luetzey/who2be

Self-hosted hub for the personas, playbooks and knowledge your agents run on —
versioned, reviewed, and served to them over MCP.
Quickstart ·
Connect an agent ·
Documentation ·
Changelog ·
Discussions
Why Who2Be
Instead of scattering system prompts, workflows, and knowledge documents
across chat histories, notes, or repositories, Who2Be keeps them in one
self-hosted configuration store (the "AgentDB") and serves them to your
agents at runtime through an MCP server.
- **Stop re-typing the same instructions.** Personas, playbooks and reference
knowledge live in one place, and every agent loads them from there at
runtime.
- **See what each agent runs on.** Every agent is one persona, one system
prompt template and an explicit tool policy, and every change leaves a
version and a status trail.
- **Standards over probability.** A new version starts as a draft and has to
pass review before it goes live; there is no shortcut from draft to active.
Agents can propose changes, but by default they cannot publish them.
Features
- **Personas** — an agent's identity, tone, boundaries, and modes, versioned
with a status workflow and diff view
- **Playbooks** — step-by-step workflows with trigger keywords, composable
into composite bundles
- **Resources** — knowledge documents with a block editor (BlockNote),
block refs, and reverse lookups
- **Agents** — concrete agent configurations with an expanded system prompt,
tool policy, and curated long-term memory
- **System prompt templates & external tools** — reusable prompt building
blocks and MCP tool bindings with tool-ref placeholders
- **Review workflow** — every version moves draft → review → active (and
later inactive); draft → active is not a permitted transition. Putting a
version live takes the admin role, each change is recorded in a status
history, and any two versions can be diffed or an earlier one restored
- **MCP server** — 100 tools: read, write, full-text + semantic search,
discovery, and the feedback flywheel (record_usage/submit_feedback);
connect via stdio or the OAuth 2.1 remote connector, e.g. to Claude Code
or Claude.ai. Writing personas, playbooks, resources or agents is off per
agent until you enable it
- **Agent work area & knowledge base** — an unversioned workspace per agent
(notes, file/URL ingest, read-only SQL tables, timeline) next to the
curated resource axis, plus an evidence-backed knowledge base with typed
edges; promotion into resources is an explicit step
- **Multi-tenancy & RBAC** — organizations → workspaces, roles
admin > editor > viewer, magic-link invitations, MFA step-up
- **Two editions from one codebase** — on-prem (signed license key) and
cloud (billing package), isolated at build time
The design decisions behind these features are recorded as ADRs in
docs/adr/.
Quickstart
**Docker is the only prerequisite** — no Python, no Node, no .env file.
Open , create an account (sign-ups are auto-confirmed
locally, so no mail server is involved), and you land in a personal workspace
that is created on first login.
Useful follow-ups:
The first start builds the API and web images from source, which takes a few
minutes. To pull prebuilt images instead:
Connect an MCP client
The MCP server runs in the stack as well, on http://localhost:8765/mcp
(Streamable HTTP, ADR-0034) and behind the web origin at /mcp. It
authenticates with an ordinary Who2Be token, so no OAuth setup is needed:
1. In the web UI go to **Settings → Tokens**, create a token (w2b_…) and copy