Grounding: this article is built directly from the transcript_content, summary_content, and summary_key_takeaways columns of public.videos for the §1.6 videos, plus the original Course 1: Picking Your Agent Harness archive. Pricing, model names, and feature claims are quoted from the source. Where the source is silent, the article flags it with a > NOTE:.

The sixth and final subtopic is the one the rest of the course is built to defend: OpenClaw is not dead. The channel's framing in the migration video was "essentially OpenClaw rebuilt by Nous Research," not "OpenClaw is obsolete." OpenClaw is the substrate that Mavis and Hermes both build on — the skills, the cron, the Discord bridge, the nightly context survival pattern, the food-tracking-app proof-of-ship, the "AI Agent Team that runs your business for $3/month" project. ~302K of the channel's ~764K total long-form views are OpenClaw-tagged. The §1.6 subtopic is the case for keeping OpenClaw — when it still wins, what it still does better than Mavis or Hermes, and the power-user patterns the channel itself uses day-to-day.

What you'll learn

  • OpenClaw's "skills" are persistent recipe books that survive the nightly context wipe, and they're the reason an OpenClaw coding agent gets more reliable over time, not less — the host calls them "the number one most important thing you should get set up on your open claw" (obET69yycFc, transcript).
  • You can run OpenClaw on a $2/month Zebra VPS with 2 GB RAM, swap the backend from Opus ($5/M tokens) to a Chinese model like Minimax (~$0.30/M tokens) — roughly a 1/16th price cut. The channel's own Opus run burned $30 in a single hour.
  • The OpenClaw 3.7 → 3.8 release arc added context engine plugin (Discord thread context survives), ACP persistent binding (Claude Code remembers across sessions), and ACP provenance (every coding task gets a session ID + origin tracing). These are the features the Claude Code + OpenClaw track in Course 4 §4.3 is built around.
  • NemoClaw is the privacy layer on top of OpenClaw — an open-source security wrapper developed with Peter Steinberger, the OpenClaw creator. It's a wrapper, not a fork: it adds isolated sandboxes, policy-based guardrails, network isolation, data privacy controls, a privacy router, and a hybrid model strategy. The 31,868-view install video (qEFaeLlfLmk) is the canonical entry point.
  • The "AI Agent Team that runs your business for $3/month" project is the channel's worked example of the multi-agent OpenClaw pattern at full power. The video (t6997xGb0gg, 1,495 views) is the channel's playbook for a multi-agent business rig.
  • MaxClaw (42,714 views, the #1 most-viewed video on the channel) is the hosted OpenClaw — the right starting point for users who want the substrate without the install.

Why OpenClaw still wins

The argument for OpenClaw in 2026 is not "OpenClaw is the best harness" — the §1.4 matrix makes clear that Mavis, Hermes, and Claude Code each win on at least one axis. The argument is that OpenClaw is the substrate that Mavis and Hermes both build on. Mavis is a curated multi-agent product that sits on top of MiniMax's hosted stack; Hermes is a builder's tool that sits on top of the OpenClaw skill pattern. The Skills architecture, the cron pattern, the Discord bridge, the food-tracking-app schema — these are OpenClaw primitives, and they show up in Mavis and Hermes under different names.

NOTE: the "substrate" framing is a synthesis from the source videos, not a direct quote. The channel's framing in the migration video is "Hermes is essentially OpenClaw rebuilt by Nous Research," which implies substrate continuity. The "Mavis = curated Mavis = curated on top of Mavis's own stack" framing is a synthesis from the Mavis video's "they already set it up for you" line. Treat the substrate framing as a reading, not a citation.

The Skills architecture — why OpenClaw beats a one-shot CLI

The core architectural difference between Claude Code and OpenClaw-as-coding-agent is persistence. Claude Code is a Node CLI — you start a session, hand it a brief, get a diff, the session ends. OpenClaw is an always-on agent, and the host's position is that the context window clears overnight, so anything stored only in chat memory is gone by morning. A skill persists. The channel's summary_verdict on the setup video names skills as "the single most important thing to set up on your OpenClaw bot" (obET69yycFc, public.videos.summary_verdict).

