Hermes Agent Skills Catalog: 118 Skills Mapped — Local vs Bundled

· 5 min read ai productivity

TL;DR: Hermes Agent v0.8.0 ships with 77 bundled skills from the Skills Hub and I’ve accumulated 41 local skills through daily use — totaling 118 skills across 14 categories. Here’s the full map of every skill, what it does, and which ones came from where.

Hermes Agent has a skills system that works like procedural memory — reusable instructions the agent loads when a task matches a trigger. Skills come from two sources: the Skills Hub (bundled with the agent, auto-updatable) and local creation (written by the user or self-created by the agent during use). I recently audited my full ~/.hermes/skills/ directory to understand what I actually have installed. Here’s the breakdown.

How Skills Work in Hermes

Every skill is a markdown file (SKILL.md) with YAML frontmatter containing a name, description, and optional triggers. When Hermes encounters a task that matches a skill’s trigger or description, it loads the skill’s instructions into context. Skills live in ~/.hermes/skills/<category>/<skill-name>/.

Bundled skills are tracked in ~/.hermes/skills/.bundled_manifest with content hashes — they update via hermes skills sync. Local skills are invisible to the Hub and never get overwritten. Both types coexist without conflict.

Bundled Skills (77 from Skills Hub)

These ship with Hermes Agent and are maintained by Nous Research and the community.

Apple / macOS

SkillDescription
apple-notesManage Apple Notes via the memo CLI — create, view, search, and edit notes. macOS only.
apple-remindersManage Apple Reminders via remindctl CLI — list, add, complete, and delete reminders. macOS only.
findmyTrack Apple devices and AirTags via FindMy.app using AppleScript and screen capture. macOS only.
imessageSend and receive iMessages/SMS via the imsg CLI. macOS only.

Autonomous AI Agents

SkillDescription
claude-codeDelegate coding tasks to Claude Code (Anthropic’s CLI agent) — features, refactoring, PR reviews, iterative coding. Requires claude CLI.
codexDelegate coding tasks to OpenAI Codex CLI — features, refactoring, batch issue fixing. Requires codex CLI + git repo.
hermes-agentComplete guide to using and extending Hermes itself — CLI, setup, config, spawning agents, gateway platforms, skills, voice, tools, profiles, and contributing.
opencodeDelegate tasks to OpenCode CLI for feature implementation, refactoring, and long-running autonomous sessions. Requires opencode CLI.

Creative

SkillDescription
ascii-artGenerate ASCII art using pyfiglet (571 fonts), cowsay, boxes, toilet, image-to-ascii, and remote APIs. No API keys needed.
ascii-videoConvert video, audio, or images into colored ASCII character video output (MP4, GIF, image sequence). Includes audio-reactive visualizers and text overlays.
excalidrawCreate hand-drawn style diagrams as .excalidraw files — architecture diagrams, flowcharts, sequence diagrams, concept maps. Open at excalidraw.com.
manim-videoCreate mathematical and technical animations using Manim Community Edition — 3Blue1Brown-style explainers, algorithm visualizations, equation derivations.
p5jsCreate interactive and generative visual art using p5.js — browser-based sketches, data visualizations, WebGL scenes, audio-reactive visuals. Export as HTML, PNG, GIF, MP4, or SVG.
popular-web-designs54 production-quality design systems extracted from real websites (Stripe, Linear, etc.) — generates HTML/CSS matching their visual identity.
songwriting-and-ai-musicSongwriting craft, AI music generation prompts (Suno focus), parody/adaptation techniques, and phonetic tricks.

Data Science

SkillDescription
jupyter-live-kernelUse a live Jupyter kernel for stateful, iterative Python execution via hamelnb. For exploration, iteration, and inspecting data in-progress.

Gaming

SkillDescription
minecraft-modpack-serverSet up a modded Minecraft server from CurseForge/Modrinth packs — NeoForge/Forge install, Java version, JVM tuning, firewall, LAN config, backups.
pokemon-playerPlay Pokemon games autonomously via headless emulation — reads game state from RAM, makes strategic decisions, sends button inputs from the terminal.

GitHub

SkillDescription
codebase-inspectionAnalyze codebases with pygount — LOC counting, language breakdown, code-vs-comment ratios.
github-authSet up GitHub authentication via git or gh CLI — HTTPS tokens, SSH keys, credential helpers. Auto-detects the right method.
github-code-reviewReview code changes by analyzing git diffs, leaving inline comments on PRs, and performing pre-push review. Uses gh CLI or REST API fallback.
github-issuesCreate, manage, triage, and close GitHub issues — search, label, assign, link to PRs.
github-pr-workflowFull PR lifecycle — create branches, commit, open PRs, monitor CI, auto-fix failures, merge.
github-repo-managementClone, create, fork, configure, and manage repos — remotes, secrets, releases, workflows.

MCP (Model Context Protocol)

SkillDescription
mcporterUse the mcporter CLI to list, configure, auth, and call MCP servers/tools — HTTP or stdio transports, ad-hoc servers, config edits.
native-mcpBuilt-in MCP client that connects to external MCP servers, discovers tools, and registers them as native Hermes tools. Supports stdio and HTTP.

Media

SkillDescription
gif-searchSearch and download GIFs from Tenor using curl. No dependencies beyond curl and jq.
heartmulaSet up and run HeartMuLa, an open-source Suno-like music generation model — full songs from lyrics + tags with multilingual support.
songseeGenerate spectrograms and audio feature visualizations (mel, chroma, MFCC, tempogram) from audio files via CLI.
youtube-contentFetch YouTube video transcripts and transform them into structured content — chapters, summaries, threads, blog posts.

