Source note (re-pulled 2026-06-17). All nine source videos now have summary_content, summary_key_takeaways, transcript_content, and summary_verdict populated in public.videos. Comments are available for c3bd0HiE3pg (27 comments), GL2FhteoPBA (7), and KKgEmpEh7zM (8). All concrete claims below are tied to either the source summary_content (via direct quote or paraphrase) or summary_key_takeaways. Dates and view counts come from published and views.

The §2.8 spine is the v0.8 → v0.16 release arc. The thematic names follow a deliberate progression: Intelligence → Everywhere → Tool Gateway → Interface → Foundation → Velocity → Surface (with two interstitials: v0.070 and the MCP server mode release at v0.060). Read in order, the names track a real engineering arc: make the model smart and stable (v0.8 Intelligence), distribute it everywhere (v0.9 Everywhere, 487 commits in 5 days), monetize the tool surface (v0.10 Tool Gateway, $10/month), replace the chat-app front-end with a real TUI (v0.11 Interface, 1,500 commits), plug in the data sources that matter (v0.14 Foundation, X via XAI OAuth), refactor the core for speed (v0.15 Velocity, 16k → 3.8k lines), ship the UI surfaces (v0.16 Surface, desktop + web admin). The MCP server mode release at v0.060 is the last piece of the arc — it flips the integration direction so other clients (Claude Code, Cursor, VS Code) can drive a Hermes agent.

Timeline at a glance

Date Version Thematic name Headline change Source
2026-03-31 v0.060 MCP Server Mode Hermes exposes itself as an MCP server; Yolo mode; official Docker image ZmbnZr0R8SU
2026-04-08 v0.070 (interim) Pluggable memory, credential pool, Camel Fox stealth browser HZD4by8OV1c
2026-04-09 v0.8 Intelligence notify on complete, live model failover, self-patched tool-calling dvv_rVxVj80
2026-04-14 v0.9 Everywhere 487 commits in 5 days; local web dashboard on port 9119; background monitor V39D46byKkc
2026-04-17 v0.10 Tool Gateway $10/month News Portal: Firecrawl search, image gen, TTS, browser automation VIpMz5uz4Cc
2026-04-24 v0.11 Interface 1,500 commits; React Ink TUI rewrite; AWS Bedrock transport; sub-agent spawning eZHO8L5GlAk
2026-05-18 v0.14 Foundation XAI OAuth for X browsing (Grok + Super Grok); native Windows; debloat command KKgEmpEh7zM
2026-05-29 v0.15 Velocity run_agent.py 16k→3.8k lines; session search 90s→20ms; kanban swarm; brainworm GL2FhteoPBA
2026-06-08 v0.16 Surface Desktop app; web admin system tab; trimmed default skills; undo command c3bd0HiE3pg

What you'll learn

  • The 5-day, 487-commit release cadence is the new normal, and the CLI's "commits behind" counter on boot is the signal to watch, not version announcements.
  • Each release has a thematic name and the names track a real arc from raw model power, to distribution, to monetization, to UX, to safety, to velocity.
  • Three of the biggest features — the $10/month tool gateway, the desktop app, and the X-via-XAI-OAuth integration — landed in separate releases but compound: together they replace roughly $200/month of third-party subscriptions.
  • The platform's stance on security hardening changed visibly: v0.15's "brainworm" mitigations, the Yolo-mode escape hatch, and the v0.16 undo command are all responses to the same underlying risk profile.
  • Migration is non-trivial. The MCP server mode release (v0.060) lost most of one creator's profiles on the dry run — Hermes upgrades are not turnkey.

Part F — Release-arc summary — the one-paragraph version

Across nine releases (v0.060 → v0.16, 2026-03-31 → 2026-06-08), Hermes has shipped:

  • One new external surface (v0.060 MCP server mode, exposing Hermes as a tool for other clients).
  • One new internal surface (v0.9 dashboard, the local web admin on port 9119, expanded in v0.16 with the system tab).
  • One monetization release (v0.10 News Portal Tool Gateway at $10/month, with Firecrawl, image gen, TTS, browser automation).
  • One UX rewrite (v0.11 React Ink TUI, 1,500 commits, the live status bar with context window consumption).
  • One integration release (v0.14 XAI OAuth for X browsing, replacing $19–$200/month in X API tooling).
  • One refactor release (v0.15 Velocity, run_agent.py 16k → 3.8k lines, session search 90s → 20ms).
  • One UI surface release (v0.16 Surface, the desktop app + system tab + trimmed default skills + undo command).
  • Two interstitials (v0.070 pluggable memory + credential pool + Camel Fox; v0.8 Intelligence, the model's smart-and-stable release).
  • One security escalation pattern (v0.060 Yolo mode → v0.15 brainworm mitigations → v0.16 undo command, all responses to the same expanding risk profile).

The release cadence is the new normal. The CLI's "commits behind" counter on boot is the signal to watch. The thematic names are not marketing — they are an engineering roadmap.

Part A — The model & memory arc (v0.060 → v0.8 → v0.070)

This is the part of the arc where the agent stops being a thin wrapper around a chat API. Three releases, two months, one conclusion: Hermes is an agent platform, not a chatbot.

v0.060 — MCP Server Mode (2026-03-31, ZmbnZr0R8SU). This release flips the MCP direction. Instead of Hermes calling out to MCP tools, other clients connect in to it. The creator says it "exposes Hermes model to cloud desktop, to cursor, to VS code," so you can drive a Hermes agent from Claude Code and see exactly what it's doing — including a world-master demo where Claude Code commands Hermes in-game (source: summary_content, ZmbnZr0R8SU). The same release also ships:

  • Profile isolation — multiple users share one orchestrator and one presentation-maker agent without collisions. On the migration test, "only three of my profiles survived after the dry run" (source: summary_content, ZmbnZr0R8SU).
  • Official Docker image — "a one-click sort of thing" rather than per-server setups (source: summary_content, ZmbnZr0R8SU).
  • Yolo mode — opt-out toggle to skip Hermes's hardened security prompts. Source: summary_key_takeaways ("Enable Yolo mode only on throwaway profiles; it disables the hardened security prompts and Hermes agents can touch a lot of files").

v0.8 — Intelligence Release (2026-04-09, dvv_rVxVj80). Three workflow changes that still shape how people use Hermes:

  • notify on complete flag on long-running processes — fire-and-forget instead of polling (source: summary_key_takeaways).
  • Live model switching with aggregator-aware failover — CLI, Telegram, and Discord all expose the picker. The router prefers OpenRouter and Nous Portal, then falls back automatically. When Opus 4.6 limits are hit, Hermes falls back to Minimax without manual intervention (source: summary_content).
  • Self-patched tool-calling failures — Hermes ran automated behavioral-match benchmarks against GPT and Codex, surfaced five failure patterns, and patched them inline, including execution-discipline guidance and thinking_only prefill for structured reasoning (source: summary_content).

Two smaller items that matter more than the headline: the browser provider migrated from browserbase to browseruse, and execute_code now runs on remote backends (Docker, SSH, Modal) instead of local-only (source: summary_content). The shared-thread-session default on Telegram/Discord gateways also lands here — no more per-user /start setup.

