How Mozilla’s CQ Aims to Build a Stack Overflow for AI Agents

Mozilla, the company behind the Firefox browser, is developing an open‑source project called CQ, led by engineer Peter Wilson. The initiative is described as a “Stack Overflow for agents,” aiming to enable AI agents to discover and share collective knowledge.

Wilson points out that agents often encounter the same problems repeatedly, leading to unnecessary work and token consumption. By using CQ, agents first query a shared knowledge base and can contribute new solutions, reducing duplication.

Current developer workflows rely on static context files such as agents.md, skill.md, or claude.md for Anthropic’s Claude Code. Wilson advocates a dynamic approach that builds trust over time rather than depending on static instructions.

The CQ codebase is written in Python and remains in an exploratory, testing phase. It can be installed locally and includes plugins for Claude Code and OpenCode. The project ships a Docker container that runs an API, an SQLite database, and an MCP (Model Context Protocol) server.

According to the architecture documentation, CQ stores knowledge in three tiers: a local tier, an organization tier, and a “global shared tier,” the latter referring to publicly available CQ instances. Knowledge units start with low confidence and cannot be shared initially; confidence increases as other agents or humans confirm them.

Mozilla is discussing whether to host a public, development‑focused CQ instance. Wilson notes internal debates about distributed versus centralized sharing and their implications for the community. He suggests that Mozilla.ai could help launch an initial central platform, but stresses the need for pragmatic validation of user value and awareness of hosting trade‑offs and risks.

The workflow includes potential security concerns such as malicious content and prompt‑injection attacks, where attackers could direct agents to perform harmful tasks. Mitigation mechanisms include anomaly detection, diversity requirements (confirmation from multiple sources), and human‑in‑the‑loop (HITL) verification.

Agents assign confidence scores to knowledge entries, which other agents then use, raising concerns about hallucinations. HITL supervision can help mitigate these issues.

Wilson uses the term “matriphagy” to describe how large language models have consumed Stack Overflow content, arguing that agents now need to build their own Stack Overflow.

Mozilla’s broader AI strategy includes Mozilla.ai, part of the Mozilla Foundation, with projects like Octonous for managing AI agents and any‑llm for providing a unified interface to multiple LLM providers. Mozilla also continues to operate MDN, a popular documentation site for web technologies, which currently does not use AI.

For more information, see the blog post at https://blog.mozilla.ai/cq-stack-overflow-for-agents/ and the GitHub repository at https://github.com/mozilla-ai/cq.