Internal AI agents
What internal AI agents are, examples of the work they do, how to build one from your own knowledge and rules, and when to build, buy or use open source.
An internal AI agent is an AI agent that works for your own team rather than for your customers. It answers questions and does recurring work from your knowledge, your rules and the tools your team already uses, and the people who rely on it can see and correct what it knows.
- Your knowledge and rules
- Agent
- Tools you allow
- A person reviews
How it differs from a chatbot
A general chat assistant knows a lot about the world and little about your business. An internal agent is the reverse: narrow, but grounded in the files and systems that define how your team works.
| General AI assistant | Internal AI agent | |
|---|---|---|
| Who it serves | One person at a time | A team, with access set for each person |
| What it knows | Public knowledge and what you paste in | Your maintained knowledge, rules and examples |
| What it can do | Answer and draft | Answer, run tools you grant, and do recurring work |
| How it changes | You re-explain each time | Someone edits a file and every answer follows |
| Who is responsible | The person chatting | A named owner, with run history and logs |
Customer-facing agents, such as a support bot on your website, are a different job with different risks. This page is about agents your own people use.
What internal agents do
Good first agents take a job your team already does by hand, often every week, where the rules are known and a person checks the result. These are the jobs Hubzoid's ready-to-use agents do:
| Agent | The work | Where a person stays in charge |
|---|---|---|
| Software Factory | Turns GitHub issues and engineering requests into draft pull requests, checked by an independent reviewer | A person reviews the draft and merges |
| Engineering Desk | Reviews pull requests against your team's rules and answers questions about your code | Nothing posts to GitHub until you allow it |
| Order & Invoice Desk | Reads orders and supplier invoices from Gmail into sourced records and reports | Unclear items come back as questions |
| Customer Desk | Drafts replies to customer requests in Zoho Desk or Freshdesk from approved knowledge | A person approves every reply |
Other common jobs: a weekly report built from the same sources each time, a check that flags exceptions before month end, and a team assistant that answers policy questions and names the document it used.
What an internal agent needs
Most of the work is not the model. It is everything around it.
- Knowledge and rules in one place. Definitions, procedures, thresholds and examples of a good result, kept where people can read and correct them. In Hubzoid this is the Hub, a folder you can keep in Git.
- Tools with limits. Read-only lookups for everyone, and sensitive actions behind a permission granted to named people. Checks belong in code, not in the prompt. See identity and access.
- A person at the right moment. The agent asks before a consequential action, or prepares a draft that a person sends. See restrict a tool and the approvals guide (opens in a new tab).
- The places people already work. Web chat, Slack, scheduled reports and the assistant someone already uses, such as Claude Code over MCP.
- Operation after launch. Run history, failures you can explain, usage and cost, and an owner who keeps the knowledge current. See observability.
How to build one
- Pick one job. Choose a task someone does regularly, with known rules and a person who checks the result. Write down who reads the output and what a good result looks like. Design an agent has a one-page template.
- Collect the material. The policy document, a few past examples with known outcomes, and the systems the work touches.
- Turn it into files. Instructions in
AGENTS.md, rules and examples inknowledge/, repeatable procedures inskills/, and tools for the systems. A coding assistant such as Claude Code or Codex can write these with you. - Test it on known cases. Ask the questions you already know the answers to. Write the important ones down as evals so a later change cannot quietly break them.
- Correct a rule and check again. Edit the file, not the chat. The next answer should follow the new rule.
- Give it to the team. Grant access to the people who need it, connect the surfaces they use and, for recurring work, put it on a schedule.
The quickstart does steps 3 to 5 with a small sample in about five minutes.
Build, buy or use open source
| Route | Good when | What you take on |
|---|---|---|
| A packaged AI assistant from a vendor | The job is general writing, search or summarising, and your data may sit with that vendor | Limited control over tools, rules and where it runs |
| A no-code agent builder | The work is a visual workflow across SaaS apps | The builder's hosting and pricing, and logic spread across canvases |
| An agent framework (for example Mastra, Agno or LangGraph) | Your engineers are building an AI application with its own interface | Access control, chat surfaces, scheduling and operation are yours to build |
| An open-source internal agent platform such as Hubzoid | The people who know the process should own the knowledge and rules, and the agent should run where you choose | Hosting and upkeep, and the honest limits of a young project |
These overlap, and teams often combine them. A framework can call a Hub over MCP, and a Hub can use tools that wrap other systems.
Where Hubzoid fits
Hubzoid is an open-source platform for internal AI agents. Give it your team's knowledge, rules and tools once, then use them in chat, workflows and the AI assistants you already use.
- The Hub is a readable folder, so the person who knows the process can change a rule and see the result.
- One Hub serves web chat, Slack and other channels, scheduled workflows, and assistants over MCP.
- Sensitive tools are checked against each person's grants before they run, and approvals can be required per tool.
- It is Apache 2.0, including team controls, and runs on your laptop or a server you control.
Next steps
- Quickstart: build a small Hub and ask your first question.
- Share your Claude Code agent with your team: start from a setup you already have.
- Ready-to-use agents: clone a complete agent and adapt it.