Getting Started¶

Adding a Cognition Protocol to an existing multi-agent system

A Cognition Protocol is designed to be added to a multi-agent system that already exists, without modifying the application. The design assumption is minimal: agents communicate using some messaging mechanism, whether native to the framework or an external CLI. No further assumption is made about the application.

The messaging mechanism can be anything that lets one agent send a message to another and read its reply. Pick the setup that matches your system.

Claude Agent Teams¶

The quickest way to try a protocol: an initiator and its teammates in a single Claude Code session, with no message bus. The initiator spawns teammates directly and exchanges messages with them using SendMessage. Works on a single machine with no network between agents.

Tutorial: running Cognition Protocols with Claude Agent Teams

AgentBridge¶

Independent agent processes joined by AgentBridge (A2A, here over a SLIM message bus), each sandboxed. Agents can live in separate processes, containers or harnesses — and use different runtimes — and communicate over the network.

Tutorial: running Cognition Protocols over AgentBridge

MAS Lab¶

Agents defined as YAML manifests (kind: Agent, kind: MAS) and run with mas-ctl. No coding harness required: add the protocol skills to the agents' manifests and run the MAS.

Tutorial: running Cognition Protocols with MAS Lab

Other coding harnesses and Claw-like agents¶

Each protocol is distributed as skills — an initiator skill, a receiver skill, and the scripts that implement the deterministic portion of the protocol. Install the initiator on any agent that may trigger the protocol and the receiver on every agent that may be asked to take part. If any agent in the system can trigger the protocol, every agent should carry both skills: each one can then start a run, and each one can answer when another starts it.

Getting the skills¶

Claude Code marketplaceAI CatalogGitHub releases

For Claude Code, the recommended install path is the plugin marketplace. One command registers this repository as a marketplace, and one more installs each protocol.

Install both skills (both roles):

claude plugin marketplace add ioc-project/cognition-protocols
claude plugin install concord@cognition-protocols
claude plugin install accord@cognition-protocols

Each protocol plugin contains two skills — concord-initiator and concord-receiver (or accord-initiator and accord-receiver). Their activation conditions are precise: the initiator skill only activates when the agent is orchestrating a protocol run; the receiver skill only when the agent receives a protocol message. An agent carrying both will activate the right skill automatically based on which messages it receives.

Install for a single role:

The plugin marketplace does not support per-skill enable/disable (Claude Code's skillOverrides setting applies only to .claude/skills/ directory skills, not to plugin skills). If an agent plays only one role and you want to reduce context cost, install only the skill you need via the skills directory — see Installing a skill below.

Skills are published to an AI Catalog and can be installed with the ai-catalog CLI.

Register the catalog once:

ai-catalog catalog add ioc-protocols https://ioc-project.github.io/cognition-protocols/ai-catalog.json

Then pull and install a skill. Skills are distributed as zip archives, so unzip after pulling:

ai-catalog pull urn:air:ioc-project:skill:concord-initiator -o /tmp/skill-download/
unzip /tmp/skill-download/concord-initiator.zip -d /tmp/skill-download/
mv /tmp/skill-download/concord-initiator ~/.claude/skills/

Repeat for each skill you need (concord-receiver, accord-initiator, accord-receiver).

Download individual skill zips from the GitHub releases page, then unzip and copy the skill folder into the harness's skills directory.

Installing a skill¶

Copy (or symlink) each skill folder into the harness's skills directory.

Harness	User-wide	Project only
Claude Code	`~/.claude/skills/`	`.claude/skills/`
Codex CLI	`~/.codex/skills/`	`.agents/skills/`
Other agentskills-compatible	see harness docs	same

Symlinks stay updated on git pull. Agents running in separate processes or containers each need their own copy of their skill, including its scripts/ folder.

Agents whose code you control¶

Frameworks such as LangChain, Google ADK, and others support agentskills-compatible skill loading. If your agent framework supports skills, use the same skills above.

If skills are not available, implement the protocols directly against their specifications.