cq: A Shared Knowledge Commons for AI Agents

cq: A Shared Knowledge Commons for AI Agents

Written by Nikos Vaggalis

Thursday, 16 July 2026

Mozilla.ai has introduced cq, an open-source standard and platform designed to facilitate shared learning among AI agents. This promises to eliminate a lot of duplicated effort on the part of AI agents, thereby making savings in terms of cost and resources.

cq (which stands for the radio call sign CQ meaning "any station, respond") is an open standard and shared knowledge commons for AI agents. In general, think of it as Stack Overflow for AI agents; just as human developers use Stack Overflow to share programming solutions, AI agents (such as Claude Code, Cursor, or GitHub Copilot) use cq to broadcast what they have learned and listen to what other agents already know. This is where the "any station, respond" fits.

But why is cq necessary? Currently, AI agents operate in isolation; when an agent hits a roadblock like an undocumented API quirk, a library version conflict, or a configuration bug, it has to figure out the solution from scratch. It writes failing code, triggers broken builds, reads files, and slowly debugs. Imagine now this being encountered by thousands of agents.

When thousands of agents encounter the exact same error, they collectively repeat these identical failures daily, burning massive amounts of tokens, time, compute power, and electricity. Worse, because agents traditionally lack persistence, they might even forget what they learned in your own project once the session resets. cq solves this "agent amnesia" by allowing agents to securely persist, query, and verify collective experience. By utilizing a common knowledge format, the system allows agents to exchange experience-driven insights, preventing them from repeating costly technical mistakes in isolation.

An agent integrated with cq does not just write code, it follows a skill-guided query/propose workflow:

Query (before acting). When the agent is about to start an unfamiliar task, like integrating the Stripe API or configuring a Rust CI/CD pipeline, the cq skill directs it to search the knowledge store first.

Act (with guidance). If another agent has already solved this problem (e.g., "Stripe returns HTTP 200 instead of a 429 error when rate-limited"), your agent retrieves this "Knowledge Unit" immediately and writes the correct error handling on its first attempt, saving time and tokens.

Propose (when discovering something new). If your agent encounters a novel error and successfully debugs a workaround, it drafts and proposes a new Knowledge Unit (KU) to save future agents from repeating the struggle.

Session reflection. Real-time logging misses things when agents are deep in a debugging rabbit hole. At the end of a coding session, running /cq:reflect prompts the agent to retrospectively scan the entire conversation history, identify debugging breakthroughs or undocumented workarounds, and present them to the human developer as structured KUs for easy, one-click contribution.

Under the hood, cq works by providing a standard set of Model Context Protocol (MCP) tools for the agent to call:

query-Searches the store before executing actions.

propose-Submits a new learning.

confirm-Endorses an existing KU that proved correct, something that boosts its confidence rating.

flag-Marks an outdated, incorrect, or duplicate KU to lower its confidence.

status-Shows store statistic

Knowledge in cq moves through a tiered "graduation" process to keep local systems lightweight and ensure only secure, verified, generalizable insights reach the public. The infrastructure operates across three tiers, including private local storage, organizational namespaces for teams, and a Global Commons for public knowledge:

Tier 1-Local agent/machine level. Private to your computer and saved in a local SQLite database. This solves basic agent memory loss across sessions and never shares data unless explicitly nominated.

Tier 2-Team/ organization. Shared privately within a company or department. It tracks internal microservices, proprietary APIs, and development environments. It runs as a hosted Postgres database using vector search to find relevant entries.

Tier 3-Global Commons (public). A completely free, community governed public knowledge pool managed by Mozilla.ai at cq.exchange, immediately available to any AI agent in the world.

Before wiring cq into any specific AI assistant, you must first install the core cq CLI and make sure it is added to your system's PATH. You can do this via Homebrew (for macOS/Linux) or Scoop (for Windows).

Once the CLI is installed, you use the cq install tool to register the cq Model Context Protocol (MCP) server, install the shared cq skill, and add the necessary always loaded instruction block to your assistant's configuration.

Then to confirm that your assistant can talk to cq, open a terminal session with your AI agent and run the status slash...