Tools
RocCLAW
By AI Agent Dashboard
Install
npm install -gREADME
<div align="center">
<img src="public/logo.png" alt="rocCLAW" width="280" />
# rocCLAW
**Run AI agents on your hardware. Use cloud only when you need it.**
The operator dashboard for [OpenClaw](https://github.com/openclaw) โ manage a hybrid fleet of local and cloud agents from any browser. Your GPUs stay busy, your cloud tokens go only where they matter.
[](https://nodejs.org)
[](https://github.com/kiritigowda/rocCLAW/releases)
[](LICENSE)
<img src="public/screenshots/app-main.png" alt="rocCLAW dashboard" width="900" />
</div>
---
## Table of Contents
- [Why rocCLAW?](#why-rocclaw)
- [Quick Start](#quick-start)
- [Local + Cloud Hybrid Fleet](#local--cloud-hybrid-fleet)
- [What You Can Do](#what-you-can-do)
- [Monitor Your Hardware](#monitor-your-hardware)
- [Built-in Skills](#built-in-skills)
- [Use Cases](#use-cases)
- [Dashboard at a Glance](#dashboard-at-a-glance)
- [Installation](#installation)
- [Setup Guides](#setup-guides)
- [Requirements & Compatibility](#requirements--compatibility)
- [Development](#development)
- [Troubleshooting](#troubleshooting)
- [Documentation](#documentation)
---
<a id="why-rocclaw"></a>
## ๐ค Why rocCLAW?
<div align="center">
<img src="public/screenshots/bot-family.png" alt="A hybrid fleet of AI agents โ local and cloud, managed from one dashboard" width="680" />
</div>
Most AI tools wait for you to type a prompt, return a response, and stop. An agent is different โ it takes an objective, breaks it into steps, executes across tools and systems, and keeps running on a schedule without manual intervention. OpenClaw agents can monitor log files, run CI pipelines, triage issues, sync data between services, and operate continuously on schedules you define.
The problem: running agents around the clock on cloud models gets expensive. If an agent is checking system health every five minutes or processing a queue of routine tasks, those tokens add up โ especially when open-weight models running on your own hardware can handle the same work at zero marginal cost.
**rocCLAW lets you build a hybrid agent fleet.** Local models handle the daily workload at zero token cost. Cloud models step in only for the tasks that need them โ complex reasoning, multi-step planning, deep context. You control the split per-agent, and the dashboard shows you exactly where every token goes.
Point rocCLAW at any OpenClaw gateway โ on your desk, across the network, or SSH-tunneled from a remote server โ and your entire fleet is right there. Chat, configure, schedule, monitor. No SSH, no terminal juggling, no guessing what your agents are doing.
```
Browser โโโ HTTP / SSE โโโ rocCLAW Server โโโ WebSocket โโโ OpenClaw Gateway
(React) (Next.js + SQLite) (local GPU / cloud API)
```
Your browser never talks to the gateway directly. rocCLAW proxies everything securely โ authentication, event replay, rate limiting โ and your tokens never leave the server.
---
<a id="quick-start"></a>
## ๐ Quick Start
**Prerequisites:** Node.js 20.9+ and a running OpenClaw gateway.
Install via npm, pre-built package, or from source โ see [Installation](#installation) for all options.
```bash
npm install -g @kiritigowda/rocclaw
rocclaw
```
Open [http://localhost:3000](http://localhost:3000), enter your gateway URL (`ws://127.0.0.1:18789`), paste your token, and click **Connect**.
```bash
openclaw config get gateway.auth.token # Get your token
```
See also: [full install guide](docs/INSTALL.md) ยท [setup guides โ](#setup-guides)
---
<a id="local--cloud-hybrid-fleet"></a>
## ๐๏ธ Local + Cloud Hybrid Fleet
**Local agents** run on your hardware with open-weight models via [Ollama](https://ollama.com), vLLM, or any local provider. They handle the predictable workload โ log monitoring, scheduled reports, file processing, data syncing, health checks. Zero token cost, and they retain memory across sessions so they improve without burning cloud credits.
**Cloud agents** use high-capability models (Claude, GPT, Gemini) for tasks that need it โ complex reasoning, multi-step planning, code generation with deep context.
**Per-agent model selection** โ Assign each agent exactly the model it needs. Your cron agent runs locally on Kimi K2. Your planning agent calls Claude. Pair it with the right [built-in skills](#built-in-skills) โ Plan First and Agent Debate for cloud agents, ReAct Loop and GitHub for local. Mix and match.
**Token usage dashboards** โ See spend per agent, per model, in real time. Know exactly which agents are consuming cloud tokens and whether they should be.
**The result:** maximum hardware utilization, minimum cloud spend. Your local GPUs stay utilized instead of idle. Cloud tokens go only to tasks that need them.
---
<a id="what-you-can-do"></a>
## โก What You Can Do
**Chat with any agent** โ Real-time streaming with thinking traces, tool call visibility, and inline exec approvals. Approve or deny shell commands right in the chat โ allow-once, allow-always, or deny.
