These two GitHub rocket ship projects aren’t open-source victories at heart

Hundreds of thousands of stars — that’s not an open-source victory.

It’s the agent battlefront of two tech giants.

Let me state my judgment up front so you don’t misunderstand halfway through: this is not anti-open-source. It’s against the narrative that “open source” is being reclaimed by capital. The MIT license hasn’t changed, but whether you can actually access the model, access the distribution, access the retention — these have long since stopped being decided by the license.

1. April 4th

On April 4th I was running workstreams for a client using OpenClaw.

OpenClaw’s backend was connected to Claude Pro, and the frontend dispatched tasks through a WhatsApp gateway. One command, three weeks of automation running. Around 10 a.m. that day, the subscription suddenly disappeared — Anthropic unilaterally cut off Claude Pro/Max access on OpenClaw. A month later, OpenAI reciprocated by opening ChatGPT subscriptions to all OpenClaw users. About one month separated the two moves.

At that moment I re-examined both projects.

Last week I said Awakening Log #3 would be about “evaluation is the bottleneck” — I didn’t toss that piece; I moved it to #4. This week something more urgent than evaluation surfaced: OpenClaw and HermesAgent, these two GitHub rockets, are being celebrated by the market as an “open-source explosion,” but standing beneath them are the two agent battlefronts of OpenAI and Anthropic.

In the May 2 article “Anthropic’s Restraint Story is Falling Apart”, I laid a thread: Anthropic wants to guard the model, OpenAI wants to grab distribution. This week that thread bore its first fruit at the agent layer.

2. Let’s clearly lay out both projects

Readers who haven’t used them, here are two quick summaries.

OpenClaw: Austrian independent developer Peter Steinberger released it in November 2025. Originally named Clawdbot, it was renamed Moltbot on 2026-01-27 due to an Anthropic trademark complaint, and three days later renamed again to OpenClaw. It’s a project that wraps an agent runtime inside a messaging gateway — in plain terms, it first unifies 24+ communication channels like WhatsApp / Telegram / Slack / iMessage / WeChat / QQ / Feishu via a WebSocket gateway, then lets an LLM-agnostic agent (i.e., not tied to a specific model, supports Claude / GPT / DeepSeek) run behind it. On GitHub it has hundreds of thousands of stars, millions of monthly active users, and the community has contributed tens of thousands of skills. As of this writing, it’s one of the fastest-growing projects in GitHub history.

HermesAgent: Nous Research (a VC-backed open-source AI lab) released it under the MIT license on 2026-02-25. The architecture is exactly reversed — it’s an agent execution engine wrapping a gateway. Its core isn’t broad channel support; it’s a closed-loop learning cycle: after an agent completes a complex task, it enters an automatic reflection phase, writes the successful workstream into a Markdown skill, and stores it in ~/.hermes/memories/. The next time a similar task appears, it calls the skill directly. All sessions are stored in SQLite + FTS5 full-text search (a local database with full-text indexing); you can search things discussed weeks ago. The latest version, v0.10+, ships with over 100 built-in skills, a three-layer memory architecture, and 6 messaging platform integrations. As of April 2026, there have been no agent-specific CVE disclosures.

One grows wide, the other grows deep.

3. Architectural difference is philosophical difference

Remember one analogy.

OpenClaw is like installing doorbells on every door and routing all the rings to a single head butler; HermesAgent is like raising a chef who figures out recipes on his own — doorbells are just his tools.

The most visible cost of this difference falls on security. In March 2026, OpenClaw disclosed multiple CVEs (vulnerabilities recorded in public vulnerability databases) all at once, one of which was a WebSocket cross-site hijacking type — meaning an attacker could use a malicious page in the browser to hijack the WebSocket connection between OpenClaw and external services, diverting or replaying messages. The CVSS score (Common Vulnerability Scoring System) was near the maximum tier. It’s not an engineering flaw; it’s a structural problem — a gateway architecture by nature lays the attack surface of 24+ communication protocols bare on the table; every additional integration adds another exposed patch.

During the same period, HermesAgent had no agent-specific CVE disclosures. It certainly has its own issues (I’ll talk about them shortly), but the exposure surface is an order of magnitude smaller.