The notify on complete flag, in detail. v0.8's "fire-and-forget instead of polling" pattern. Before v0.8, a long-running process (e.g. a Kanban parent task that takes 30 minutes) required the user to either poll the Kanban UI or stay in the TUI and watch the progress. After v0.8, the user can launch the task with notify on complete and the agent sends a notification when the task completes. The notification lands in the dashboard's sessions tab; the user reads the result. The user is not polling; the user is notified. The same pattern is the basis for the §2.4 cron tab's log session surface — the cron job fires, the agent runs, the result is logged, the user reads the log.

The v0.8 self-patched tool-calling failures, in detail. The source video's framing: "Hermes ran automated behavioral-match benchmarks against GPT and Codex, surfaced five failure patterns, and patched them inline." The five failure patterns, paraphrased from the source:

  • Premature finalisation. The agent declared a task "done" before all child tasks completed. The patch: thinking_only prefill for structured reasoning — the agent must show its work before declaring done.
  • Tool call without parameter validation. The agent called a tool with invalid parameters (e.g. a path that didn't exist). The patch: execution-discipline guidance — the agent must validate parameters before calling tools.
  • Infinite retry on transient errors. The agent retried a 5xx error indefinitely. The patch: backoff and retry limits — the agent must respect backoff and stop after N retries.
  • Skipping a step in a multi-step workflow. The agent skipped a sub-task in a Kanban pipeline. The patch: explicit step enumeration — the agent must enumerate each step before executing.
  • Wrong tool for the job. The agent used curl when wget was the right tool, or python when node was the right tool. The patch: tool selection guidance — the agent must pick the right tool for the job, not the first tool that comes to mind.

The five patches are the v0.8 "Intelligence" framing made concrete. The release is not about the model getting smarter; it is about the agent getting more disciplined.

The browseruse migration. v0.8 migrated the browser provider from browserbase to browseruse. The two providers are similar in API; the migration was a low-friction swap. The reason for the migration: browseruse is more reliable on Mac (cross-reference §2.6 Computer Use — the macOS-only platform support story), and the channel's recommendation is to use browseruse for any Mac-based browser automation.

v0.070 — pluggable memory, credential pool, Camel Fox (2026-04-08, HZD4by8OV1c). Slotted between v0.8 and v0.9 chronologically (published the day before v0.8) and ships three features that address the "agent resurfaces stale context" failure mode:

  • Pluggable memory system — swap memory providers and flush stale state (source: summary_content, HZD4by8OV1c).
  • Credential pool — replaces the OpenClaw habit of storing passwords inside the skill file itself, which the host says "made it very easy to extract passwords" if you ever pushed to GitHub (source: summary_content).
  • Camel Fox — a browser that "pretends to be a human browsing," built to defeat anti-bot detection on Amazon, airline pricing, and competitor scrapes. Does not reliably work on X/Twitter but does help on flights and retail (source: summary_content).

This is also the release where the team publicly committed to migrating off OpenClaw — the deciding factor was Hermes' self-improving skills loop beating OpenClaw's media-creation tools for their internal workflow (source: summary_content).

The pluggable memory system, in detail. v0.070's "pluggable memory" means the memory provider is a swappable backend. The default memory provider is the SQLite + markdown vault that ships with Hermes (similar to the §2.1.3 OpenHuman memory tree). The user can swap in Honcho memory, or a custom memory provider, via the dashboard's memory tab or the ~/.hermes/config.yaml memory.provider field. The "flush stale state" operation is a one-line command (hermes memory flush) that clears the memory slot without affecting the skill library or the persona. The user-facing win: when the agent's context slot is full of stale memories, the user can flush without re-installing.

The credential pool, in detail. v0.070's "credential pool" is a separate file (default ~/.hermes/credentials.yaml or the equivalent on the current build) where API keys, OAuth tokens, and passwords live. The skill files reference the credential pool by name (e.g. credentials: youtube_api_key) instead of inlining the key. The host's framing: on OpenClaw, the host had the habit of storing passwords inside the skill file itself; if the user ever pushed a skill file to GitHub, the password was exposed. The credential pool fixes that — the skill file references the pool, the pool lives outside the dotfiles repo. The v0.16 system tab (§2.4.6) surfaces the credential pool as a first-class admin surface.

The Camel Fox browser, in detail. Camel Fox is a stealth browser that "pretends to be a human browsing." The use case: anti-bot detection on Amazon, airline pricing, and competitor scrapes. The browser spoofs user-agent strings, mouse movements, and timing patterns to defeat the most common anti-bot checks. The source video's honest caveat: does not reliably work on X/Twitter. The use case: Amazon product pricing scrapes, airline fare scrapes, competitor pricing scrapes. The use case not covered: anything that requires a logged-in session on a platform with strong anti-bot (X, Instagram, TikTok).

Part B — The distribution & monetization arc (v0.9 → v0.10)

Two releases, ten days. v0.9 ships the surface; v0.10 ships the revenue model.

v0.9 — Everywhere Release (2026-04-14, V39D46byKkc). 487 commits in 5 days — the biggest in project history (source: summary_content). The "Everywhere" name signals a push into China via WeChat and iMessage support. Two things from this release became permanent fixtures:

  • Local web dashboard on port 9119 — a browser UI to manage settings, sessions, skills, and the gateway without touching config files. By default it loads locally; do not expose it to the public, because the creators cite a 12 seconds SSH-attack window as realistic (source: summary_key_takeaways). This is the §2.4 dashboard.
  • Background process monitor — streams errors, port-listen messages, and terminal activity into chat, replacing OpenClaw's silent handling. This is the moment the channel publicly migrated from OpenClaw to Hermes for the dashboard plus the self-evolving skills loop (source: summary_content).

The release also added Hermes backup and Hermes import CLI commands for snapshotting config, sessions, skills, and memory, plus a plugins command for swapping in custom context-engine logic. The CLI now tells you how many commits you are behind on boot, and the channel's rule is to update from the CLI, not by asking the agent to self-update (source: summary_key_takeaways).

The 12-second SSH-attack window, in detail. The v0.9 release notes cite a 12 seconds SSH-attack window as realistic. The framing: a brute-force SSH attack on a public-facing port can succeed in roughly 12 seconds if the attacker has a known username + a weak password. The implication: a VPS with SSH on 0.0.0.0:22, a username of ubuntu, and a weak password is compromised within 12 seconds of going live. The fix: use Tailscale (§2.2.2), or use a non-standard SSH port, or use SSH key-only authentication. The §2.4 dashboard's port 9119 has the same issue — exposed on 0.0.0.0, the dashboard is brute-forced in seconds.

The "Everywhere" name — what it actually means. The v0.9 release is the "Everywhere" release because it pushed Hermes into China (WeChat, iMessage) and added the infrastructure for many messaging platforms. The release ships WeChat support, iMessage support, and 15 other messaging platform integrations. The 17th platform (QQ Bot) lands in v0.11. The "Everywhere" framing is not just marketing — the release is the moment Hermes becomes a multi-platform agent, not a Discord-only agent.