The setup is deliberately low-friction: the channel's own presentation skill was built by telling the bot his preferences and saying "save this up as a skill" (Michael, obET69yycFc, transcript). The bot wrote the spec, structure, host location, and chained a deep research skill — none of it required hand-writing markdown. Refinement stays conversational: "update your skill on making presentations," "I like light theme," or "you didn't do a good enough job on research… update that in your skill" are the actual commands.

A second skills video deepens the rationale.

The deeper claim is that skills beat memory. The channel's bot can do it right once, maybe twice, and then the third time it messes it up — hallucinating tweets, shuffling presentation slides. The fix isn't a better prompt; it's moving the workflow out of memory and into a skill file. Skills push only metadata into the context window at session start, then auto-reload the full directory when the agent needs them, which means you can ship "hundreds or thousands of skills" without bloating the prompt (txTowBmYXMc, public.videos.summary_key_takeaways). The channel's own bot already runs more than ten skills.

The threshold is concrete: once context passes ~40%, the agent enters what the host calls the "dumb zone" (txTowBmYXMc, public.videos.summary_key_takeaways: "if it crosses 40%, your agent enters the dumb zone") — instructions get under-weighted, output drifts. The fix is to move every recurring workflow from a memory point into a skill. Cheap Chinese models on a Minimax Mac setup are the channel's preferred backend specifically because they perform well with the skill-metadata pattern; "spending a bunch of money and getting lost" is the alternative the channel wants to avoid (txTowBmYXMc, summary_verdict).

Two warnings from the setup video that are easy to miss

(obET69yycFc, summary_key_takeaways):

  • Clawhub has had "issues with viruses or whatnot," and the channel does not "fully say, 'Hey, browse skill hub.'" Use it for inspiration only — clawhub is treated as a reference catalog, not a package registry.
  • Someone else's skill encodes their preferences, not yours. Have the bot write the skill to your spec, not install someone else's.

The 3.7 → 3.8 update arc — context engine, ACP binding, ACP provenance

The 3.7 release dropped on March 7th (3.7) with a marketing-friendly "40+ features and 30+ bug fixes" changelog. The honest take from the channel is that life after the update is "slightly" better — the headline is the headline, but the workflows that actually changed are concentrated in three places. The first two are deeply relevant to the Discord bridge: the context engine plug-in system and ACP persistent binding.

Context engine plugin — fixes the Discord thread problem. Before 3.7, opening a Discord thread for a new task meant the main branch "doesn't update the main chat" — your agent had no idea what you'd just discussed. The new plugin connects to the main branch itself, so "everything ties together. It's more coherent, more cohesive." If you run more than one OpenClaw instance (the channel runs the Avengers cast), this is the change that actually improves your daily flow, not one of the other 36 bullet points on the changelog.

ACP persistent binding — makes Claude Code usable across sessions. The second relevant change is for users who wire Claude Code into their OpenClaw agent. The host built the boxminingai.com site itself this way: he brainstormed the dashboard layout with his agent, then told the agent to use Claude Code to execute. Before 3.7, every time you started a new session, Claude Code had to be re-briefed on what it was building. ACP persistent binding "remembers what Claude Code has been building" across sessions, which "substantially" improves the loop when you're bouncing between planning (OpenClaw) and execution (Claude Code / Windsurf). The takeaway: enable this before you start any Claude Code project, or your agent will re-brief Claude on every session and you'll feel like nothing is sticking.

Model cost reality check. The host runs the most expensive tier (Opus) and is burning $30–$60/day on it. The co-host is on the Minimax developer plan, which is a flat $20/month, and reports it working "substantially better" — but only after about two weeks of context optimization. The 2-week number is the one to remember: don't judge a new model setup on day one.

OpenClaw 3.8 dropped on March 8 — one day after 3.7 — with a single workflow-changing feature: ACP provenance. With ACP provenance, every coding task gets a session ID plus origin tracing, so you can see what passed, what failed, and where it broke. The host pair it with is Cloud Code — "really good at programming" — used to vibe-code the new boxmoneyai.com site (Windows 95 UI included). For a track that hands work from OpenClaw to Claude Code to Cloud Code, provenance is the thread that ties the three logs together.

