ClawPanel: a desktop control room for OpenClaw, Hermes Agent and DeepSeek Harness

Three agent engines behind one Tauri window

ClawPanel is a management panel, not an agent framework. It targets three third-party engines: OpenClaw, which the README links through a Chinese translation repository, Hermes Agent from nousresearch, and DeepSeek Harness from deepseek-ai. The app ships as a cross-platform desktop application built on Tauri v2, with JavaScript as the primary language and Rust underneath per the project topics, and the README is translated into eleven languages including English, Japanese, Korean, Spanish, Russian, French and German. The pitch is a built-in AI assistant that handles one-click installation, configuration diagnosis and error repair through what the README describes as 8 tools, 4 modes and interactive question answering. The audience is explicit: people who want to run these frameworks without living in a terminal, and the project clearly aims at a Chinese-speaking base first, with QQ, WeChat, Douyin, Feishu, Discord and Telegram communities all linked from the README.

The DeepSeek Harness integration shows the real engineering

The most concrete mechanism in the README is the DeepSeek Harness section. A managed runtime installs, starts, stops and uninstalls a pinned acceptance version of the harness, deliberately avoiding pollution of the global npm environment. An independent workbench opens the full DSH web interface inside the panel, with direct conversation, workspace selection, model switching and session management. The model channel is unified: provider, API key, context window, output limit and default model are synced from ClawPanel into DSH, and the panel reads back the result to confirm. The embedding story is the part worth studying: DSH keeps listening on loopback, while the remote web version reaches the full interface through short-lived tokens and a sandbox proxy, so no additional public port has to be opened. That is a considered answer to a real security question, and it is described in enough detail to be checked against the code.

Hermes Agent memory as three editable Markdown files

For Hermes Agent, ClawPanel focuses on making the agent durable rather than disposable, in the README's own framing of upgrading an agent from a one-off chat window to a continuously operated system. Long-term memory is visualized as three Markdown layers, Notes, User Profile and Soul, editable in place, so facts, preferences and personality persist as reviewable files rather than opaque state. Session management surfaces conversations, message flows, run status and tool-call details in one view, which addresses a genuine pain point with agent frameworks: understanding what a run actually did. Persona maintenance turns expression style, values and preferences into editable assets. Channel management centralizes connections to external messengers such as QQ, Telegram and Discord. None of this changes what Hermes Agent does; the panel's claim is that it makes the engine's existing state legible and operable, and the screenshots in docs/hermes-agent.md are there to back that up.

A web mode for ARM boards, and a Node version trap

Deployment flexibility is better documented here than in most hobby projects. A pure web mode drops the GUI dependency entirely, and the README lists Orange Pi, Raspberry Pi and RK3588 boards as targets, with npm run serve as the run command. A Docker ARM64 image exists at ghcr.io/qingchencloud/openclaw:latest, and an Armbian, Debian and Ubuntu Server one-click script auto-detects the architecture, documented further in docs/armbian-deploy.md. The version constraint deserves a warning label of its own. The ClawPanel web backend requires Node.js 18 or newer, but running the OpenClaw Gateway brings in OpenClaw's own engines.node requirement, and the README quotes OpenClaw 2026.7.1 as needing >=22.22.3 <23, or >=24.15.0 <25, or >=25.9.0. A board or server provisioned at Node 18 will start the panel and then fail on the gateway, so matching versions to the engine requirement up front is the practical first step for any deployment.

Documentation or storefront: the qt.cool ecosystem inside the README

The README blends technical documentation with commerce, and adopters should read it with that in mind. A free test platform called 晴辰云, hosted at gpt.qt.cool, is described as the companion service for model trials, with the panel itself explicitly not reading its keys automatically; users must register, claim quota on a check-in page, and paste the API key manually. Below that sits a sponsored promotion for 词元 API, a third-party OpenAI-compatible provider offering a promotional ¥1 to $1 credit top-up, complete with a Base URL, https://ciyapi.79tian.com/v1, wired into a configuration walkthrough. The project homepage itself is claw.qt.cool, the community links live under qt.cool, and an AtomGit mirror serves users in China. None of this is hidden, and the README does label the API section as third-party sponsored promotion with the note that its rules govern. But the overlap between the project's home domain and its recommended test platform means the documentation and the business interests are adjacent, and pricing or availability claims belong to those pages, not to the panel's engineering.

An AGPL badge over a licence GitHub cannot classify, at v0.21

Two maturity markers sit uneasily together. On one side, the release cadence is strong for a small project: v0.21.4 and v0.21.5 shipped on 2026-09-01, v0.21.6 on 2026-09-02, with a CI workflow badge and the last push on 2026-09-02, nine days before this review's reference date. On the other side, the version number itself says pre-1.0, and the licence situation is muddled: the README badge says AGPL-3.0, but the repository metadata classifies the LICENSE file as Other, a custom licence GitHub cannot categorise. If the file does turn out to be AGPL-3.0, that carries obligations around source availability for modified network-facing deployments that matter to anyone embedding the panel into a service; if it is custom, standard assumptions go out the window entirely. Read the file before shipping anything derived from it. As for alternatives, the honest one is going without a panel: DeepSeek Harness ships its own web interface, which ClawPanel itself embeds, and each engine can be run and configured by hand. The difference is exactly the product: ClawPanel unifies three engines' sessions, memory and model channels behind one window, at the cost of adopting a v0.21 tool with unresolved licence labelling.

Editorial conclusion

Operators already running OpenClaw, Hermes Agent or DeepSeek Harness who want one GUI for sessions, memory and model channels should try the web mode with npm run serve, or the Docker image, and confirm their Node version satisfies OpenClaw's engines.node range before anything else. It is the wrong pick for teams needing a stable 1.0 tool with a licence their policy team can accept, since the project sits at v0.21.x, the README badge claims AGPL-3.0 while GitHub classifies the file as custom, and the docs lean Chinese-first despite eleven README translations. Verify three things first: the actual LICENSE file contents, whether the pinned DeepSeek Harness version the panel installs matches what you need, and the terms of the sponsored API and companion test platform, whose promotions sit inside the documentation itself.

Official sources

1. Issues
2. Project website
3. qingchencloud/clawpanel on GitHub
4. README
5. Releases