The background process monitor. v0.9's background process monitor streams errors, port-listen messages, and terminal activity into the chat surface. Before v0.9, OpenClaw handled errors silently; the user had to read the logs to find out what went wrong. After v0.9, the agent's chat surface shows the error in real time. The use case: a long-running task fails, the user sees the failure in the chat, the user can respond immediately. Cross-reference the §2.4 dashboard's log session tab — the same errors land in the dashboard's log session.

The Hermes backup and Hermes import commands. v0.9 shipped CLI commands for snapshotting and restoring the agent's state. The hermes backup command dumps the agent's config, sessions, skills, and memory to a tarball. The hermes import command restores from a tarball. The use case: a weekly cron for backups; a restore on a fresh VPS after a hardware failure. The archive guide 29-hermes-vps-setup.md recommends a weekly tarball of ~/.hermes/:

tar -czf hermes-backup.tar.gz ~/.hermes/

The hermes backup command is a wrapper around the same tarball pattern. The hermes import command is a wrapper around the inverse.

The plugins command. v0.9's plugins command lets the user swap in custom context-engine logic. The use case: a user has a custom context window strategy (e.g. a sliding-window summarisation scheme) that is not the default; the user writes a plugin that implements the strategy; the user registers the plugin via hermes plugins register <name>. The plugin runs as part of the context engine. The plugins command is the bridge between Hermes and the user's custom code.

v0.10 — Tool Gateway Release (2026-04-17, VIpMz5uz4Cc). This is the monetization release. The News Portal Tool Gateway unlocks four tools — web search (Firecrawl), image generation, TTS, and browser automation — for a single $10/month basic plan (source: summary_content). The host calls Firecrawl "insanely strong" (source: summary_content).

The key architectural choice is that you can subscribe to the gateway and still bring your own model key. The demo runs GLM 5.1 through a flat-rate coding plan — the $10 only buys the tools, the model traffic stays on your own plan. The host's rule: "Pay no more than $10" for the base (source: summary_content).

Two operational gotchas worth knowing (source: summary_key_takeaways):

  • Re-route bug — after enabling the tool gateway, the agent sometimes defaults back to News Portal's hosted models and errors out. Fix: re-select your provider in the UI, or edit ~/.hermes/config.yaml and flip the provider field away from newsresearch.
  • Core infrastructure cleanup — 180 commits in 2 days targeting "agent core infrastructure, CLI tooling, and tool reliability" to fix wrong-tool calls and the roughly every-6-to-9-hour behavior glitches. If your agents were misbehaving on that schedule, give v0.10.0 a day or two to settle.

The v0.10 Tool Gateway's tools — what each one does. The four tools unlocked by the News Portal Tool Gateway at the $10/month basic plan:

  • Web search (Firecrawl). The host calls Firecrawl "insanely strong" — it scrapes and structures web content, returning clean markdown instead of HTML. The agent invokes Firecrawl instead of curl | grep; the result is structured and ready to summarize.
  • Image generation. A hosted image generation API. The agent invokes it to produce images for the daily briefing, the thumbnail, the social post. The image gen is fast enough for a daily cron.
  • Text-to-speech (TTS). A hosted TTS API. The agent invokes it to produce audio summaries of the daily briefing, for users who prefer audio over text.
  • Browser automation. A hosted browser-automation API. The agent invokes it to drive a real browser for workflows that need a UI (e.g. filling out a form, navigating a SaaS dashboard that has no API). Cross-reference §2.6 Computer Use — the v0.10 browser automation is a hosted alternative to the on-Mac Computer Use path.

The four tools together replace a stack of paid third-party subscriptions (Brave Search, separate image-gen API, separate TTS API, separate browser-automation API). The $10/month is the cost-replacement story the §2.8.4 "What it means for cost (now)" framing references.

The "BYOK + Tool Gateway" pattern. The key architectural choice: you can subscribe to the $10/month tool gateway and still bring your own model key. The model traffic stays on your own plan; the $10 only buys the tools. The demo in the source video runs GLM 5.1 through a flat-rate coding plan — the $10 only buys the Firecrawl + image gen + TTS + browser automation; the GLM 5.1 traffic is on the user's coding plan. The host's rule: "Pay no more than $10" for the base.

The hosted-model default, the gotcha. When the user enables the tool gateway, the agent sometimes defaults back to News Portal's hosted models (the v0.10 release added a hosted-model layer for users who do not bring their own key). The re-route bug is the failure mode: the user has MiniMax M2.7 configured, but the tool gateway's hosted-model layer takes over and the agent errors out. The fix: re-select the user's provider in the UI, or edit ~/.hermes/config.yaml and flip the provider field away from newsresearch. The fix is one line; the bug is a re-routing issue, not a tool issue.

The standing safety rule that lands here. From VIpMz5uz4Cc, the v0.10 release notes spell out the safety ruling the channel keeps repeating: "Run Hermes on a VPS, not a local Mac mini, and don't grant it blanket system permissions — the creator frames this as the right risk trade-off given current agent security maturity." This is the §2.6 Computer Use house rule, stated as a v0.10 release principle.

Part C — The interface & integration arc (v0.11 → v0.14)

From 1,500-commit TUI rewrite to X-via-OAuth. This is where the agent becomes a daily-driver tool.

v0.11 — Interface Release (2026-04-24, eZHO8L5GlAk). 1,500 commits and a full React Ink TUI rewrite — roughly 300 of those commits went into the TUI alone (source: summary_content). The headline feature is hermes --tui, a React Ink rewrite of the CLI backed by a Python JSON-RPC gateway called TUI gateway.

What it actually does in practice (source: summary_content):

  • The live status bar shows git branch, current working directory, per-prompt stopwatch, and — most importantly — live context window consumption, which the Discord interface has never exposed.
  • Sub-agent spawns appear in the activity pane in real time. You can watch child agents branch off.
  • New slash commands include /help, /clear, /skills, and the standout /steer. Unlike /btw (side questions), /steer injects a directive mid-run that "arrives after next tool call" without breaking flow.

The model surface also expanded significantly (source: summary_content): AWS Bedrock on the Converse API (a real architectural move for AWS shops), plus five new inference providers (NVIDIA NIM, RCAI, Step Plan, Google Gemini CLI via OAuth, Vercel AI Gateway). The catalog adds Claude Opus 4.7, Kimi 2.6, Xiaomi MiMo V2.5 Pro, and GLM 5V Turbo. QQ Bot is the 17th messaging platform.

The operational warning (source: summary_key_takeaways): sub-agents can now spawn their own worker agents with file-based coordination (the "CEO → manager → worker" pattern). Set a max spawn depth immediately or the new orchestrator pattern can recursively collapse your gateway.

The "CEO → manager → worker" pattern, in detail. v0.11's sub-agent spawn chain is recursive: a CEO-level agent can spawn a manager-level agent, which can spawn a worker-level agent, which can spawn its own sub-agents. The chain is bounded by max spawn depth; without a cap, the chain can recursively collapse the gateway. The default max spawn depth is 3 (CEO → manager → worker); a 4-level chain is possible but not recommended. The 5-level chain (CEO → manager → worker → sub-worker → sub-sub-worker) is a footgun.