**Put agents on autopilot** โ Schedule recurring jobs with drag-and-drop โ run every 5 minutes, daily at 9am, or any cron expression. Agents retain context across sessions and act on heartbeat schedules independently.
**Configure without SSH** โ Edit any agent's personality files and permissions directly in the browser. Each agent has 7 personality files that define its behavior:
<details>
<summary>IDENTITY ยท SOUL ยท USER ยท AGENTS ยท TOOLS ยท HEARTBEAT ยท MEMORY</summary>
`IDENTITY.md` โ name, creature type, vibe, emoji, avatar ยท `SOUL.md` โ core truths, boundaries, personality ยท `USER.md` โ context about you (name, pronouns, timezone) ยท `AGENTS.md` โ operating rules and workflows ยท `TOOLS.md` โ tool usage guidelines ยท `HEARTBEAT.md` โ periodic check configuration ยท `MEMORY.md` โ persistent memory and learned context
</details>
**Access from anywhere** โ Connect to any gateway via LAN, Tailscale, or SSH tunnel. Your gateway stays secure; you stay mobile.
**Stay in control** โ Per-agent exec permissions, sandbox isolation, and cryptographic device authentication. See [Permissions & Sandboxing](docs/permissions-sandboxing.md) for the full security model.
---
<a id="monitor-your-hardware"></a>
## ๐ Monitor Your Hardware
When your agents run on local hardware, you need to see how that hardware is doing. rocCLAW provides live system metrics so you know whether your GPUs are earning their keep or sitting idle.
**Live gauges** โ CPU, memory, GPU utilization, VRAM, disk, and network. Works for local machines **and** remote gateways โ "Remote" vs "Local" labels are applied automatically.
**Time-series graphs** โ Track resource usage over 5m, 10m, or 30m windows. Spot bottlenecks, see when your GPU is maxed out, and decide whether a task should move to cloud.
**AMD GPU support** โ ROCm-first detection with automatic sysfs fallback. Full metrics for AMD GPUs including VRAM, temperature, power draw, and clock speeds. See [Requirements & Compatibility](#requirements--compatibility) for details.
---
<a id="built-in-skills"></a>
## ๐ง Built-in Skills
<div align="center">
<table>
<tr>
<td align="center"><img src="public/screenshots/bot-before-skills.png" alt="Agent before skills โ basic tools" width="420" /><br/><em>Before: basic tools</em></td>
<td align="center"><strong>+ Skills</strong><br/>โ</td>
<td align="center"><img src="public/screenshots/bot-after-skills.png" alt="Agent after skills โ rocket scientist" width="420" /><br/><em>After: rocket scientist</em></td>
</tr>
</table>
</div>
Same agent, same hardware. The right skills change what it can do.
rocCLAW ships with 12 featured skills you can assign per-agent directly from the dashboard โ no config files, no CLI. Give your local agent the skills it needs for routine work, and equip your cloud agent for complex reasoning.
| Category | Skill | What it does |
|----------|-------|-------------|
| **Agent Behavior** | Proactive Agent | Anticipates needs, self-schedules crons, maintains a working buffer |
| | Self-Improving Agent | Self-reflection, self-criticism, self-learning โ evaluates and improves permanently |
| **Problem Solving** | Plan First | Generates a detailed plan before execution (Plan-and-Solve research) |
| | ReAct Loop | Interleaves reasoning with actions, observing results to inform next steps |
| **Quality & Accuracy** | Agent Debate | Multiple agents debate answers to reduce hallucinations |
| | Self-Critique | Structured self-review against quality criteria before finalizing |
| **Development** | Team Code | Coordinate multiple agents as a dev team working in parallel |
| | Skill Creator | Build new skills from scratch, validated against the AgentSkills spec |
| | GitHub | Issues, PRs, CI runs, code review via `gh` CLI |
| | Git Workflows | Rebasing, bisecting, worktrees, reflog recovery, merge conflicts |
| **Multi-Agent** | Agent Team Orchestration | Defined roles, task lifecycles, handoff protocols, review workflows |
| | Multi-Agent Collaboration | Intent recognition, intelligent routing, reflection across agent teams |
Skills are **per-agent** โ assign different combinations to each agent to match its role in your fleet. Need more? Browse and install additional skills from [ClawHub](https://clawhub.ai) โ integrated directly into rocCLAW with one-click install.
---
<a id="use-cases"></a>
## ๐ก Use Cases
<div align="center">
<img src="public/screenshots/bot-field-guide.png" alt="Agent field guide โ each bot cataloged by role and specialization" width="680" />
</div>
A hybrid fleet makes sense anywhere you have repetitive work alongside tasks that need deeper reasoning.
- **DevOps & infrastructure** โ A local agent monitors logs, resta
... (truncated)
tools
Comments
Sign in to leave a comment