Mozilla AI, backed by the non-profit Mozilla Foundation, has launched a new open-source project called Project CQ. This ambitious project aims to create a shared, open knowledge base dedicated entirely to AI software coding agents. Peter Wilson, an engineer at the company, described the system as a Stack Overflow exclusively for AI agents. This innovative system allows coding agents to search through previous solutions, contribute new knowledge, and most importantly, avoid repeating the same mistakes in isolation from other systems.
The resource consumption and inefficiency challenge
The project addresses a well-documented problem in the tech community regarding inefficiencies in the current working environment. Currently, AI programming tools from various platforms—such as Claude Code, Cursor, and Copilot—consume massive amounts of tokens and cause frequent build and continuous integration failures because they independently and repeatedly rediscover and solve the same programming problems.
To illustrate the scale of this issue, developers on Hacker News pointed to a glaring example of a Cursor user who consumed nearly 170 million tokens in just two days, reflecting a massive waste of resources, time, and financial cost associated with running these advanced models.
How the new collaborative system works
To curb this waste, the system relies on an intelligent, proactive mechanism. Before an AI agent begins processing an unfamiliar coding task, it first queries the platform’s shared knowledge base to search for any pre-existing solutions. When the coding agent discovers a new solution or an innovative approach, it presents that knowledge to the system to be added as a proposal.
Next, other coding agents confirm the effectiveness of these solutions or report information that has become outdated or unusable. New knowledge units start with a low confidence level and gradually rise as they are verified by other agents or human developers.
Knowledge and development levels
The system organizes this knowledge across three core levels to ensure efficiency and security:
- Local level: Data and solutions specific to the individual user and their development environment.
- Enterprise level: Allows teams to share knowledge within the company boundary.
- Global commons: A level that may evolve in the future into a public, open platform for all coding agents globally.
The Python-based prototype is currently available on GitHub and includes custom plugins for multiple technical tools. The project also features a Model Context Protocol server for managing local knowledge stores, a dedicated team API, and a human-in-the-loop user interface to review and audit contributions. Wilson explained that the company began building this system in early March, influenced by a post from prominent researcher Andrew Ng questioning the feasibility of creating a knowledge platform for coding agents.
Security risks and protection solutions
Reactions from the developer community quickly divided. On Hacker News, contributors acknowledged the real problem of poor efficiency while simultaneously pointing out severe security risks. If these tools blindly search for solutions, attackers could exploit this to introduce poisoned knowledge units containing backdoors.
These concerns gain additional weight given recent security incidents. In February 2026, researchers uncovered a vulnerability allowing remote code execution via manipulated configuration files, and 1,184 malicious skills were discovered in an AI framework marketplace.
To address these concerns, Mozilla’s architectural document points to the integration of advanced protection mechanisms against knowledge poisoning, including anomaly detection techniques, requiring diverse confirmation sources, and enforcing direct human-in-the-loop verification steps.
The collapse paradox and building the alternative
In his remarks, Wilson used a tone of deliberate paradox, noting that the volume of questions asked on traditional platforms has collapsed dramatically, dropping from over 200,000 questions per month at its peak in 2014 to just 3,862 questions in December 2025 due to the rise of AI models.
“It is a case akin to filial cannibalism, where the offspring swallowed the human society that fueled its data, and now this offspring seeks to build its own version of the resource that contributed to its destruction.”
Finally, it is worth noting that the system is currently classified as not ready for production, with Mozilla AI testing it internally to discover and correct any issues or vulnerabilities before moving toward broader adoption.
FAQ
What is the new system launched by the Mozilla Foundation?
The foundation launched an open-source initiative aimed at creating a collaborative knowledge library dedicated to AI coding agents, working to reduce resource consumption and prevent the repetition of errors.
How does this platform contribute to reducing coding waste?
The platform allows AI tools to search for pre-stored programming solutions before starting to write code, preventing them from consuming millions of tokens to solve problems already solved by other tools.
What are the most prominent security risks associated with this project?
Concerns focus on the possibility of malicious actors injecting the platform with harmful code or backdoors, which AI tools might retrieve and apply blindly, exposing systems to breaches and data leaks.
How did the company handle potential security concerns?
The company adopted defense mechanisms that include anomaly detection techniques, requiring multiple confirmation sources for new solutions, and establishing a human auditing layer to ensure no malware is passed through.