The TUI's /steer command. The v0.11 source video's standout slash command: /steer. Unlike /btw (side questions), /steer injects a directive mid-run that "arrives after next tool call" without breaking flow. The use case: the agent is in the middle of a long-running task, and the user wants to redirect. The user types /steer focus on X instead of Y; the agent's next tool call uses the new directive. The flow is not broken; the agent continues. The /btw command is for side questions (the agent answers the side question and then continues); the /steer command is for redirection (the agent's next tool call uses the new directive). The two coexist in the same release.

The TUI's live context window consumption. The v0.11 TUI's live status bar shows context window consumption — the user can see the context window fill up as the agent's response grows. Cross-reference the §2.4 long-context gotcha: bigger context windows are not always better, and the live context window consumption is the visual signal that the user has hit the wall. The v0.11 release's other load-bearing detail: the Discord interface has never exposed context window consumption; the TUI does.

The model surface expansion in v0.11. Five new inference providers joined in v0.11:

  • NVIDIA NIM. NVIDIA's inference microservices. The NIM path is the bridge between Hermes and the local NVIDIA box (cross-reference §2.9.4). The provider is the path for users who have an NVIDIA card and want to run Hermes locally.
  • RCAI. A research-lab-affiliated inference provider. The channel does not detail the provider; verify on the RCAI site.
  • Step Plan. A step-function-style inference provider. Useful for workflows that need step-by-step reasoning.
  • Google Gemini CLI via OAuth. The Gemini CLI is Google's command-line tool for Gemini models; OAuth integration means the user can authenticate with their Google account, no API key needed. Cross-reference the v0.14 XAI OAuth pattern.
  • Vercel AI Gateway. Vercel's hosted AI gateway. The Vercel path is the bridge between Hermes and the Vercel deployment platform.

The catalog also added four new models: Claude Opus 4.7, Kimi 2.6, Xiaomi MiMo V2.5 Pro, GLM 5V Turbo. QQ Bot is the 17th messaging platform (Discord, Telegram, Slack, Microsoft Teams, iMessage, WeChat, plus 11 others).

v0.14 — Foundation Release (2026-05-18, KKgEmpEh7zM). Dominated by one cost-replacing integration: XAI OAuth support lets your agent browse X directly through your account. No API keys, no tokens, no per-call fees. You log in once via OAuth and both regular Grok search and Super Grok come bundled with the plan you already pay for (source: summary_content).

The cost math (source: summary_content): Grok access requires X Premium Plus (note the plus), not the base Premium tier. Premium Plus runs about $30 a month. The host compares that favorably to "X APIs now can cost up to $200 per month or even the unofficial ones" at $19. Net: the v0.14 X integration replaces roughly $19–$200/month in tooling.

Smaller items in the same release (source: summary_content):

  • Native Windows support lands (the old WSL path was "not local").
  • Linux and Windows builds of the computer use feature are promised "in the next few weeks." Until then, "Mac is still probably the best way to use everything." — this is the §2.6 platform-support snapshot.
  • A new debloat command strips unused components like camel flock from your install.
  • Microsoft Teams gets end-to-end support.
  • One-hour cloud sessions are available for support escalations.

Comment-grounded note on v0.14. The audience read on the X integration: top-liked viewer @davidbayliss3789 (2 likes) is a UK-based viewer put off by X's politics, not by the integration; viewer @nomadicvelocities writes "Ok fine. You've convinced me. Time to move from Openclaw finally" — i.e. v0.14 was the prompt for a real OpenClaw-to-Hermes migration. Viewer @Jayander-n3k asks "Why do you use kimi and not deep seek?" — answered by @XxQuidanxX with a one-word "Context?" indicating the audience assumes the choice is context-window-driven. Source: public.youtube_comments for KKgEmpEh7zM, comments Ugwcw1D94f5jVmoanbZ4AaABAg, UgwIrNGch7CbVDviL8N4AaABAg, Ugw_hgaIYRiCoRX0wc94AaABAg.

The XAI OAuth integration, in detail. v0.14's headline cost-replacement integration. The setup:

  1. The user has an X Premium Plus subscription ($30/month) — the Premium Plus tier is required, not the base Premium tier.
  2. The user runs hermes config and selects the X integration.
  3. The user clicks "Connect X account" — the user is redirected to X's OAuth flow.
  4. The user authenticates with their X credentials.
  5. The user is redirected back to Hermes. The OAuth token is stored in the v0.070 credential pool.
  6. The user can now invoke the agent to browse X directly through the user's account. The agent's X browsing is rate-limited by X's terms, not by Hermes.

The cost replacement: the v0.14 X integration replaces $19–$200/month in X API tooling. The Premium Plus subscription is $30/month; the X API alternatives are $19/month (unofficial) to $200/month (official). The user saves money and gets a more reliable integration.

The x_search first-class tool. v0.14 also shipped x_search as a first-class Hermes tool. The agent invokes x_search to search X for a query; the tool returns structured results. The tool is built on top of the XAI OAuth integration — the user's X account is the authentication. The tool is rate-limited by X's terms; the user can monitor the rate limit in the dashboard's analytics tab.

The 808-commit, 633-PR v0.14 release. The v0.14 release was the largest in project history at the time: 808 commits, 633 merged PRs, 1,393 files changed, 165,061 insertions, 545 issues closed, 215 community contributors. The release shipped the XAI OAuth, the native Windows support, the Microsoft Teams end-to-end, the debloat command, and the x_search first-class tool. The release cadence at the v0.14 mark: roughly 1 major release every 7–10 days, with intermittent hot fixes (the v0.15 hot fix is the most-cited example).

Part D — The velocity & surface arc (v0.15 → v0.16)

Two releases, ten days. Refactor first, then put a UI on it.

v0.15 — Velocity Release (2026-05-29, GL2FhteoPBA). Headline refactor: run_agent.py dropped from 16,000 lines to 3,800 lines in roughly 8 days, redistributed across 14 focused modules. User-facing behavior stays compatible, so future updates are less likely to break your agent the way OpenClaw's used to (source: summary_content).

Three concrete speedups (source: summary_content):

  • hermes --version went from 700ms to 258ms.
  • Per-turn framework overhead fell from 399,000 function calls to 213,000 — a cut you can actually feel mid-conversation, with fewer hangs on simple tasks and less need for /goal.
  • Session search dropped from ~90 seconds to 20ms — about 4,500× faster — by removing the auxiliary LLM that previously powered it. Bonus: hermes sessions list now lets you resume a previous session and see the verbatim conversation.

The standout feature is hermes kanban swarm: a graph with a root task, parallel workers, a verifier, and a synthesizer. It ships with work-tree-per-task (so coding jobs don't collide in one directory), TTL claims, retry fingerprinting, and stale-task respawn guards (source: summary_content). The creator flags this as the update's standout feature and is doing a dedicated video on it (cross-referenced in §2.3). The kanban swarm is the §2.3 multi-board update's natural extension — the same parent-child retry loop (§2.3 anchor) generalised to a verifier lane.