A small but important version note: thread awareness (where the agent can read what happens inside a Discord thread, so context no longer dies at the main chat) was actually shipped in 3.7, not 3.8. The video muddles the two versions; the rest of the 3.8 surface area (Telegram streaming, native OpenAI GPT support) is 3.8-specific.

NemoClaw — the privacy layer

NemoClaw Setup Guide: FASTEST Way to Install (31,868 views) is the second-most-viewed video on the channel and the canonical entry point for the privacy layer on top of OpenClaw. NemoClaw is an "open-source security privacy layer for openclaw" developed with Peter Steinberger (the OpenClaw creator). It is a wrapper, not a fork — it adds isolated sandboxes, policy-based guardrails, network isolation, data privacy controls, a privacy router, and a hybrid model strategy on top of the OpenClaw foundation.

The minimum stack the channel ran:

  • Zebra server: $2/month, 2 GB RAM is sufficient
  • Discord bot token: a new one per NemoClaw agent
  • Model: free tier works — nvidia.com/nemotron-3-super on OpenRouter (the creator mis-spoke it as "Neotron 3 Super" in the video) at $0/M input and $0/M output tokens. Alternatives: Hunter Alpha or minimax-m2.5 for stronger task execution

One hard line from the transcript: the creator refused to retrofit NemoClaw onto his existing production OpenClaw agents ("we have a good thing going already"). For now it runs as a separate agent on Discord only — a deliberate non-disruptive rollout. This is the substrate pattern: NemoClaw runs alongside your existing OpenClaw install, not on top of it.

MaxClaw — the hosted on-ramp

MaxClaw: One-Click to Set Up Openclaw FULLY (42,714 views) is the #1 most-viewed video on the channel and the canonical hosted-OpenClaw walkthrough. The hosted variant is a sandboxed OpenClaw running on MiniMax servers for $20/month — the recommended starting point for most people. Sign in at agent.minia.io with Google. No install, no OpenAI embeddings to enable, no chat-app wiring, no memory configuration. Pay $20/month and the box is live.

The sharpest claim in the video is about sandboxing. The creator warns that installing OpenClaw directly on a Mac Mini means granting the agent "system access," which he calls "extremely dangerous" because current models still hallucinate. He has the receipts: a message the agent sent to his girlfriend at 3am while running with full system access. MaxClaw keeps the agent sandboxed while still giving it web access, so you can run research and due-diligence tasks without handing over your life. The channel used it to have MiniMax do due diligence on MiniMax itself and got a usable report back.

The deployed OpenClaw is v2.15 at test time — explicitly a "slightly older version" than the main build. If you need the latest features, verify the version before relying on the hosted instance for production agent teams. Your identity, SOUL, and bootstrap files are pre-filed and downloadable — download them on day one so you can audit exactly what the pre-configured agent is operating on. A Mini Max mobile app (Apple + Android) mirrors the same agent. Phone and browser stay in sync.

The "AI Agent Team that runs your business for $3/month" project

How to Build an AI AGENT TEAM That RUNS YOUR BUSINESS for $3 month (1,495 views) is the channel's worked example of the multi-agent OpenClaw pattern at full power. The 1,495-view count is the channel telling you this is a niche audience — most users don't need to run a business on OpenClaw. But for the users who do, the video is the channel's playbook for a multi-agent business rig: a $3/month VPS, the Avengers-themed agent setup (Stark, Banner, Loki, etc.), the skills + cron + Discord bridge, and the worked examples (daily briefing, food-tracking app, presentation skill) that prove the pattern.

The video is the capstone for the OpenClaw track. The cross-course capstone (the §1.4 matrix) is the harness decision; this video is the proof that the OpenClaw substrate can carry a real workload.

The "dumb zone" at 300 lines of soul.md

This is the most important operational detail in the OpenClaw-as-coding-agent setup. Once the soul.md file swells, the model degrades sharply. The host's original soul was 300 lines. At that size the agent enters a "dumb zone" where the model "starts messaging your girlfriend instead of building a presentation" (Michael, 258R3kzDRAQ, transcript). Cap soul.md at 15–30 lines and trim agents.md (the bootstrap file) to the minimum.

