# OpenClaw and Hermes — verified at primary, and what comparison is possible Phase 2, 6 August 2026. Everything below was opened at source on that date. Secondary write-ups were used only as a finding aid and **none is cited**. --- ## Both exist, and both are real | | **OpenClaw** | **Hermes Agent** | |---|---|---| | Repository | `github.com/openclaw/openclaw` | `github.com/NousResearch/hermes-agent` | | Verified via | GitHub REST API, 6 Aug 2026 | GitHub REST API, 6 Aug 2026 | | Language | TypeScript | Python | | Stars at retrieval | 385,338 | 226,388 | | Licence | **MIT** (LICENSE file, "Copyright (c) 2026 OpenClaw Foundation") | **MIT** ("Copyright (c) 2025 Nous Research") | | Last push at retrieval | 2026-08-06T12:43:53Z | 2026-08-06T12:06:41Z | | Homepage | `openclaw.ai` | `hermes-agent.nousresearch.com` | | Description | *"Your own personal AI assistant. Any OS. Any Platform. The lobster way."* | *"The agent that grows with you"* | Both were pushed to on the day of retrieval. Neither is dormant. **A metadata correction worth recording.** The GitHub API reports OpenClaw's licence as `NOASSERTION`. The repository's own `LICENSE` file is a standard MIT licence. Several secondary write-ups assert "MIT" — which is right, but they were right by luck, not by checking. The API field is wrong and the file is authoritative. --- ## OpenClaw's architecture, from its own README - A **Gateway** — *"the local control plane for sessions, tools, events, and channel connections"*. - A **Control UI, CLI and TUI**, all of which connect to the Gateway. - **Channels** — WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage. - **Companion apps and nodes** adding voice, canvas, camera, screen and device-local actions. - Hosted and local **model providers**. - **Tools, skills and plugins** as the extension surface. Install is `npm install -g openclaw@latest`, requiring Node 22.22.3+, 24.15+ or 25.9+. ## Hermes's architecture, from its own README Installs via a one-line script that provisions uv, Python 3.11, Node.js, ripgrep, ffmpeg and a portable MinGit. Ships a CLI, gateway, TUI and tools, with native Windows support. **Hermes carries a `Migrating from OpenClaw` section** — `hermes claw migrate`, which detects `~/.openclaw` and imports settings, memories, skills and API keys. The first item in its list of what gets imported is: > **SOUL.md** — persona file --- ## The finding that matters for the paper **All four harnesses share a persona-file convention, and that is evidence of a common lineage rather than convergent design.** | Harness | Persona file | |---|---| | OpenClaw | `SOUL.md` | | Hermes | imports OpenClaw's `SOUL.md` directly | | **Axium** | `soul.example.md`, loaded as the cached static block (`sonnet.rs:536`) | | **Orange** | `soul_path` setting, `_load_soul()` with `_FALLBACK_SOUL` (`agent.py:64-86`) | The architectural parallel runs further than the filename. OpenClaw's Gateway/CLI/TUI/channels split is the same shape as Axium's `tui/server.rs` + `channels/{cli,telegram}.rs` + `worker.rs`. Both expose tools, skills and plugins as three distinct extension points. Axium's own README describes it as competing with agentic tools of this class. So the brief's description of Axium and Orange as OpenClaw-lineage harnesses is **supported by the source**, and the paper can say so with citations rather than as an assertion. --- ## What comparison is feasible The plan required classifying each third-party harness into one of three states. Both land in the same place, and it is better than expected. ### State (a) — runnable here: **technically yes, but it is not my call** The runtimes are present on this machine, verified 6 Aug 2026: | requirement | available | |---|---| | Node (OpenClaw needs 22.22.3+, 24.15+ or 25.9+) | **v24.18.0** ✓ | | npm | 11.16.0 ✓ | | Python (Hermes bundles 3.11) | 3.14.4 present ✓ | | **cargo — needed to build Axium at all** | **1.96.1** ✓ | Nothing blocks installation on capability grounds. Three things block it on judgement grounds, and they are recorded rather than quietly resolved: 1. Both official install paths are **`curl … | bash` / `iwr … | iex` remote scripts**, or a global npm install. That is third-party software being installed system-wide on the author's working machine. It is a consequential, outward-facing change and needs an explicit decision from him, not a loop tick. 2. A meaningful head-to-head needs **API keys wired into two more harnesses** and real token spend across a full task suite. 3. Each needs a **benchmark adapter** written against its CLI, which is a substantial build in itself. ### State (b) — architectural comparison: **available now, and well-sourced** Both projects' READMEs, licences, documentation sites and repository metadata are in hand and citable. A documented-design-against-documented-design comparison can be written today, clearly labelled as architectural rather than measured. ### State (c) — unverifiable: **does not apply** Neither had to be cut. The Phase 2 rule about recording cuts is not triggered. --- ## Decision, and what it means for the paper's framing **The paper does not promise a measured industry comparison, and must not.** What it can honestly deliver, in descending order of evidential strength: 1. **Measured**: Axium's and Orange's own components, against their own claims. This is where the real results are, and two are already in hand — the classifier at 36.3% against a claimed 60–70%, and the brownfield ablation showing the expensive coder bought nothing. 2. **Architectural, cited**: how the four harnesses solve the same problems — routing, persona, memory, channels, tool surface, extension model — with every claim about OpenClaw and Hermes sourced to their own repositories. 3. **Not attempted**: a measured head-to-head against OpenClaw or Hermes. Feasible, but requiring installs and spend that were not authorised. **Stated as not attempted rather than dressed up as a limitation of the tools.** The title and abstract must reflect that. A paper claiming to benchmark against leading harnesses would be overclaiming; a paper that measures two self-built harnesses honestly and situates them in a cited architectural landscape is what the evidence supports. **Open question for Nikolaos:** authorise the installs and the API spend and item 3 becomes possible. It would be the strongest section in the paper. Without that authorisation it stays out. --- ## Sources opened, 6 August 2026 | URL | What it is | |---|---| | `https://api.github.com/repos/openclaw/openclaw` | Repository metadata | | `https://api.github.com/repos/NousResearch/hermes-agent` | Repository metadata | | `https://raw.githubusercontent.com/openclaw/openclaw/main/README.md` | 111,109 bytes | | `https://raw.githubusercontent.com/openclaw/openclaw/main/LICENSE` | MIT | | `https://raw.githubusercontent.com/NousResearch/hermes-agent/main/README.md` | 17,688 bytes | | `https://raw.githubusercontent.com/NousResearch/hermes-agent/main/LICENSE` | MIT | | `https://docs.openclaw.ai/` | Documentation site, 200 | | `https://hermes-agent.nousresearch.com/docs/` | Documentation site, 200 | Not cited, used only to locate the above: assorted third-party blog posts and video write-ups surfaced by the search APIs. Several make claims about these projects that were not checked against the repositories, including the licence claim noted above.