Meet Claude Agents

AI teammates you can give real work to

They run on your machine, work from your own database, and nothing you own leaves your absolute control.

Requires a Claude Code subscription. Runs on macOS and Windows.

The Claude desktop app with a sidebar of named agent sessions, and the phone app showing a bookkeeping entry held in draft for approval.

Give each agent a job

One agent, one role, one conversation

  • Bookkeeper
  • Document Librarian
  • Data Analyst
  • Code Reviewer
  • Database Admin
  • Report Writer
  • Staff Engineer
  • Chief of Staff
  • Operations Manager
  • Finance Assistant
  • Customer Support
  • HR Assistant
  • Legal Clerk
  • Creative Assistant
  • Product Manager
  • Sales Clerk

Close the books without the chasing.

Files each invoice against the right supplier and period, drafts the journal entries, checks they balance, and leaves them in draft for you to post. Nothing gets done without your approval.

See more ways to use Claude Agents

How it works

Your machine, your data, your rules

Your data never leaves

Everything runs on your local device under your own account. A PostgreSQL database is installed for you. Nothing leaves unless you connect something that can send it, and approve it each time.

Nothing new to download

Under the hood every agent is a Claude Code session. You reach them in the Claude apps you already have: the desktop app, claude.ai, or the phone app. Each agent appears in your session list by name, and you message it the way you message Claude today.

No new app in your dock. No new interface to learn. No second place to check.

Work the way you want to

Your agents work from an ontology you write — the terms you use for the data in your business and how they relate to each other. Add a relationship in your own language and every agent uses it the same way, every time.

Files stay where they are

Your files stay in your file system. The database records their location, metadata and what they are about. Ask the data manager and it will find a file by name or the meaning of its contents and its relationship to other data.

Every answer carries its source

Each fact an agent works out records what produced it and when. Walk back from a decision to the report, to the analysis, to the original data source, with fingerprints proving none of it changed.

Agents that stay in their lane

No agent can do anything outside the scope you provide — the folders it can access, whether it can write to them or only read, whether it can search the web. Everything is under your direction via the admin bot.

Agents that keep working

Each agent starts at sign-in and keeps its own conversation. Stop one and it pauses. Start it again and it picks up where it left off. They only sleep when you turn your machine off, then wake up again when you restart it.

How Claude Agents is different

Six decisions that set it apart

01

Data sovereignty

Most AI products keep your context on somebody else's servers. Claude Agents installs PostgreSQL on your own machine and everything is recorded in that one place. It is a database file on a disk you control, in a format any tool can read if you grant it permission.

You can open it directly via a human user interface so you never lose access to your data. You can back it up. You can take it with you. You don't need permission from any third party to get to your own data.

02

Your ontology

Most systems make you fit your business into their categories. Here you write your own.

The database holds the kinds of record your business deals in and the relationships allowed between them, as rows you can edit. An invoice may be billed to a person or an organisation. A document may be about a project. A decision may depend on a report. You add what your business actually has, in the words your business actually uses, and an agent that tries to connect two things in a way you have not permitted is refused.

No more expensive, complex CRMs, ERPs, or other applications that need acronyms to describe them.

03

Each agent sees only its work

Every database record carries a label. A table says which agent may read and write which labels. The database applies it to every query, row-by-row, including ones the agent writes itself.

The rule lives in the database rather than in the software, so it holds even when an agent has been talked into asking for something it should not have. An agent can read the rules that bind it. It cannot edit them.

04

Agents do not message each other

Grok Bot lets its agents pass work between themselves. Claude Agents does not, on purpose.

A request travels as a file, and a person reads it on the way. That means a document carrying instructions aimed at a machine reaches nobody who can act on it, and every widening of what a machine can touch passes a human. You trade some autonomy for a boundary that holds.

05

Swap the engine, keep the record

Every session records which engine and which model did the work. Change the model and the record says which conclusions came from which, with dates.

The record does not belong to the engine. It belongs to you, and it outlives whichever model wrote into it.

06

Based on the knowledge framework

Data is the file. Information is that file read into shaped facts: this invoice, this amount, this supplier. Knowledge is how those facts connect and what they explain. Judgement is the decision you took, recorded with what it rested on.

Then the loop closes. The action taken sits beside the decision. The consequence sits beside the action, holding what you expected and what actually happened. A decision nothing followed from is visible as exactly that.

What it does not do

The boundaries are the product

It does not run in the cloud

One machine, your account, your disk. Everything runs on your local device under your own account, and a PostgreSQL database is installed for you.

It does not sign into your apps and click around on your behalf

It works through your database and your files, not by driving your screen. Nothing leaves unless you connect something that can send it, and approve it each time.

It does not let agents instruct each other

A request travels as a file, and a person reads it on the way. That means a document carrying instructions aimed at a machine reaches nobody who can act on it.

It does not let an agent widen its own reach

No agent can grant itself a folder, a command, or a wider view of your data. An agent can read the rules that bind it. It cannot edit them.

It does not decide anything

It records what you decided and what came of it. The action taken sits beside the decision, and the consequence sits beside the action, holding what you expected and what actually happened.

It does not add another app to your dock

Under the hood every agent is a Claude Code session. You reach them in the Claude apps you already have: the desktop app, claude.ai, or the phone app.

Meet your first agent

Install it, then open the desktop app, claude.ai or your phone and find the session named claude-admin. Ask it for the agent you want and it makes it.

Requires a Claude Code subscription. Runs on macOS and Windows. PostgreSQL is installed for you if it is not already there.