Security also got a real pass. The release blocks the recent "brainworm"-class prompt injection at three choke points: tool output, recalled memory, and stored skills — plus 15 new threat patterns and a security guidance plugin (source: summary_content). Pin to v0.15 + the same-day hot fix together; the hot fix restores the 19,000-entry skills catalog, fixes the Docker dashboard reload loop, corrects kanban worker behavior, and re-enables MCP bear command resolution in Docker (source: summary_key_takeaways).

Comment-grounded note on v0.15. Top-liked viewer @luckyjc3 (2 likes) is pointed: "If I were on that team, I wouldn't be advertising that I wrote 16k lines in 1 file" — the audience's read on the refactor is "this was overdue." Viewer @Badmavs (0 likes) reports a real migration friction: "Release 13 broke few things for delegation and model switching, my agent had to patch the code. This time I am not upgrading but installing a brand new alongside." Viewer @timothy.everyday (1 like) reports the cost concern: "I use hermes for 3 days. Already burned through my allowance 😂 10$ gone because I wanted a finance dashboard." Viewer @enzopaupau2302 (1 like) is asking for the kanban swarm deep-dive explicitly — cross-confirms the "creator's standout feature" framing. Source: public.youtube_comments for GL2FhteoPBA, comments UgwQ87PgEfYz-jrAsAN4AaABAg, UgxYx8aou_FgPgTqYgl4AaABAg, UgzweEpl8P9VDuC0zfN4AaABAg, Ugzofzds2qqjk0DYeP14AaABAg.

The run_agent.py 16k → 3.8k refactor, in detail. The v0.15 refactor was a top-to-bottom rewrite of the agent's core orchestration code. The original run_agent.py had grown to 16,000 lines in a single file — a maintenance nightmare, the audience's "this was overdue" framing. The refactor redistributed the code across 14 focused modules:

  • hermes/orchestrator.py — top-level agent orchestration. The user-facing entry point.
  • hermes/profile.py — Kanban profile management. The §2.3 profile create / edit / delete operations.
  • hermes/kanban.py — Kanban board management. The board dispatch, parent-child retry, Worker Logs.
  • hermes/skill.py — Skill registry. The native skill set, the agent-created skills, the Curator's domain.
  • hermes/bundle.py — Skill Bundle primitive. The §2.5 YAML parsing, the /bundles create and /bundles reload commands.
  • hermes/curator.py — The Curator. The §2.7 background maintenance.
  • hermes/dashboard.py — The Dashboard. The §2.4 port 9119 server.
  • hermes/mcp_server.py — The MCP server mode. The §2.8 v0.060 server.
  • hermes/cli.py — The CLI entry point. The hermes command, the hermes --tui command.
  • hermes/memory.py — Memory management. The §2.8 v0.070 pluggable memory.
  • hermes/credentials.py — Credential pool. The §2.8 v0.070 secret store.
  • hermes/computer_use.py — Computer Use. The §2.6 vision-capable model + Hermes routing layer.
  • hermes/desktop.py — Desktop app. The §2.2.3 Electron + React shell.
  • hermes/web_admin.py — Web admin panel. The §2.4.6 v0.16 system tab.

The 14 modules are each focused on a single concern; the cross-module dependencies are explicit. The result: future updates are less likely to break the agent, the refactor is reviewable, and the velocity improvements (700ms → 258ms for hermes --version, 399k → 213k function calls per turn, 90s → 20ms for session search) are the user-visible effects.

The "brainworm" mitigation, in detail. v0.15's headline security escalation. "Brainworm" is the channel's name for a class of prompt-injection attack where a malicious payload in a tool output, a recalled memory, or a stored skill is interpreted as a user instruction. The v0.15 release blocks brainworm attacks at three choke points:

  • Tool output. The agent's tool output is sanitized before being added to the context window. The sanitization is pattern-based: known brainworm patterns (e.g. "ignore previous instructions", "you are now a helpful assistant") are stripped.
  • Recalled memory. The agent's recalled memory is sanitized before being added to the context window. The sanitization is the same pattern-based approach.
  • Stored skills. The agent's stored skills are sanitized before being invoked. The sanitization is the same pattern-based approach.

Plus 15 new threat patterns (e.g. encoded payloads, indirect injections via filenames, command-substitution attacks) and a security guidance plugin that warns the user when a high-risk pattern is detected. The v0.15 hot fix is the receipt that the brainworm mitigations are the load-bearing security escalation of the release; pin to v0.15 + the same-day hot fix together.

v0.16 — Surface Release (2026-06-08, c3bd0HiE3pg). About UI surfaces, not features. Three layers landed (source: summary_content):

  • Desktop app — drag-and-drop config, skills, and sessions without touching a terminal. The cost the channel flags: it "bleeds tokens on heavy UI schemas," so coding sessions should stay in the TUI. Community user "Tom in Tampa" confirms MCP tools are not discoverable in desktop sessions yet — the fix is an open pull request. A "thin client to your headless server" now lets a VPS-hosted agent connect back to the desktop UI. (Cross-reference §2.2.3 — the desktop-vs-TUI piece.)
  • Web admin panel — the dashboard's new system tab exposes the Hermes server, messaging channels, MCP catalogs, credentials, webhooks, and memory settings. Ron calls this "a very big deal" because routine admin leaves the config files and lives in the browser. (Cross-reference §2.4.6 — the v0.16 system tab.)
  • Skills trimmed, not curated — default built-in skills were pruned; niche and heavy skills are now optional installs, and a --no skills blank-slate install is available. The Skills Hub adds "Nvidia skills" as a trusted source alongside OpenAI, Anthropic, and Hugging Face — relevant for CUDA-X and NIM workflows.

Loop-level wins worth knowing (source: summary_content): a new undo command rolls back the last several user turns (faster than re-prompting), Read file token usage drops by about 14%, the model picker supports fuzzy search (V4 FL finds Deepseek V4 Flash), and the Hermes prompt-size diagnostic exposes context usage directly. Max spawn depth is "effectively uncapped" for deeper delegation trees — combine with the v0.15 brainworm mitigations, not against them.

Comment-grounded note on v0.16. Top-liked viewer @TheRicoRick (2 likes) confirms the desktop-app caveat from the channel: "Yeah I'm TUI camp the desktop is pretty pretty but I cant get as much work done personally." Viewer @VictorVovchenko-g5w reports a real bug — "Hermes desktop can't connect to vps?" — confirmed by @FelixKraut's detailed follow-up: "It can but can not use any of the GUI settings in the desktop app to change any setting. You can not even switch the AI model as the Desktop settings do not translate into your Hermes setting of the VPS. Tried it. Useless as of now if you ask me." So: the desktop's "thin client to your headless server" claim is partially true (it can connect) but the settings round-trip is broken as of the source recording. Viewer @chrish281 adds a real integration pattern: "using herdr with the tui on a local model works quite well once you set hermes up to orchestrate with herdr properly...obviously running locally you can't have agent swarms, but getting hermes to create monitored tasks in herdr really seems to work well." Source: public.youtube_comments for c3bd0HiE3pg, comments UgwwcGhFLdLoLH0Bq8R4AaABAg, UgzL1o69bTQul_Jb5WV4AaABAg, UgwCcAxfTe-J9SXSuWR4AaABAg, UgzXncSQ3x1biRpmDLt4AaABAg.