The 40% context threshold from the txTowBmYXMc video is the same idea in different units: once context passes ~40%, the agent enters the dumb zone. The two metrics are correlated but not identical — soul.md is the static persona file, context window usage is the dynamic per-session budget. Keep both capped.

The cheap self-hosted path

The cheapest non-free path the channel endorses: a $2/month Zebra VPS with MiniMax M2.5. The creator picks Zebra over DigitalOcean and AWS because the entry tier is $2/month for 2 GB RAM and 40 GB storage; a promo bumps that to 4 GB RAM for $3. The instance is online 24/7, reinstallable, and serves as a throwaway sandbox — if OpenClaw goes rogue, you reset the box and you're back to a $2/month bill.

The install itself is short:

  1. Connect to the Ubuntu box via Termius
  2. Grab the Mac/Linux install command from openclaw.ai and paste it
  3. The installer runs in about 2–3 minutes
  4. If you see a path missing warning, export the path and run openclaw tui to launch

Two commands worth memorizing for the long haul:

  • openclaw onboard — full re-setup if you break it
  • openclaw configure — edit model, gateway, skills, or channel connections later, without re-onboarding

The 2,810-line case for OpenClaw

The brief for this course is the 2,810-line archive of the original Course 1: Picking Your Agent Harness — every OpenClaw video, every release-cadence deep-dive, every update-track the channel covered from v2.25 to v4.7. That's not a course you delete when the flagship moves. That's a legacy depth that the new harness is built on top of.

The implication: even if you migrate to Hermes, the OpenClaw material is still relevant. The four channels of legacy depth are:

  1. The update cadence (the v2.25 → v4.7 release arc, ~10 weeks). The original course covers every release in chronological order, with the security/agent-loop/context/memory/UI/install axes mapped. Hermes has its own release cadence, but the pattern — fast releases, 24-hour hotfixes, /verbose flags, snapshot habits — is the same.
  2. The Skills architecture (the persistent skills, the 40% context threshold, the "dumb zone" failure mode). Hermes's Skill Bundles are structurally similar to OpenClaw's skills, but the philosophy — skills beat memory, ship hundreds not thousands — is OpenClaw's contribution.
  3. The Discord bridge (channel-per-agent, threads-per-subtask, the 3.7 context engine plugin, the two intents that have to be on). Hermes has its own chat-platform story, but the Discord-bridge pattern is OpenClaw's contribution.
  4. The "dumb zone" failure mode (the 300-line soul.md problem, the 40% context threshold, the 24-hour context window). Every long-running agent hits the dumb zone eventually. The OpenClaw material is the canonical reference for the failure mode, and the lessons transfer to Hermes.

The 2,810-line archive is not obsolete. It's the foundation. The Hermes course (Course 3) cites back to it. The Mavis article (§1.1) cites back to it. The Claude Code article (§1.3) cites back to it for the OpenClaw-as-coding-agent framing. The migration article (§1.5) cites back to it for the OpenClaw → Hermes story. The decision matrix (§1.4) cites back to it for the hosted vs self-hosted axis.

Why the substrate framing matters for the harness decision

The substrate framing is not just a rhetorical move — it has a concrete implication for the harness decision. The implication is: the harness you pick in 2026 should be the one with the deepest substrate underneath it. If the substrate is OpenClaw, then the harness you pick is one that builds on OpenClaw — Mavis (which uses OpenClaw primitives), Hermes (which is "OpenClaw rebuilt by Nous Research"), or OpenClaw itself.

If the substrate is Claude Code (a CLI), then the harness you pick is one that wraps Claude Code — Kilo Code, Open Claude, Grok CLI, the various Anthropic-compatible clients. These harnesses are Anthropic-shaped but not OpenClaw-shaped; the model layer is the same, the agent loop is different.

The 2026 substrate ranking, distilled from the channel's coverage:

  1. OpenClaw substrate — the deepest. ~302K views, the most videos, the longest release arc, the most documented failure modes. Mavis and Hermes both build on this substrate.
  2. Claude Code substrate — the most composable. The settings.json swap pattern means any Anthropic-compatible backend can route through Claude Code. Kilo Code, Open Claude, Grok CLI all build on this substrate.
  3. Mavis substrate — the most curated. Single-product, no composition story, but the curated presets are a real advantage for non-technical users.
  4. Hermes substrate — the most ambitious. The Kanban + Dashboard + Skill Bundles + MCP server mode is the most feature-rich substrate in the channel, but it's also the newest and least battle-tested.

