Your AI agents now live in a Matrix room you control

What it does

HiClaw is a multi-agent runtime built around a Manager-Workers architecture. It deploys containers for the Manager and individual Workers, then wires them into a shared Matrix room alongside human users. A built-in Higress AI gateway holds the actual API keys and GitHub PATs, handing Workers only consumer-scoped tokens so they can collaborate without ever seeing the real secrets.

The interesting bit

The platform treats the Matrix room as its central nervous system: agents exchange messages, share files through an integrated MinIO filesystem, and pull skills on demand from a community registry. By letting deterministic OpenClaw or QwenPaw agents lead tasks while autonomous Hermes Workers handle code execution, it matches the runtime to the job instead of forcing one engine to do everything.

Key highlights

• Credentials never reach Workers; the gateway keeps real API keys and PATs locked away.
• Built-in Tuwunel Matrix server and Element Web client mean no external IM approvals or bot registrations.
• Kubernetes-native control plane with declarative YAML for Worker, Team, and Human resources.
• MinIO shared filesystem cuts token bloat by letting agents swap files instead of pasting contents into prompts.
• Multi-region Helm charts are available, though the default image registry points to Alibaba Cloud China unless overridden.

Caveats

• Minimum footprint is 2 CPU cores and 4 GB RAM; realistic multi-agent setups want 8 GB.
• It orchestrates agent containers rather than implementing agent logic itself—you bring your own LLM and runtimes.
• Default image registry is China-based; deployments outside Asia should switch to the North America or Southeast Asia mirror.

Verdict

Worth a look if you need an on-premise, auditable control plane for agent teams. Skip it if you’re after a single-agent coding assistant or a fully managed SaaS.