The single-user ceiling
Upstream OpenClaw is brilliant for one person. One device, one user, one session store, one sandbox, one set of channel credentials. That assumption is wired through the codebase — `~/.openclaw/sessions/`, `~/.openclaw/memory/`, one set of cron jobs. The moment a colleague wants their own thread, you're either spinning up a second daemon or vendoring patches across the codebase.
OpenClawMU does the boring, careful work of slicing those state-bearing surfaces by tenant: per-tenant session and memory directories, per-tenant sandbox roots, per-tenant cron, per-tenant channel pairings, per-tenant skills config (with admin-only key protection so a tenant can't override platform secrets). One gateway, many users, isolated workloads.
The cloud-platform ceiling
The cloud bot platforms — Botpress, Chatfuel, ManyChat, Voiceflow, Cognigy — are great until you hit one of three walls:
- Per-active-user pricing. Your unit economics get worse as you grow.
- Bring-your-own-key constraints. You pay them a margin on top of your already-paid Anthropic/OpenAI bill.
- Data residency / compliance. Their database is theirs. Your audit team will not love it.
OpenClawMU sits in your infra, talks to your LLM, stores state on your disks, and never sends anything to a Neul Labs server because no such server exists.
The "roll your own" ceiling
You could just take upstream and add multi-tenancy yourself. We did the math. Tenant directory layout, hashed token issuance and constant-time verification, dispatcher token routing across 59 gateway methods, sandbox spawning per tenant (including bubblewrap detection, seccomp profile generation, mount namespace hardening), per-tenant quota accounting and cost snapshots, S3-compatible backup with path-traversal protection, a control-plane HTTP API for admin resource management, and a tenant-scoped OpenAI-compatible endpoint — that's the 6–8 weeks OpenClawMU saves you.
Every change is fork-marked with OPENCLAWMU ADDITION. When
upstream ships, you rebase and the markers move; you never lose the diff,
you never lose the boundary.
The bet, stated plainly
Self-hosted, multi-tenant, open-source AI infrastructure is the inevitable layer between the model providers and the bot/agent products every team will build in the next 18 months. OpenClawMU exists to make that layer free, forkable, and audit-friendly — so the products on top of it can be the differentiated work.
If you're shipping a bot product, a messaging-fronted agent service, or an internal AI gateway for your company — start here, save the eight weeks, ship the differentiated thing.