The harness decision is, in part, a substrate decision. The §1.4 matrix's "best for" column is downstream of the substrate ranking.

The food-tracking app and the worked examples

The OpenClaw substrate is not just a list of features — it's a set of worked examples that prove the pattern. Three worked examples from the source material:

  1. The food-tracking app (referenced in the Skills videos and the MaxClaw video). The channel's bot built an end-to-end food-tracking app using a local SQLite database, a photo column for audit, and a persistent skill so the agent doesn't re-learn the schema each session. The proof: "no-code" doesn't mean "no real software." A self-hosted agent can ship a real product.
  2. The "AI Agent Team that runs your business for $3/month" project (t6997xGb0gg, 1,495 views). The channel's capstone for the OpenClaw track — a multi-agent business rig on a $3/month VPS, with the Avengers-themed agent setup (Stark, Banner, Loki), the skills + cron + Discord bridge, and the worked examples (daily briefing, food-tracking app, presentation skill).
  3. The daily news briefing (referenced across the cron + Kanban videos in Course 3 §3.3). A 14-step sequential cron that became a 9-worker parallel pipeline after the cron + Kanban pairing. The proof: the same workload that took 14 sequential steps now runs in 9 parallel workers, with the parent-child retry loop absorbing the failures.

The three worked examples are the receipts for the OpenClaw substrate. They prove that the substrate ships real software, not just a chat interface.

The NemoClaw privacy layer

NemoClaw is the under-discussed piece of the OpenClaw substrate. The 31,868-view install video (qEFaeLlfLmk) is the second-most-viewed video on the channel, which is the channel's signal that privacy is a real concern for the OpenClaw audience. NemoClaw is a wrapper, not a fork — it adds isolated sandboxes, policy-based guardrails, network isolation, data privacy controls, a privacy router, and a hybrid model strategy on top of the OpenClaw foundation.

The NemoClaw pattern is structurally similar to a Docker + network-namespaces setup, but with a higher-level policy language. The implication: NemoClaw is the right OpenClaw substrate for users handling sensitive data (medical, legal, financial) who want the self-hosted control without losing the safety properties of a sandbox.

NOTE: the exact NemoClaw policy language is not in the source video transcript excerpt I read. The 31,868-view count is the channel's signal that this is a popular install path; verify the policy language on the current NemoClaw docs.

Try it yourself

The hands-on goal: stand up an OpenClaw install that does one real 24/7 job, prove the skills + cron pattern, and only then consider whether to migrate to Hermes or layer Mavis on top.

  1. Pick a starting point. If you only want to test the agent lifecycle, sign up for MaxClaw (42,714-view video) at the MiniMax-hosted URL with Google auth, burn the welcome credit, and use the export buttons for SOUL.md, HEARTBEAT.md, and MEMORY.md when you're ready to migrate. If you want production control, skip MaxClaw and provision a $2/month Zebra VPS (2 GB RAM, 40 GB) — the 4 GB promo tier is $3 if you need the headroom.
  2. Install OpenClaw. Connect via Termius and paste the Mac/Linux command from openclaw.ai. The installer takes 2–3 minutes. If you see a path missing warning, export the path and run openclaw tui.
  3. Memorize the two recovery commands. openclaw onboard (full reset) and openclaw configure (edit model / gateway / skills / channels). These are the two commands you'll use most.
  4. Pick a model key deliberately. For MaxClaw, do not paste an existing Minimax coding key — the bot will stop responding. For the self-hosted path, use the $20/month MiniMax coding plan key (M2.5, 300 prompts / 5 hours) and skip the Lightning variant. Switch to nvidia.com/nemotron-3-super on OpenRouter ($0/M tokens) if you only need a free test model.
  5. Build one skill. Tell the agent your preferences and say "save this up as a skill." Verify the skill landed in the skills directory. Refine it conversationally — "update your skill on making presentations" is the actual command.
  6. Wire a cron job. Use the cron jobs tab in the OpenClaw dashboard (or openclaw configure from the terminal). Pick a daily 9:00 a.m. briefing, a per-run prompt, and a delivery channel (Discord or Telegram). The cron pattern is the OpenClaw advantage over Claude Code (§1.3) and Mavis (§1.1).
  7. Cap soul.md at 15–30 lines. The 300-line soul put the channel's bot in the "dumb zone." Yours will too. Trim.
  8. Update via terminal. openclaw update from your VPS terminal, not from the agent. Agent-driven updates hit permission errors and disconnects.
  9. Audit after every update from 3.2 forward. Run openclaw doctor to flag missing dependencies.
  10. Decide: stay on OpenClaw, migrate to Hermes, or layer Mavis on top. The §1.4 matrix is the decision frame. The §1.5 migration script is the cross-platform path. The §1.1 Mavis video is the curated-on-top path. Pick one and commit for 30 days.

