Boxmining · Courses
Structured syllabi built from the BoxminingAI YouTube channel — the highest-viewed videos on each topic, in learning order. 18 courses · 545 videos.
The channel's coverage of the CLI-coding workflow, framed as the coding track's front door. Claude Code is Anthropic's agentic coding CLI; the channel's takes on it split into two camps: (a) is it the best coding agent *right now*? and (b) how do you route it through cheaper models so you're not paying Opus prices for every keystroke? This course is the entry point for anyone who wants to ship software with an agent.
The new front door for the site. The first decision in 2026 is not which model to use — it is which **harness** wraps the model. OpenClaw, Mavis, Hermes Agent, and Claude Code all sit on top of roughly the same set of frontier models; what differs is the agent loop, the multi-agent topology, the install footprint, the cost ceiling, and the audit surface. This course walks through the four harnesses the channel covers most, names the use case each one wins, and finishes with the migration path from OpenClaw to Hermes — the move the channel itself is making.
The 3-way race between Claude Code, Perplexity Computer, and Kilo Code — the three coding CLIs the channel's coverage keeps coming back to. The pitch in one sentence: Claude Code gives you control, Perplexity Computer gives you parallel-agent orchestration, and Kilo Code gives you a BYOK OpenRouter on-ramp. The mid-session escape hatch that ties the three together is the `/model` hot-swap.
The channel's multi-agent successor to OpenClaw. Where OpenClaw is one agent that you can extend with skills, **Hermes Agent is a team of agents** organised on a Kanban board, with a Dashboard, a Desktop App, a VPS option, a Skill Bundle primitive, a Computer Use mode, and a release arc the channel has covered in full (v0.060 → v0.16) across April–June 2026. If OpenClaw is "give one assistant a personality and memory," Hermes is "run a small company of assistants that hand work to each other, schedule themselves, and ship the receipts to a browser."
The unglamorous but essential course: *where does the agent actually run?* The channel's coverage splits into three camps — hosted (MaxClaw, Mavis, KiloClaw), VPS / cloud (NemoClaw, Hermes VPS, Nut Studio), and local (Mac Mini, NVIDIA, Qwen on your laptop). This course is the decision framework + the install guides, and it is the operational backbone of every other course on the site.
The framework the BoxminingAI channel uses to pick a model for an agent stack. Not a leaderboard, not a vibe — a routing rule. The pitch: every model is a tool, the right tool depends on the slot (orchestrator / executor / auxiliary), and the channel's coverage gives you the receipts for each.
Where do you actually talk to the agent? Discord and Telegram are the two surfaces the BoxminingAI channel uses — Discord as the always-on team surface (channel-per-agent, threads-per-subtask) and Telegram as the on-the-go mobile surface (streaming replies, quick queries from the phone). This course walks through both, the bot build for each, and the cross-surface discipline (one thread per topic) that keeps an agent from collapsing into a megathread after week one.
The channel's take on Anthropic in 2026 is unusually sharp: a single thesis ("you don't need Opus"), a single controversy (the plan-throttling saga, where Anthropic called the new limits a "feature, not a bug"), and a single exception (Fable 5 + Loop Designs, the one place where Claude-class output is genuinely worth the spend). This course collects all three threads into a coherent migration story — what changed in March–April 2026, what to do about it, and where the cheap Claude window closes. Across seven videos in the [Claude section of Course 2](03-model-tier-list.md) and the standalone Opus guide, the channel's coverage is the most systematic on the page, and it forms the empirical backbone of every recommendation in this course.
The "your agent broke" landing page. Every "my agent forgot," "my agent is dumb," "my agent fabricated data," and "my agent did something dangerous" video the channel has ever shipped lives here — consolidated from the nine source files into one troubleshooting track. The fixes split into memory layering (context window hygiene, embeddings, Obsidian, Honcho, `/goal`-style bootstrap) and behaviour hygiene (`SOUL.md` size, skill count, decomposition, sandboxing). If your OpenClaw or Hermes install has ever ghosted you, fired off a 3 a.m. message, or started hallucinating mid-task, start here.
The channel's coverage of Minimax, the Chinese model family trained on the OpenClaw Agent Harness framework. This course pulls together every video on Minimax in one place: the M2.7 review, the M3 drop, the Claude Code env-var swap, the Mavis beginner on-ramp, and the operational failure modes the channel has hit on its own stack — including the "dumb zone" that anyone running Minimax on a long-lived agent will eventually see.
The four Chinese labs the channel actually covers, side by side. This isn't a geopolitics course and it isn't a "they're copying Claude" think piece. It's the model-card summary, the routing decision, and the honest answer to "which one wins when" — Kimi as a model you wire in (skip the hosted wrapper), Qwen as the local-inference champion with a credible Plus cloud tier, DeepSeek as the family that puts reasoning *inside* the tool call, and GLM 5.1 as the new "Opus-at-the-McDonald's-price" play that just got 140% more expensive.
The two features that turn a chatbot into a *system*. Skills are the recipe book your agent follows on demand — each one is a markdown file in `~/.openclaw/skills/` that the agent loads only when the task matches. Cron is the trigger that fires those skills on a schedule, so the recipe runs while you sleep. The channel treats the pair as the channel's automation heartbeat: every morning-briefing, every Twitter-scrape, every sub-agent fan-out starts with a skill file and a cron line.
The channel's coverage of the high-volume auxiliary slot — the free-tier providers that power Hermes Agent's cheapest workflows, the Xiaomi Mimo V2 Pro model that became the "high-volume king" on OpenRouter, and the broader question of when a "cheap and good enough" model beats a flagship. This course pulls together the channel's framing of the auxiliary role, the BYOK free-tier pattern, and the Mimo V2 Pro free promotional period — the slot in the model hierarchy that you actually want to fill *first* before paying for anything else.
The course on the channel's two-pronged fix for "vibe-coded slop." Sub-agents give you context isolation — each worker sees only what its task needs, so cheap models stay in the smart zone. The Kanban gives you an inspectable orchestrator — durable, retryable, named-profile tasks on a board, with worker logs and a parent-child retry loop that a vanilla orchestrator does not ship. Together they turn multi-agent work from "the bot is gaslighting me" into "step 3 of 9 is broken, here is the artefact."
How to connect your agent to the rest of your world. The agent is no good if it can't reach Notion, YouTube transcripts, or any of the third-party APIs that hold your data — this course is the 6-step pattern for wiring them up, the .env discipline that keeps the secrets out of chat, Notion as a second brain for everything the agent shouldn't re-derive from scratch, and the "save the working integration as a skill" loop that makes the next integration free.
The channel's *non-tutorial* track — Anthropic / OpenAI / Meta / xAI news, IPO coverage, leaked models, and the "Chinese labs are copying Claude" thread. These are the videos the channel posts when it has a *take* on the industry, not a how-to. This is the Stream 2 capstone: everything the model- and tool-coverage in the prior six courses points at is named out loud here.
The agent has your data, your API keys, your shell, and your contact list. The defaults are unsafe. This course is the channel's "don't get pwned" syllabus — five sub-articles that turn the most-quoted incidents on the channel (the 3am girlfriend message, the Facebook/Meta contact-list blast, the .env file in chat) into a concrete operating discipline.
Three vendors all claim a 60-second OpenClaw install on Windows. One of them is closer to 60 seconds than the others. This course is the legacy "1-click OpenClaw" path — the route you take when the recommended [Course 1.03 clean VPS path](01-picking-your-agent-harness.md) is not an option (Windows-only shop, no SSH fluency, zero patience for `wsl.conf`).