ClawFleet is an open-source multi-agent fleet management tool, officially positioned as an "AI company" solution running on personal hardware. The project centrally manages AI agents running in isolated Docker containers through a browser dashboard, and currently supports two runtimes, OpenClaw and Hermes. The official website states that deployment can be completed in about 10 minutes (the introduction also writes 5 minutes), and emphasizes that you can use your own ChatGPT subscription, thereby avoiding cloud bills; data remains on local hardware, with no SaaS subscription required.
Purpose and Positioning The official description compares ClawFleet to the idea of buying N dedicated Mac Minis, running one AI agent on each (some OpenClaw, some Hermes), and having them collaborate in Discord. ClawFleet makes this idea a reality for free on an existing Mac or Linux machine. Each agent runs in an independent Docker container with an isolated file system and network. The official resource usage reference: each OpenClaw instance uses about 500 MB of memory, and each Hermes instance about 150 MB (idle memory data measured on an M4 MacBook Air with 16 GB of memory; OpenClaw's memory rises by about 3 times when the agent actively browses and loads Chromium, while Hermes remains basically flat).
Main Capabilities
- Two runtimes, one dashboard: OpenClaw and Hermes are both managed as first-class citizens, sharing an asset pool (models, channels), with each instance having container isolation, real-time statistics, logs, event streams, and clawfleet shell access.
- Sandboxed instances: each agent is in an independent Docker container, isolated from the host machine and other instances.
- Any LLM provider: supports OpenAI, Anthropic, Google, DeepSeek, or use a ChatGPT subscription (OpenClaw).
- clawfleet shell: can enter the terminal of any instance for Hermes TUI chat or OpenClaw bash shell operations.
- Version locking: locks tested runtime versions to prevent upstream breaking changes from affecting use.
- Role system: reusable role settings (introduction, backstory, style, traits), currently for OpenClaw instances.
- Skill management: OpenClaw has 52 built-in skills, and 13,000+ community skills can be obtained through ClawHub.
- Soul Archive: snapshots roles + memory + configuration, and clones them as new members, currently exclusive to OpenClaw.
- Browser graphical desktop: view an instance's graphical desktop in the browser via noVNC, which can be used to observe agent behavior, manually guide workflows, or demonstrate; currently applicable to OpenClaw images with the XFCE desktop, and equivalent visibility for Hermes is on the roadmap.
Installation and Usage The official installation command performs the following steps: install Docker as needed (Colima on macOS, Docker Engine on Linux), download and install the clawfleet CLI, pull the prebuilt sandbox image (about 1.4 GB), start the dashboard as a background daemon, and open it in the browser http://localhost:8080。Mac Users are strongly advised to install Docker Desktop first for the best experience; otherwise, ClawFleet will automatically install Colima as a substitute Docker runtime.
By default, the dashboard listens on 0.0.0.0:8080 on Linux, can be restricted to localhost, and can also be accessed via an SSH tunnel established from a laptop. The official documentation states that OpenClaw's built-in Control Panel requires a secure context, which an SSH tunnel can satisfy; the Hermes native dashboard does not require a secure context; other dashboard features can be used without a tunnel. For manual installation methods, see the Getting Started wiki page.
The "Run Your Agent Company" workflow The official documentation uses a company metaphor to describe usage: Assets are the tools owned by the company, and Fleet is the team of AI employees. First put tools into the toolbox—Assets → Models to register LLM API keys (each model is verified before saving, and models are shared across all runtimes); Assets → Characters to create reusable roles (such as job descriptions for CTO, CPO, and CMO, including bio, background, communication style, and personality traits; currently oriented toward OpenClaw, while Hermes uses the personality system in its native dashboard); Assets → Channels to connect messaging platforms (verified before saving; OpenClaw supports 24+ channels, while Hermes currently supports Discord / Telegram / Slack). Then create instances in Fleet → Create (OpenClaw or Hermes), and assign models, roles, and channels in Fleet → Configure; install skills for instances in Fleet → Skills (52 built-in + one-click installation of 13,000+ from ClawHub; the skill manager is currently exclusive to OpenClaw, while Hermes manages skills through its native dashboard); save an instance's soul in Fleet → Save Soul (personality, memory, model configuration, and conversation history), browse saved souls in Fleet → Soul Archive, and load them via Load Soul when creating a new instance, so that new members inherit the original knowledge and personality. The official documentation also demonstrates connecting the fleet to messaging platforms, after which engineers, product managersA scenario of autonomous collaboration with marketing personnel in a Discord group chat.
Differences between the two runtimes OpenClaw: The upstream is github.com/openclaw/openclaw, a local-first assistant gateway that lets a single agent be accessed from 24+ messaging platforms (Telegram, Discord, Slack, Lark, WhatsApp, Signal, Matrix, etc.). The bot exists in conversations as a participant, supporting pairing codes, @ mentions, and roster-aware coordination with other bots. It is suitable for scenarios where you want the bot to exist as a participant in group chats, for collaboration among multiple different roles (such as a CEO + CTO + PM in one Discord), and for installing skills from a community directory. In ClawFleet, models + roles + channels can be configured through the dashboard, roles are hot-reloaded as reusable roles (SOUL.md), and the skill manager provides 52 built-in items + ClawHub installation, with support for Save Soul snapshots and cloning. Hermes: The upstream is github.com/NousResearch/hermes-agent, built around a learning loop: the agent writes new skills from experience, improves them through use, and maintains memory accumulated across sessions (FTS5 cross-session search + Honcho user modeling). The primary interface is the TUI, and messaging platforms are a secondary remote control method. It is suitable for scenarios where you want a personal assistant that learns about you over time, need first-class cron / scheduled automation delivered to messaging platforms, remote execution on SSH / Daytona / Modal (serverless) backends, use of long-tail LLM providers (Nous Portal, GLM, Kimi, MiMo, MiniMax, 200+ on OpenRouter), and prefer a TUI-first workflow. In ClawFleet, models + channels (Discord / Telegram / Slack) can be assigned through the dashboard, using the same asset pool as OpenClaw, and it provides one-click links to the Hermes native dashboard (credential pool, cron, personality, terminal backend) as well as clawfleet shell hermes-1 to enter the interactive TUI.
Documentation and commands The official Wiki provides complete documentation, including Getting Started (prerequisites, installation, first realexample), Dashboard Guide (sidebar, asset management, fleet management), LLM provider guides (Anthropic, OpenAI, Google, DeepSeek), channel guides (Telegram, Discord, Slack, Lark), CLI Reference, and FAQ. Every command supports --help. The quick reference lists commands to destroy all instances (including data), stop the dashboard, and remove all build artifacts.
License and participation The project uses the MIT license, and contributions through issues or PRs are welcome.
Limitations and information not yet provided The official documentation does not clearly state ClawFleet's own charging model or pricing; it only states that cloud bills can be avoided and no SaaS subscription is required. Some capabilities (role system, skill manager, Soul Archive, noVNC graphical desktop) are currently exclusive to OpenClaw or only apply to the OpenClaw image, and Hermes's feature scope in the dashboard is newer and narrower (container lifecycle and Configure), with deeper features requiring access through the Hermes native dashboard. Achieving full feature alignment between the two runtimes on the dashboard is the direction of progress stated officially. Other features, pricing, platform compatibility details, and effects not explained in the materials are not yet provided.