Multi-agent behavior differences form another watershed. OpenClaw supports persistent collaborative teams — multiple agents can run continuously, leave messages for each other, and form long-term working relationships. HermesAgent uses parent-child isolated execution — a child agent disappears once the task is done, results are sent back to the parent, and agents cannot talk to each other directly. The former resembles a permanent project team, the latter a one-time outsourcing.

Do you want an assistant that’s everywhere, or a companion that grows alongside you?

Architecture choice isn’t a technical preference; it’s a philosophical preference.

4. Stars are the cheapest signal on GitHub

I need to pause here.

The number “hundreds of thousands of stars” shouldn’t be the basis for discussing “open-source prosperity.” A single line of JS can farm ten thousand stars; you can buy a thousand for five dollars on Fiverr; hanging on the GitHub trending page for half a day rolls in ten thousand organic clicks. It’s an entry metric, not a health metric.

What actually tells you “does this project truly belong to the community” are the three things below.

First, commit concentration.

The vast majority of commits in the OpenClaw main repo come from the top 5 maintainers. Steinberger is the core among them. Since he joined OpenAI in February, no new maintainers have been publicly announced — not that no one is committing, but the commit list hasn’t been updated. A project that claims to have been handed over to a foundation can’t even say “who is in charge now.”

Second, the governance vacuum of the foundation.

In February 2026, Steinberger announced he was joining OpenAI to lead a “next-generation personal agent strategy” and at the same time publicly said OpenClaw was being transferred to an independent nonprofit foundation. Three months later, as of the day I’m writing this article, none of the foundation’s governance documents, board member list, or funding sources have been made public. An “independent foundation” that in three months hasn’t held an open meeting, hasn’t disclosed donors, hasn’t changed maintainers — what is the difference between that and an empty label slapped on a Notion page?

Third, the invisible moves of model binding.

OpenClaw is verbally LLM-agnostic, and the documentation clearly states that Claude / GPT / DeepSeek can all be connected. But the de facto binding isn’t in the documentation; it’s in the subscription path.

• In early April, Anthropic cut off Claude Pro/Max access on OpenClaw.
• In early May, OpenAI opened ChatGPT subscriptions to all OpenClaw users.

One cut, one open, a month apart. On the surface, it’s still LLM-agnostic. In reality, for an ordinary user, the only line that can be reliably called is the ChatGPT line. That’s de facto binding — not in the license, but the moment you top up.

In the May 2 Awakening Log I said, “OpenAI is grabbing distribution, Anthropic is guarding the model.” The agent layer is the first time this chess game hit the table.

The license is MIT, but the distribution channel is a single point of failure. In this era, the definition of “open source” has already been half swapped out.

5. Two battlefronts, two kinds of contest

Lift your gaze up a level.

OpenClaw is taking the “distribution absorption route,” with OpenAI behind it. The logic is simple: rather than clash head-on with Anthropic at the model layer, it’s better to swallow millions of end users directly at the agent layer. Steinberger joining OpenAI to lead the “personal agent strategy,” the foundation remaining undisclosed three months later, and users getting access to ChatGPT subscriptions — this isn’t a coincidence; it’s a roadmap. In the May 9 Awakening Log I calculated that trillion-dollar capex bill — Big Tech poured money into upstream compute; who will take the downstream? OpenClaw’s millions of monthly active users are the entry point that catches a piece of the downstream.

HermesAgent is different.

Nous Research is a VC-capital-backed open-source AI lab. The VC path has its own exit cycle — generally within 5–7 years an LP must be given an exit, either an IPO or an acquisition. Hermes v0.10 has been out for less than a year; its valuation hasn’t reached IPO scale yet. So within 6–18 months, there will inevitably be a move of “the team being acquired en bloc by a major lab” or “the core team being poached en masse by a major lab.”

This is a specific, falsifiable prediction. Not a vague “sooner or later,” but “within 18 months.” You can save this article and come back to reconcile in November 2027.

Anthropic’s side is the “model protection route.” The moment it cut OpenClaw’s access in April was the moment it publicly told the market: I will not subsidize the agent layer. Subscription profits from Claude Pro/Max are the lifeblood; the agent layer can be left to third parties, but you can’t use my subscriptions to feed someone else’s gateway — this move is cold, even restrained. But restraint is not neutrality. It just isn’t grabbing distribution; that doesn’t mean it will support Hermes. Hermes has so far received no official support from either Anthropic or OpenAI precisely because it isn’t worth either side’s attention right now — but when Nous Research’s VC exit window arrives, at that moment it will become very worth their while.

