Mozilla cq: Stack Overflow for AI agents

With cq, Mozilla AI has introduced an open-source project intended to serve as a shared knowledge base for AI coding agents. The name is derived from “colloquy,” specifically a structured exchange of ideas. The stated goal: agents should no longer work in isolation, repeatedly encountering the same errors, but should be able to learn from each other.

As Peter Wilson explains in a blog post at Mozilla, AI agents currently always work independently of each other. If an agent encounters an unknown problem – such as an API with unexpected behavior or a faulty CI/CD configuration – it must solve it independently: write code, trigger errors, diagnose, start over. If another agent encounters the same problem, the entire process is repeated. This costs tokens and computing power.

According to Mozilla, the situation is exacerbated by the fact that the models' training data becomes outdated. At the same time, platforms like Stack Overflow, which once served as a central knowledge source and whose content flowed into the models' training data, are affected by a massive decline in users. Specifically, Mozilla points to a drop from over 200,000 questions per month at its peak in 2014 to under 4,000 in December 2025.

How cq works

cq relies on decentralized knowledge exchange: Before an agent tackles an unknown task, it queries the so-called “cq commons.” If another agent has already solved the problem, the solution is immediately available. If an agent learns something new, it can feed this knowledge back into the database. Other agents confirm it through practical use or mark it as outdated. Knowledge is thus intended to build trust through application, not mere authority.

In this context, Mozilla points to a trust gap: 84 percent of developers use AI tools or plan to do so, but 46 percent do not trust the accuracy of the results – an increase from 31 percent the previous year. Knowledge confirmed by multiple agents across different codebases could carry more weight here than the individual answer of a single model, Mozilla hopes.

A first functional prototype of cq includes plugins for the coding agents Claude Code and OpenCode. It also includes an MCP server (Model Context Protocol) for the local knowledge store, a team API for sharing within organizations, a user interface for human review, and containers for setting up the entire system. According to Mozilla, development only began in early March, so it is officially an early proof of concept.

Technical details about cq can be found on the project page on GitHub.

Open Standard Instead of Vendor Lock-in

Mozilla explicitly positions cq as a vendor-independent project. Not everyone uses the same coding agents, and developers should not be prescribed a specific tool, it states in the blog post. The current approach of storing knowledge in Markdown files within repositories has its limits. Instead, a dynamic system is needed that builds trust over time.

Peter Wilson explicitly points out that the idea aligns with a recently published post by AI researcher Andrew Ng. He also suggested a “Stack Overflow for AI coding agents.” Accordingly, Mozilla sees this as a confirmation of its own approach and calls on the developer community to participate in the design of cq.

(fo)