The v0.16 undo command, in detail. v0.16's user-facing loop-level win. The undo command rolls back the last several user turns — faster than re-prompting. The use case: the user asked a question, the agent gave a wrong answer, the user wants to rewind. Before v0.16, the user had to start a new session. After v0.16, the user types undo and the agent rewinds the conversation to the point before the wrong answer. The undo command is bound to the v0.15 brainworm mitigations: if the wrong answer triggered a brainworm attack, undo rolls back the context window to before the attack.

The v0.16 Read file token usage drop. v0.16 dropped Read file token usage by about 14%. The mechanism: the Read file tool's output format was optimised to use fewer tokens for the same content. The user-visible effect: 14% fewer tokens per Read file invocation. On a long coding session with many Read file invocations, the savings are real. The mechanism is transparent to the user — the agent's Read file tool just uses fewer tokens.

The v0.16 model picker fuzzy search. v0.16's model picker supports fuzzy search — typing V4 FL finds Deepseek V4 Flash. The user no longer has to type the full model name. The fuzzy search is a small quality-of-life improvement that compounds on long sessions: the user is switching models mid-session more often (cross-reference the §2.9 v0.8 hot-swap mechanic), and the fuzzy search reduces the friction of each switch.

The v0.16 prompt-size diagnostic. v0.16's prompt-size diagnostic exposes context usage directly. The user can see how much of the context window is consumed by the current prompt, and how much is left. The diagnostic is the visual signal for the long-context gotcha (cross-reference the §2.4 long-context rules): when the diagnostic shows the context window is full, the user knows to flush (/clear) or start a new session.

The v0.16 trimmed default skills. v0.16 pruned the default built-in skills. The niche and heavy skills are now optional installs, and a --no skills blank-slate install is available. The user-visible effect: the install is faster, the default skill library is smaller, the user opts in to the skills they want. The v0.15 hot fix restored the 19,000-entry skills catalog; v0.16 trims the default back down to the essentials.

The v0.16 "Nvidia skills" trusted source. The Skills Hub adds "Nvidia skills" as a trusted source alongside OpenAI, Anthropic, and Hugging Face. The Nvidia skills are CUDA-X and NIM workflows — relevant for users running Hermes on a local NVIDIA box (cross-reference §2.9.4). The trusted-source designation means the user can install Nvidia skills without the security warnings that third-party skill sources trigger.

Part E — The MCP server mode release (v0.060)

The v0.060 release is the last piece of the §2.8 arc, but it lands chronologically first (2026-03-31). The release flips the integration direction: instead of Hermes calling out to MCP tools, other clients connect in to it. That is the §2.8 anchor for the MCP story.

What MCP server mode actually does. Per ZmbnZr0R8SU summary_content: "exposes Hermes model to cloud desktop, to cursor, to VS code." The host says it in plain language elsewhere in the source: you can drive a Hermes agent from Claude Code, Cursor, or VS Code, and see exactly what the agent is doing. The world-master demo in the source video has Claude Code commanding Hermes in-game; the receipt for that demo is the MCP server mode transport.

The two architectural choices that matter.

  1. Profile isolation. Multiple users share one orchestrator and one presentation-maker agent without collisions. On the migration test, "only three of my profiles survived after the dry run" (ZmbnZr0R8SU summary_content). The profile isolation is per-request, not per-session; that is the part that needs care when multiple users are driving the same Hermes instance.
  2. Official Docker image. "A one-click sort of thing" rather than per-server setups. The Docker image is the path for anyone who wants MCP server mode without maintaining a custom install.

The Yolo mode escape hatch. Per ZmbnZr0R8SU summary_key_takeaways: "Enable Yolo mode only on throwaway profiles; it disables the hardened security prompts and Hermes agents can touch a lot of files." The Yolo mode toggle is the explicit acknowledgement that the hardened security prompts are not always what the user wants. For MCP server mode, the user is explicitly handing control of a Hermes agent to a client like Claude Code; some workflows (e.g. dev exploration) want the security prompts off. Yolo mode is the way to do that — but the channel's rule is to keep Yolo on throwaway profiles only.

The v0.060 migration cost. The source video's "only three of my profiles survived after the dry run" is a real first-party signal. The migration cost of v0.060 is not turnkey. The §2.8.4 "What it means for migration" framing applies: back up every profile, point Claude Code at the v0.060 docs, install dependencies first, then run the dry run. The Boxmining team lost most of their profiles in one pass; assume the same will happen to you.

The MCP server mode release's relationship to the rest of the §2.8 arc. v0.060 lands first chronologically (2026-03-31), before v0.070 (2026-04-08), v0.8 (2026-04-09), v0.9 (2026-04-14), v0.10 (2026-04-17), v0.11 (2026-04-24), v0.14 (2026-05-18), v0.15 (2026-05-29), and v0.16 (2026-06-08). The MCP server mode release is the "we can be driven" primitive; the v0.9 dashboard is the "we can be inspected" primitive; the v0.16 system tab is the "we can be administered" primitive. Together, the three layers are the platform's external surface — the rest of the arc is the internal surface (TUI, Kanban, Skill Bundles, Curator, Computer Use).

The audience's read on the v0.060 release. The source video has 5,130 views, which is high for an interim release. The audience's read is that the MCP server mode is the most consequential integration primitive in the arc — it turns Hermes from a standalone agent into a tool other clients can drive. For developers who already use Claude Code, Cursor, or VS Code, that is a real value proposition: keep your existing client, route a Hermes agent underneath it.

What this means

The v0.8 → v0.16 arc answers a question the platform was dodging in 2025: what is Hermes Agent, structurally? By v0.16, the answer is concrete: a TUI-first agent harness (v0.11), with a local web admin surface (v0.9, expanded v0.16), an optional desktop client (v0.16), a $10/month managed tool gateway (v0.10), native integrations to the data sources that matter (X via OAuth in v0.14), and a refactored core that runs ~4,500× faster on session search and ships inside an MCP server other clients can drive (v0.060 + v0.15). The thematic names are not marketing — they are an engineering roadmap you can read top-to-bottom.

What it means for installation choices (now). The v0.16 desktop app's settings-round-trip is broken for VPS-hosted agents (confirmed in public.youtube_comments, @FelixKraut on UgwCcAxfTe-J9SXSuWR4AaABAg). Stay on the TUI for any task where token cost matters, including coding. The desktop app is for project overviews and config browsing, not for daily drive.

