Company brain
What a company brain for AI agents is, what belongs in it, how agents and assistants use it over MCP, and how to start one as an open-source Hub.
A company brain is the maintained, shared context that lets AI agents and people answer the way your business works: its definitions, rules, procedures, decisions and owners, plus governed access to the systems that hold live records. In Hubzoid a company brain is a Hub, a folder of files you can read, review and keep in Git.
- Definitions and rules
- Procedures
- Tools to live systems
- One Hub
- Chat, workflows, MCP
What belongs in it
The test for each piece of material: would an experienced colleague need it to do the work correctly?
| Material | In a Hub | Example file |
|---|---|---|
| Definitions, policies, thresholds | Reference knowledge | knowledge/net-sales.md |
| How a recurring job is done | A skill | skills/weekly-briefing/SKILL.md |
| Who the agent serves and its rules | The definition | AGENTS.md |
| Live records such as orders and stock | Reached through a tool, not copied in | tools_local/orders.py or an MCP connector |
| Large exports and background documents | Raw material, searched as text | raw_data/2026-q3-export.csv |
| Passwords and API keys | Never. Protected configuration only | .env or a secret store |
Name an owner and a date in the documents that define the rules. Retire the old version rather than leaving two policies for the agent to choose between.
How agents use it
The same Hub serves three ways of working, each with its own history and access:
- Chat. A teammate asks in the web app or Slack. The agent answers from the Hub and names the file it used.
- Workflows. A scheduled task or code workflow prepares the weekly report from the same definitions.
- Your own assistant over MCP. Claude Code or another MCP client connects to the Hub, signed in as you, and gets the Hub's instructions, knowledge and the tools you may use.
A rule changed in one file reaches all three. "Provide context once" does not mean everyone sees everything: restricted tools run only for the people granted them, and separate teams can keep separate Hubs.
A company brain is not chat memory
| What it is | Where it lives | |
|---|---|---|
| Shared knowledge | Facts and procedures the whole Hub relies on | Hub files, reviewed like code |
| Conversation history | One conversation's back and forth | Per person, per chat |
| Workflow state | Which periods or items were already handled | The workflow's own records |
| Personal notes | What one person asked the agent to remember | That person only |
An agent can save a note to shared knowledge only when the deployment allows it and a person with the right grant asks. Saved notes are proposals to review, not automatic truth. See memory and history.
Why keep it as files
- Readable. The person who knows the process can open the rule, check it and fix it.
- Reviewable. Changes are diffs. You can see who changed a definition and when.
- Portable. Plain Markdown and Python move between models, runtimes and hosts.
- Open. Hubzoid is Apache 2.0, so the brain and the software that serves it stay yours.
Search over a pile of documents can find material, but it cannot tell a current policy from last year's. Ownership and dates make the brain trustworthy. Searchable raw material is still useful for detail, as long as the agent can tell rules from records.
Start one
- Run
hubzoid init my-huband openmy-hub/knowledge/. The quickstart walks through it. - Pick one question your team asks every week. Write its definition, source, owner and exceptions in one file.
- Add a skill if the answer needs several steps, then a tool if it needs live data.
- Ask the question in chat and check the agent names your file. Change the rule and ask again.
- Connect Claude Code over MCP so your own assistant uses the same brain.
Next steps
- The Hub: every part of a Hub and how the agent reads it.
- Connect an assistant: use the brain from Claude Code.
- Memory and history: keep each kind of context in the right place.