Common pitfalls

  • Going past 40% context with MiniMax. The model gets visibly dumber. Restart before you cross the line, not after. The openclaw configure flow has a Restart button that clears the context window in ~10 seconds and preserves long-term memory.
  • Granting system access on a Mac Mini without a specific reason. The 3am-message incident is real. Use the sandboxed hosted variant first. The nhDA7tcQtx0 video (4,158 views) is the full case against the Mac Mini.
  • Picking the annual MaxClaw plan during evaluation. The transcript says: monthly first. Annual plans lock you in before you've validated the workflow.
  • Pasting your existing Minimax coding key into MaxClaw. Confirmed broken by the creator — the bot stops replying. Use the welcome credit instead, or skip MaxClaw entirely.
  • Picking the Lightning variant on the MiniMax coding plan. It's billed outside the coding plan and will burn the wrong wallet. Stick to M2.5.
  • Treating skill overlap as cosmetic. On OpenClaw, the thumbnail generator and image generator clashed silently. Audit any merged-skill agents after migration. The Hermes vocal tool-use mode (see Article 1.2) surfaces overlap that OpenClaw hid.
  • Confusing 3.7 and 3.8. Thread awareness is 3.7, provenance and Telegram streaming are 3.8. Tutorials that mix them are mixing versions.
  • Letting soul.md bloat past 30 lines. The model enters the "dumb zone" — wrong targets, ignored instructions, off-topic replies. Cap it.
  • Trying to repair a polluted agent directory in place. If the agent has been running for months and behaviour is degrading, nuke the directory and reinstall from scratch via SSH. Don't try to fix it incrementally.
  • Trusting the 1-week free Hunter Alpha / Healer Alpha models as SOTA. The hosts say early tests are "not super good" — the models are stealth drops optimized for OpenClaw agentic workflows. Treat output as agentic-grade, not state-of-the-art.
  • Updating through the agent. Agent-driven updates hit permission errors and disconnects. Always openclaw update over Termius.
  • Relying on clawhub as a package registry. Clawhub has had "issues with viruses or whatnot." Use it for inspiration only.
  • Hiding raw tool errors behind 2.26's default redaction. Tool failure replies now hide raw errors unless you pass /verbose. If you didn't know that switch existed, you will waste hours debugging silent failures.
  • Trusting "allow always" after 3.31. 3.31's tightened exact approvals caused "allow always" to behave like "allow once" — the host calls this approval amnesia. 4.1 fixes the trust persistence bug, but edge cases remain for wrapper paths and inline eval. If you upgraded to 3.31 without reading the changelog, your gateway could be quietly re-prompting on every command.
  • Skipping openclaw doctor after 3.2. 3.2 improved doctor specifically to flag missing dependencies. Run it after every upgrade from 3.2 forward.
  • Reading "Hermes is OpenClaw rebuilt" as "drop-in compatible." They are not — see Article 1.5. The data layer transfers; the runtime does not.
  • Treating OpenClaw as obsolete. ~302K of the channel's ~764K total long-form views are OpenClaw-tagged. The substrate is alive. The MaxClaw video (42,714 views), the NemoClaw video (31,868 views), the Skills video (obET69yycFc, 4,975 views), and the food-tracking app are all 2026-era work. The "AI Agent Team that runs your business for $3/month" project (1,495 views) is the channel's capstone for the OpenClaw track.

Sources