The "tokens down, velocity up" arc — the v0.15 + v0.16 combined story. The v0.15 Velocity Release (16k → 3.8k lines) and the v0.16 Surface Release (desktop + system tab + undo command) together tell a story about the platform's direction:

  • v0.15 is the internal refactor — make the core faster, smaller, more reliable. The run_agent.py 16k → 3.8k lines refactor is the headline; the session search 90s → 20ms is the user-visible effect. The brainworm mitigations are the security response to the expanding risk profile.
  • v0.16 is the external surface — make the admin easy, make the desktop app usable, expose the v0.16 system tab. The undo command is the user-visible effect; the desktop app is the new surface. The trimmed default skills are the cost-saving response.

The two releases are 10 days apart (v0.15 on 2026-05-29, v0.16 on 2026-06-08). The cadence is the new normal. The CLI's "commits behind" counter on boot is the signal to watch.

The v0.15 → v0.16 install order matters. Pin to v0.15 + the same-day hot fix together. The hot fix restores the 19,000-entry skills catalog, fixes the Docker dashboard reload loop, corrects kanban worker behavior, and re-enables MCP bear command resolution in Docker. Running v0.15 alone will look broken. The v0.16 install on top of v0.15 is the right order.

What it means for security posture (now). Three independent signals escalate together: v0.060's Yolo mode (an opt-out escape hatch that disables the hardened prompts), v0.15's "brainworm" mitigations at three choke points plus 15 new threat patterns, and v0.16's undo command (recovery from a bad prompt run). The pattern: the surface area for the agent to do damage is expanding faster than the security mitigations, and the Yolo mode toggle is the explicit acknowledgement. Run Yolo on throwaway profiles only; pin v0.15 + its same-day hot fix together; verify your max spawn depth is capped after every v0.11+ upgrade.

What it means for cost (now). Three releases stack into a real cost-replacement story. v0.10's $10/month tool gateway replaces paid Brave Search + standalone browser-use + image-gen subscriptions. v0.14's XAI OAuth replaces $19–$200/month in X API tooling (assuming you already pay for X Premium Plus at $30/month). v0.16's trimmed default skills cut the install-time token spend on skills you would have ignored anyway. The headline cost question — "what does Hermes cost per month?" — now has a defensible answer: $10 tools + your existing model subscription + $30 for X (if you want it) + whatever you already pay for your VPS.

What it means for migration (now). Migration is the part the channel does not hide. The v0.060 dry run lost "only three of my profiles" for one creator; another viewer (@Badmavs on UgxYx8aou_FgPgTqYgl4AaABAg) reports skipping in-place upgrades after a bad v0.13 experience. The worked path is to point Claude Code at the target version's docs, install dependencies first, then run the dry run. Manual migration is the path that loses profiles.

What to watch next. The arc is not over. Three near-term follow-ups are already teased in the source videos: the kanban swarm deep-dive (referenced in v0.15 source, requested explicitly by viewer @enzopaupau2302 on Ugzofzds2qqjk0DYeP14AaABAg), the Linux/Windows computer-use rollout promised in v0.14 "in the next few weeks," and the desktop MCP discoverability fix (an open pull request, per the v0.16 source). If you are on the fence about installing, v0.16 is the moment the platform has all three surfaces (TUI, web admin, desktop) usable; the next release will likely close the remaining desktop-routing bugs.

Try it yourself

  1. Update the right way. Run hermes update from the CLI (not by asking the agent to self-update), then follow with hermes doctor afterwards (v0.9 source). v0.10+ ships 180-commit core infrastructure passes that take a day to settle; if your agent was misbehaving on a 6–9-hour schedule, give the patch a day before you roll back (v0.10 source).
  2. Configure live failover before you need it. Even if you only run one provider, set OpenRouter or Nous Portal as a secondary in the model picker. v0.8 ships aggregator-aware fallback that automatically switches to Minimax when Opus 4.6 limits are hit (v0.8 source).
  3. Move all credentials into the credential pool. Anything you would have stuffed in a skill file (API keys, passwords, OAuth tokens) belongs in the v0.070 credential pool. If you ever pushed a skill file to GitHub, rotate the keys now (v0.070 source).
  4. Migrate manually with a safety net. If you must move from an older Hermes version to v0.060+: back up every profile and dependency list, point Claude Code at the v0.060 docs, install dependencies first, then run the dry run. The Boxmining team lost most of their profiles in one pass — assume the same will happen to you (v0.060 source).
  5. Choose the surface per task. Coding → TUI (token-efficient, live context bar). Admin / handoff / team coordination → web admin panel. Project overviews and config editing → desktop app. The channel's rule of thumb: "Run the desktop app only for project overviews; stay in the TUI for any task where token cost matters" (v0.16 source; cross-confirmed by viewer @TheRicoRick on UgwwcGhFLdLoLH0Bq8R4AaABAg).

Common pitfalls

  • Self-updating the agent. Always run hermes update from the CLI (v0.9 source). Asking the agent to update itself is the most common path to a broken state; hermes doctor will only catch half the issues.
  • Exposing the v0.9 web dashboard publicly. Port 9119 is fine on localhost. On a remote box, the creators cite a 12-second SSH-attack window as realistic (v0.9 source). Tunnel it.
  • Leaving max spawn depth uncapped on v0.11+. The new "CEO → manager → worker" sub-agent pattern can recursively collapse your gateway. Set a depth limit immediately after upgrading (v0.11 source).
  • Trusting /fast mode by default. It routes to priority queues for roughly double the cost with the same intelligence. The hosts don't use it (v0.9 source).
  • Forgetting the v0.15 hot fix. Pin to v0.15 and the same-day hot fix together. The hot fix restores the 19,000-entry skills catalog and the Docker MCP bear command; running v0.15 alone will look broken (v0.15 source).
  • Paying for News Portal tools while routing models through it. Subscribe to the $10/month basic plan, then bring your own model key (GLM 5.1, Kimmy, MiniMax) and route through a flat-rate coding plan. The $10 should buy the tools, not the model traffic (v0.10 source).
  • Running Yolo mode on a production profile. It disables the hardened security prompts and Hermes agents can touch a lot of files. Keep Yolo on throwaway profiles only (v0.060 source).
  • Native Windows for computer use. The v0.14 build is a real native port, but the host still rates macOS as "probably the best way to use everything." Wait for the Linux/Windows computer-use rollout if you need desktop control off macOS (v0.14 source).
  • Picking Grok for agentic chains. The creator is direct: Grok is "fast" but "not your best servant." Reserve it for X-sourced news only — the X integration is the reason to enable the model, not its general capability (v0.14 source).
  • Migrating manually. Point Claude Code at the target version's docs and have it execute the move. The dependency-install step is the one that gets skipped, and that's exactly what wipes the profiles (v0.060 source).
  • Treating the v0.16 desktop app as a full settings client. Viewer @FelixKraut confirms settings do not round-trip to a VPS-hosted Hermes — the desktop can connect but the GUI settings do not apply. Stay in the TUI for any configuration change (UgwCcAxfTe-J9SXSuWR4AaABAg).

Sources (this article + aggregated sources for Course 2)

This is the §2.8 sources list, plus the aggregate-source manifest for the rest of the course. Every video referenced anywhere in Course 2 is listed here, in the order the channel released them, with the §2 subtopic that uses them.

