Onboard and lead your agents
Lead your agents like a great team.
Role, onboarding, feedback — and nothing about the way they work changes without your approval.
Free · only Docker required · your data stays with you
The problem isn’t the first run. It’s the fiftieth.
The prompt lives in a chat history, the process in someone’s head, the knowledge in three versions of a file — and when something goes wrong, nobody knows what changed.
- PersonaInvoice Checkerv3
- PlaybookPre-check an invoicev5
- KnowledgeRequired invoice detailsv2
Every change with version, approval and history
Hire your first agent.
Describe a task and play through how an agent is created, approved and improved in Who2Be.
- 1Task
- 2Draft
- 3Approval
- 4Feedback
- 5v2
This is what an agent looks like in Who2Be. With JavaScript, you can create one yourself here.
Or pick an example:
A preview in your browser — nothing is sent or stored.
Builder: Draft created: persona, playbook and tool permissions. I can’t activate it — that’s up to you.
Approved. Your agent now works with version 1.
A few weeks later, it reports on its playbook:
outdated
Draft v2 runs against your test cases
- Typical case
- Edge case
- Case from feedback
Version 2 is active. Version 1 stays readable and can be restored.
- Task
- Pre-check invoices
- Persona
- thorough, friendly, asks when something is unclear
- Playbook
- Pre-check an invoice
- 1Check required detailsv2− +
- 2Reconcile amountsv2− +
- 3Flag anything unusualv2− +
- Knowledge
- Required invoice details
- Tools
- Read allowed · write off
- History
- v1 · Activev2 · Active
Onboard your agents like a new colleague.
Job description corresponds to
Persona
Work instructions corresponds to
Playbooks
Company handbook corresponds to
Knowledge
Keys and access corresponds to
Tool permissions
Experience corresponds to
Memory
Performance review corresponds to
Feedback and cases
Probation period corresponds to
Test cases
Four-eyes principle corresponds to
Approval
Fast memory, slow behavior.
What your agent remembers changes quickly. How it works only changes with your approval.
Inner circle: memory
Remembers what helps on every run — instantly or after your approval.
Outer circle: behavior
Persona, playbooks, knowledge and prompt only change through these six steps.
- FeedbackAgent or human reports: playbook outdated.
- CaseThe situation is recorded, and the agent gives its side.
- DraftThe improvement becomes version 4. Version 3 stays active.
- Test casesVersion 4 is tested on real cases: 3 of 3 passed.
- ApprovalVersion 4 waits for you. Without your consent, nothing happens.
- ActiveVersion 4 is at work. Version 3 stays readable and retrievable.
Works with your tools.
Who2Be doesn’t run agents. Your agents fetch their role, processes and knowledge via MCP — 90 tools, locally or via OAuth.
- Personas
- Playbooks
- Knowledge
- Claude Code
- Claude.ai
- Your own agent
What Who2Be manages
Personas
Role, tone and boundaries — shared by several agents.
Playbooks
Processes with triggers, combinable into larger ones.
Knowledge
Block editor with references and backlinks.
Agents
Persona, playbooks, tool permissions and memory in one.
Work area
Notes, files and tables for each agent.
Search
Full-text and semantic, even in long documents.
All 90 MCP tools and the architecture in the repository (external site)
Today on your machine. Soon in the cloud too.
Self-host
Free, with every feature, on your own system.
- Docker is all you need
- Your data stays with you
- Every feature, no limits from us
git clone https://github.com/luetzey/who2be.git && cd who2bedocker compose up -d --waitCloud
We run Who2Be for you — sign in, get started, nothing to install.
- Sign in with Google, GitHub or Apple
- Free to start, Pro for everyday use
- Servers in Germany or Finland
Same source code · free for internal use · Apache 2.0 two years after each release
For everyone who has to sign off.
- Data locationWith you — or in Germany or Finland.
- Source code and licenseOpen to inspect. FSL-1.1, Apache 2.0 after two years.
- Roles and accessAdmin, Editor, Viewer. A second factor for sensitive actions.
- TraceableEvery change versioned, agent access logged.
Frequently asked questions
When is the cloud coming?
It’s expected in late 2026. Until then, Who2Be runs entirely on your own system — free and with every feature. You can follow new versions in the releases on GitHub (external site).
Do I need Who2Be if I already use a tool for building agents?
Probably alongside it, not instead of it. Who2Be doesn’t run agents — it keeps their configuration and delivers it at runtime via MCP. Your execution tool stays where it is.
Which models are supported?
Who2Be doesn’t talk to any model itself. It delivers personas, processes and knowledge to your agent; which model that agent uses is up to you, in your tool.
How do I connect it to Claude Code?
Via MCP: either as a local connection or as a remote connection with OAuth. In the interface, you create a token under Settings and copy the ready-made client configuration right next to it.
Can I move from the cloud to my own server later?
It’s the same software. There’s no “export everything” button for moving content today — the content lives in a Postgres database and can be backed up and restored with the usual database tools. An export button in the interface is on the list.
Is there a German version?
Yes. The interface and this website are both available in English and German — switch to German.
I found a bug or have a request.
Pick a task you do yourself too often.
Free on your own system. The cloud follows, expected in late 2026 — follow releases on GitHub (external site).
New — and verifiable in ten minutes.
No customer logos, no made-up numbers. Instead: