Compare lobu
How lobu stacks up against OpenClaw, Claude Tag, Viktor, and Town, and how event-sourced org memory and connectors work under the hood.
At a glance
OpenClaw (personal runtime), Claude Tag (platform coworker), Viktor and Town (managed hires), and lobu. Different jobs, not worse products.
Swipe to compare →
| Capability | OpenClaw | Claude Tag | Viktor / Town | lobu |
|---|---|---|---|---|
| What it is | Personal agent runtime | @Claude in Slack | Managed AI coworker | Multiplayer agent backend |
| Open source / self-host | Yes | No (cloud) | No (cloud) | Yes |
| Multi-tenant isolation | Single-tenant shared FS | Per-channel @Claude | Personal or workspace | Per channel / DM sandbox |
| Org memory | Local files / session | Channel-scoped | Learned user / workspace | Event-sourced · entities |
| End-to-end history | Session / files | Channel context | Product memory | Full event log agents can replay |
| Source ACL mirror | No | Platform-managed | Per-user OAuth | Fail-closed (Slack · GitHub live) |
| Credentials to workers | Often on the host | Platform-managed | Cloud OAuth | Never (gateway proxy) |
| Human approvals | DIY | Managed | Draft / confirm flows | Tools · writes · action modes |
| Connectors & reach | Skills + local setup | Admin-provisioned | Fixed OAuth catalog | SDK + agents can ship connectors |
| Proactive work | Heartbeats / cron | Async tasks | Routines · schedules | Scheduled + event-driven |
| Platforms | Many chat apps (local) | Slack (beta) | Slack · email · calendar · more | Slack · TG · WA · Discord · Teams · MCP |
| Embed in your product | No | No | No | Yes · API · MCP · multi-tenant |
| Model choice | Any you wire | Claude only | Vendor stack | 16 providers |
Event-sourced memory agents can trust.
Everything that happens (connector pulls, messages, agent actions, human approvals and corrections) appends to an event log. Facts project onto typed entities (companies, projects, people, incidents). Agents don’t re-derive the world from a chat window: they read the same event history and entity state end to end, then act under ACL and approval gates.
- Append-only log of what happened
- Source + run provenance on every fact
- Corrections supersede, never erase
- Full trail agents can follow end to end
- Shared current state for the org
- Typed objects + relationships
- Same graph across agents and channels
- ACL-filtered on every recall
› Event sourcing: connectors and agent actions write the log; entities are the shared picture. Slack, web, and Claude/ChatGPT via MCP all see the same history, filtered by ACL.
How entity memory works →Not limited to a fixed catalog.
Ship built-in connectors where they exist. When a source isn’t covered, use the Connector SDK: TypeScript definitions for auth, feeds, events, and operations. Agents run in a sandbox with code and tools, so they can draft and extend connectors the same way a human engineer would, then apply them so new sources feed the same event log and entities.
Auth methods, feed kinds (poll, webhook, virtual), event shapes, and write ops in the Connector SDK.
Connections append events. Agents read the log and entity graph, the same end-to-end history for everyone in the org.
If the catalog is missing a tool, an agent can write the connector in the sandbox, then you approve and ship it.