Anthropic guards the model, OpenAI grabs distribution, Hermes waits for a sugar daddy.

None of the three is doing “pure open source,” but the market loves this narrative, so the stars on GitHub keep climbing.

6. How independent developers land on solid ground

After ten thousand characters of grand narrative, back to people like us.

If you write code — Claude Agent SDK or Claude Code is the most stable choice. Use the model provider’s own SDK directly and bypass the uncertainty of third-party frameworks. You’re betting on Anthropic as a company, not on some GitHub project and its “foundation.”

If you do life automation — OpenClaw still works now, and the community ecosystem is good. But prepare a contingency plan for “the day it loses LLM access.” Copy the workstreams that run stably into pure Python or pure Markdown so they can run without OpenClaw. Spending two extra hours on this step now can save you two weeks later.

If you do research or repetitive tasks — HermesAgent’s skill learning closed loop is worth using. Markdown skill files can be shared, inherited, and versioned; this truly has value. But do not trust its self-evaluation. The community consensus is a thumbs-down: when a task gets botched, it still thinks it did beautifully. Even worse, it will overwrite your manually edited skill files with its automatic learning. Either turn off self-evaluation and write your own scoring tool, or lock ~/.hermes/memories/ as read-only and maintain it by hand.

Finally, the real moat.

It’s not which framework you use; it’s writing your skills, memories, and workstreams as portable Markdown. OpenClaw’s skills use the open standard from agentskills.io, and Hermes is also compatible with this standard — this is an unexpected window right now. Before this window closes, turn your agent assets into a format decoupled from the framework. The day OpenClaw’s access is cut or Hermes’s team gets poached, you git clone your own skills repo, connect a new runtime, and it just works.

You shouldn’t remain loyal to any particular framework; instead, make your agent assets portable across any framework.

7. Three predictions, written at the end

I’m putting down three specific predictions for this piece. They can be reconciled.

• Within 6 months (before November 2026), OpenAI will integrate OpenClaw into the official traffic distribution chain of ChatGPT Pro — possibly a “connect your agent” button, maybe embedding the OpenClaw entry directly in the ChatGPT app. Logic: millions of monthly active users just sitting there — OpenAI won’t limit itself to the light touch of merely “opening subscriptions.”
• Within 12 months (before May 2027), the HermesAgent core team will be poached in its entirety by Anthropic or a top-tier AI lab, or Nous Research will be acquired. Logic: Nous Research’s VC exit cycle won’t wait longer; Hermes’s memory system and self-improving skills are capabilities that Anthropic lacks.
• Within 18 months (before November 2027), OpenClaw’s foundation will publicly disclose its governance structure, but by then Steinberger will be substantially decoupled, and the governance documents will only be a post hoc ratification — the foundation’s board will most likely be a mix of former OpenAI executives and a few community representatives.

If you think Hermes won’t be absorbed and the OpenClaw foundation will grow into a true community governance body, feel free to leave your judgment in the comments. I’ll come back to reconcile in six months, then again at twelve months and eighteen months. That’s the tradition of the Awakening Log: leave predictions, reconcile accounts, no fluff.

In the next Awakening Log #4, I’ll finish that piece on “evaluation is the bottleneck” — why 99% of AI projects die at the landing stage not because the model isn’t strong enough, but because no one can quantify “how good is my AI’s output, really?” Hit “Wow”, and next Saturday at 8 p.m. let’s keep talking.

Forward this to that friend who’s still building automation with OpenClaw or Hermes — they deserve to know.

---

Publish log

• WeChat Official Account: Published (sent directly by user on the morning of 2026-05-16, did not use the 20:00 scheduled send)
• Draft link (preserved): https://mp.weixin.qq.com/s/g3e68sEFX7LjzKseC9gHX7-JVqdOhvE73PmmTDaaKud8grylWL1EBnL87KpkbNab (2026-05-16 00:25)
• ⚠️ Xiaofei generated 2 duplicate drafts at the time; user manually deleted the extra — incident captured in xiaofei-publisher.md “Publish idempotency hard constraint”
• ⚠️ The wikilinks of diagram1/diagram2 in the body were not converted by the remote publish tool, and bare ![[...]] text appeared in the article — captured in “Obsidian wikilink conversion hard constraint”