,000 or more repositories, where a local code change can break a downstream service even if it looks correct on its own. "Qodo will make that connection and will help Codex look at the software as a whole and not miss those P0 bugs," he says.

Under the hood, Qodo runs on OpenAI models, including GPT‑6 Astra, to gather engineering knowledge, review code, and monitor software quality. The company isn't shy about its model preferences. "We love OpenAI models. We want the frontier, the best reasoning models, but also those that are really fast," Itamar says.

How developers are using it today

Nnenna Ndukwe, Qodo's AI Developer Relations Lead, says developers are using the combined workflow to build features and tackle technical debt with GPT‑6 models inside Codex. The pattern she sees is automation of the handoff: Codex does the coding, Qodo does the independent review, and the loop closes without human mediation.

"That's what I've been seeing developers do, which is really exciting because GPT‑6 is extremely capable," Nnenna says.

Nnenna runs the tooling on herself, too. She has been using GPT‑6 Astra's Extra High and Ultra effort settings in her own automated development workflows as well as knowledge work. Her setup includes:

• Scheduling Codex to report on what she accomplished during the week and where she could focus or improve next
• Using it to research conversations about agentic engineering on social platforms and suggest ideas for developer content
• Scheduling Codex to analyze her own code changes and suggest agent skills to create or refine

Her verdict on the platform is blunt: "What does Codex not do? I love everything that Codex does. I love the idea of combining AI and automation."

Why 'yin and yang' agents could change who reviews code

Itamar frames coding and review agents as "yin and yang": one creates code, the other challenges it. Each role needs its own setup, its own context, and may involve multiple agents working in concert.

His vision goes a step further than automated review. He wants agents to discuss their findings with each other and "surface only what they don't accept" for human review. Under that model, developers stop writing every line and become planners and reviewers instead — inspecting unresolved issues alongside the agents' discussion, the underlying context, and prior decisions made by other developers.

"That's how you create trust," Itamar says.

"I think it makes the whole process overall more efficient," Nnenna adds. "We get to leverage the best parts of our expertise."

The framing matters for the broader market. If agents can genuinely negotiate disagreements about code quality before escalation, the human bottleneck in code review — long the slowest stage of the pipeline for large teams — starts to shrink.

What role does OpenAI's Decisions API play?

Itamar also sees an opening in OpenAI's recently launched Decisions API. Rather than relying on language models for every decision, Qodo could use the API to choose which rules and agent skills apply at each stage of development.

He gives a concrete example: a customer with 200 rules doesn't need a model reasoning over all of them every time. The Decisions API could first select which subset needs checking, then let a language model reason through just the relevant ones. That makes selection faster and cheaper while improving review quality.

"The Decisions API is an immense opportunity for us to bring more trust to the process," Itamar says.

The architectural point is worth noting. Moving rule selection off the LLM and onto a deterministic API layer is a small instance of a bigger industry shift — using large models only where reasoning is genuinely required, and cheaper machinery everywhere else.

Is Qodo building a 'software factory'?

Nnenna describes Qodo's approach as combining "the best aspects of artificial intelligence and the best aspects of traditional software engineering and scripting." The division of labor is explicit:

• Company knowledge guides the coding agents
• Independent review checks their work
• Human expertise resolves the questions that remain

Itamar wraps the whole system in a bigger label: a software factory. By his description, it's a system customers could assemble and own — coding agents and review agents working together, with an understanding of how software actually changes in production.

That ownership point distinguishes Qodo's pitch from fully managed platforms. The customer keeps the rules, the context, and the assembly. With OpenAI as the model provider, Qodo is betting that coding agents will put more of a company's accumulated engineering experience to work, while developers focus their judgment where it matters most.

The near-term test is adoption inside the Codex marketplace. If the cross-repository review claims hold up in 10,000-repository environments, Qodo's plugin becomes a natural checkpoint for any enterprise running Codex at scale — and a template for how context vendors attach themselves to the agent ecosystem.