Skip to main content
Version: 1.0.0

Version 1.0.0

First IndustrialClaw release: a single-device platform that installs from one .run installer and is operated entirely from a browser on the device.

Download v1.0.0 · Installation

Validated Platforms​

  • AIR-075 · NVIDIA Jetson AGX Thor · JetPack 7.1 · Ubuntu 24.04.

Platform​

  • Installs Docker, pulls the four core images, generates the device's Gateway token and self-signed TLS certificate, installs the OverlayFS driver, and places a desktop shortcut.
  • The desktop shortcut runs an environment check, starts the containers, waits for the Dashboard, and opens the browser. Closing the browser stops the platform.
  • Dashboard on https://<device-ip>:18765/. The OpenClaw Gateway binds 127.0.0.1 only.
  • Model providers: local Ollama (default qwen3.5:4b-q4_K_M), vLLM, any OpenAI-compatible API, and GitHub Copilot. Each Agent can use a different installed model.

Agents​

  • An Agent is one ZIP of TOOLS.md plus skills/, up to 20 MB, imported through Import Package → Security Scanner → Create Agent.
  • SkillSpector scans every package. HIGH and CRITICAL findings block the import; MEDIUM requires acknowledgement of that scan digest. The scan runs again on submit rather than trusting the preview.
  • Updates are delta packages, scanned as the merged result. A successful update keeps existing Tasks, Materials, Outputs, and workspace files; a failed one leaves the installed version untouched.
  • A Skill may declare its own container runtime (compose, dockerfile, or image) in SKILL.md frontmatter, subject to Administrator approval at import.
  • Up to 10 Sub-agents per device.

Tasks​

  • Tasks are Markdown work orders in schema 1.0, created from the Dashboard after the Agent exists—up to 64 KB, with an optional Material ZIP up to 1 GB.
  • The platform runs each Step's command and check, applies retry and failure handling, runs Final Verification in a fresh session, and then confirms every Expected Output exists as a regular file.
  • One Task per Agent at a time, and at most five Agents running Tasks on a device.
  • Each Task Run exposes a Log, a Session, and downloadable Outputs.

Accounts and audit​

  • Two roles, Administrator and User, fixed at account creation. Users see only the Main Agent and their assigned Agents.
  • The bootstrap admin is the only account that can delete another account or force a sign-out, and cannot itself be deleted or demoted.
  • 50 accounts per device, 10 Administrators beyond the bootstrap admin, 3 concurrent sign-ins, 24-hour sessions.
  • Audit records 42 event types across sign-in, accounts, permissions, Agents, Tasks, settings, and OverlayFS, filterable by time, user, role, event type, and result.

Settings​

  • UI (User Interface): navigation position, light or dark theme, and English or Traditional Chinese, stored per account.
  • Token and Task Run history retention: 1 day, 7 days, 30 days, or forever.
  • OverlayFS write protection, enabled or disabled by an Administrator. Both directions require the device password and a reboot, and the switch is refused while any Task is running.

Known limitations​

  • One device per installation. There is no central multi-device console in this release.
  • Sub-agents do not call one another. Roles are chained by running Tasks in sequence and passing files between them.
  • A Task definition cannot be edited in place; delete it and create it again under the same filename.
  • While OverlayFS is on, importing or updating Agents, uploading Materials, and creating or deleting Tasks are blocked.
  • Uninstalling deletes Agent workspaces, Task definitions, Materials, and Outputs.