Source videos (re-pulled 2026-06-17, all transcripts now populated):

Archive guides used in §2.2 / §2.4 / §2.7:

public.youtube_comments cited (33 total across 3 videos):

  • For KKgEmpEh7zM (8 comments): Ugwcw1D94f5jVmoanbZ4AaABAg (@davidbayliss3789, 2 likes), UgwIrNGch7CbVDviL8N4AaABAg (@nomadicvelocities), Ugw_hgaIYRiCoRX0wc94AaABAg (@Jayander-n3k, 1 like), Ugw_hgaIYRiCoRX0wc94AaABAg.AWxvSPz7WSnAWyb2Tqu37Y (@XxQuidanxX).
  • For GL2FhteoPBA (7 comments): UgwQ87PgEfYz-jrAsAN4AaABAg (@luckyjc3, 2 likes), UgxYx8aou_FgPgTqYgl4AaABAg (@Badmavs), UgzweEpl8P9VDuC0zfN4AaABAg (@timothy.everyday, 1 like), Ugzofzds2qqjk0DYeP14AaABAg (@enzopaupau2302, 1 like).
  • For c3bd0HiE3pg (27 comments; top-cited): UgwwcGhFLdLoLH0Bq8R4AaABAg (@TheRicoRick, 2 likes), UgzL1o69bTQul_Jb5WV4AaABAg (@VictorVovchenko-g5w, 1 like), UgwCcAxfTe-J9SXSuWR4AaABAg (@FelixKraut), UgzXncSQ3x1biRpmDLt4AaABAg (@chrish281).
  • For 86UIZVWkvF8 (Mavis): Ugxk6RGLqVm9oyhIv314AaABAg (@RyanDavisEdwardjr, 1 like), UgwN8DmX-YplvspuQ1p4AaABAg (@realme72only), UgzbtfQWLoce8-eyFdd4AaABAg (@BoxminingAI, pinned).
  • For bSpjLglSh34 (OpenHuman): UgwG26PDHwOur8fdEtF4AaABAg (@phantombrainm, 6 likes), UgxMLZl0OWmBGdwiwZF4AaABAg (@badr_mo, 2 likes), UgyVpz9umiFYf7hVZBR4AaABAg (@mikejuba7632), UgwRkL_xOePsmeZtYmt4AaABAg (@socal-pilot), UgzLrP3vwuJ0goWj6Fx4AaABAg (@BenoitDion-x4o), Ugyd8ha9GYvQA7IuF4d4AaABAg (@stevesmith4600).
  • For _A02brv2Csg: Ugxg2UXh4oAu6BEJckZ4AaABAg (@AdairCasey, ordering question).
  • For TCL0emWOjOQ: Ugxjewosvkmii5NxYDt4AaABAg (@zombopanda, 4 likes), Ugxjewosvkmii5NxYDt4AaABAg.AXa-aKgp8svAXa3y20WlzW (@AdamDymitruk, 1 like), Ugypvs7zFvweNvOeQit4AaABAg (@Tim-A-110, 1 like), UgxiNtm_TUe2KDgxwRJ4AaABAg (@SlasherSeven).
  • For t7rI2IBOvqI (/goal): Ugw2DG6m-2ex_9R_GGB4AaABAg (@angelmechanical, plan-vs-goal), UgxG__bOIpxMZfvJdep4AaABAg (@ashrule1, token cost), UgxG__bOIpxMZfvJdep4AaABAg.AWfWxeNScIvAWftlSClDRT (@jonpattrson), UgxG__bOIpxMZfvJdep4AaABAg.AWfWxeNScIvAWgIqwr8zh4 (@stuartpatterson1617), UgxF_POP7CoaEAwjwMh4AaABAg (@badrnaciri4504, Linux/WSL permissions), UgyuKQDUPvn9HHygxMl4AaABAg (@felixen21, broken affiliate link).
  • For Hh1sDJZ6VhY (Computer Use): 10 first-party viewer comments cited in §2.6 — UgxhVK9DQohPg32dqkB4AaABAg.AWhw8D1BOpUAWi1LglUE95 (@phinehasblack8739, channel reply, 2 likes, "use a smart model"), UgxhVK9DQohPg32dqkB4AaABAg.AWhw8D1BOpUAWiycAeCE9m (@TreBros, "brain and hands" framing), UgzL6afuDM-9LDT62mB4AaABAg (@allansolomon5132, "very deep" warning, 1 like), UgzKRYejRcNCpJUqVJ54AaABAg (@kaotictube, "virus" reaction), UgxhVK9DQohPg32dqkB4AaABAg (@MrSpice1971, Adobe Premiere Pro question).

Cross-references to other course articles:

  • §2.2 — First install (VPS, Mac, Desktop App) — for the v0.16 desktop app, the v0.16 "system tab" round-trip bug, and the token-cost trade-off flagged in v0.16.
  • §2.3 — The Kanban — for the kanban swarm pattern introduced in v0.15 and the parent-child retry loop.
  • §2.4 — The Dashboard — for the v0.9 port-9119 local web dashboard and the v0.16 admin panel / system tab.
  • §2.6 — Computer Use — for the macOS-first computer use feature whose Linux/Windows rollout was promised in v0.14.
  • §2.7 — The Curator — for the v0.12.0 "The Curator" release that named the feature and shipped autonomous skill maintenance.

Supabase query used (re-pulled 2026-06-17):

SELECT
  video_id, title, views, published,
  summary_content,
  summary_key_takeaways,
  transcript_content,
  action_intel,
  summary_verdict
FROM public.videos
WHERE video_id = ANY(ARRAY[
  '86UIZVWkvF8', '2NbfOOD2i1E', 'bSpjLglSh34',
  '5F1hFI2lZCg', 'UbK2kXygPUY', 'QSANg6VHkXI', 'nhDA7tcQtx0',
  'R_aLVXYzDac', 'iN2fD36Sgdg', 'fKoPRL0dhyk',
  'GfQEdMZ9LlA',
  '_A02brv2Csg', 'TCL0emWOjOQ', 't7rI2IBOvqI',
  'Hh1sDJZ6VhY',
  'SpFgS7WlCJc',
  'dvv_rVxVj80', 'HZD4by8OV1c', 'V39D46byKkc',
  'VIpMz5uz4Cc', 'eZHO8L5GlAk', 'KKgEmpEh7zM',
  'GL2FhteoPBA', 'c3bd0HiE3pg', 'ZmbnZr0R8SU',
  'Af7Fg1m7hRw', 'Nqs_5RLg6QA', 's3Q9hvdlrmo', 'mWJiMAN0DWk'
]);

SELECT video_id, comment_id, author_name, like_count, published_at, text_original
FROM public.youtube_comments
WHERE video_id IN (
  'c3bd0HiE3pg', 'GL2FhteoPBA', 'KKgEmpEh7zM',
  '86UIZVWkvF8', 'bSpjLglSh34',
  '_A02brv2Csg', 'TCL0emWOjOQ', 't7rI2IBOvqI',
  'Hh1sDJZ6VhY'
)
ORDER BY video_id, like_count DESC;