MLOps — Inference

SkillDescription
ggufGGUF format and llama.cpp quantization for CPU/GPU inference — flexible 2–8 bit quantization without GPU requirements.
guidanceControl LLM output with regex and grammars — guarantee valid JSON/XML/code generation via Microsoft Research’s constrained generation framework.
llama-cppRun LLM inference on CPU, Apple Silicon, and consumer GPUs without NVIDIA. GGUF quantization (1.5–8 bit), 4–10× speedup vs PyTorch on CPU.
obliteratusRemove refusal behaviors from open-weight LLMs using mechanistic interpretability — diff-in-means, SVD, LEACE, SAE decomposition. 9 CLI methods, 28 analysis modules.
outlinesGuarantee valid JSON/XML/code structure during generation with Pydantic models for type-safe outputs. By dottxt.ai.
vllmServe LLMs with high throughput using PagedAttention and continuous batching — OpenAI-compatible endpoints, GPTQ/AWQ/FP8 quantization, tensor parallelism.

MLOps — Models

SkillDescription
audiocraftAudio generation with MusicGen (text-to-music) and AudioGen (text-to-sound) via PyTorch.
clipOpenAI’s vision-language model — zero-shot image classification, image-text matching, cross-modal retrieval. 400M image-text pairs.
segment-anythingFoundation model for image segmentation — zero-shot transfer, point/box/mask prompts, automatic mask generation.
stable-diffusionText-to-image, image-to-image, inpainting via HuggingFace Diffusers. Build custom diffusion pipelines.
whisperGeneral-purpose speech recognition — 99 languages, transcription, translation, language identification. Six model sizes (39M–1550M params).

MLOps — Training & Evaluation

SkillDescription
axolotlFine-tune LLMs with Axolotl — YAML configs, 100+ models, LoRA/QLoRA, DPO/KTO/ORPO/GRPO, multimodal support.
grpo-rl-trainingGRPO/RL fine-tuning with TRL for reasoning and task-specific model training.
peftParameter-efficient fine-tuning — LoRA, QLoRA, 25+ methods. Train <1% of parameters with minimal accuracy loss for 7B–70B models.
pytorch-fsdpFully Sharded Data Parallel training — parameter sharding, mixed precision, CPU offloading, FSDP2.
trl-fine-tuningFine-tune with RL using TRL — SFT, DPO, PPO/GRPO, reward model training. Works with HuggingFace Transformers.
unslothFast fine-tuning — 2–5× faster training, 50–80% less memory, LoRA/QLoRA optimization.
lm-evaluation-harnessEvaluate LLMs across 60+ academic benchmarks (MMLU, HumanEval, GSM8K, TruthfulQA). Industry standard.
weights-and-biasesTrack ML experiments, visualize training, optimize hyperparameters with sweeps, manage model registry.

MLOps — Platforms & Research

SkillDescription
modalServerless GPU cloud — on-demand GPU access, ML model APIs, batch jobs with auto-scaling.
huggingface-hubHugging Face Hub CLI — search, download, upload models and datasets, manage repos, deploy inference endpoints.
dspyBuild complex AI systems with declarative programming, auto-optimize prompts, modular RAG via Stanford NLP’s framework.

Productivity

SkillDescription
google-workspaceGmail, Calendar, Drive, Contacts, Sheets, Docs integration via gws CLI with OAuth2 token refresh.
linearManage Linear issues, projects, teams via GraphQL API — create, update, search, organize. API key auth, all via curl.
nano-pdfEdit PDFs with natural-language instructions via nano-pdf CLI — modify text, fix typos, update titles per page.
notionCreate and manage Notion pages, databases, blocks via curl — search, create, update, query workspaces.
ocr-and-documentsExtract text from PDFs and scanned documents — web_extract for URLs, pymupdf for text PDFs, marker-pdf for OCR docs.

Research

SkillDescription
arxivSearch and retrieve academic papers from arXiv REST API — keyword, author, category, or ID search. No API key needed.
blogwatcherMonitor blogs and RSS/Atom feeds via blogwatcher-cli — add blogs, scan for articles, track read status, filter by category.
llm-wikiBuild a persistent, interlinked markdown knowledge base based on Karpathy’s LLM Wiki — ingest sources, query compiled knowledge, lint for consistency.
polymarketQuery Polymarket prediction markets — search markets, prices, orderbooks, price history. Read-only, no API key needed.
research-paper-writingEnd-to-end ML/AI research paper pipeline — experiment design, analysis, drafting, revision, submission. Covers NeurIPS, ICML, ICLR, ACL, AAAI, COLM.

Other Bundled

SkillDescription
dogfoodSystematic exploratory QA testing of web applications — find bugs, capture evidence, generate structured reports.
domain (domain-intel)Passive domain reconnaissance — subdomain discovery, SSL cert inspection, WHOIS, DNS records, domain availability. No API keys.
himalayaManage email via IMAP/SMTP CLI — list, read, write, reply, forward, search, organize. Multiple accounts, MML support.
find-nearbyFind nearby places (restaurants, cafes, pharmacies) via OpenStreetMap. Works with coordinates, addresses, cities. No API keys.
obsidianRead, search, and create notes in the Obsidian vault.
openhueControl Philips Hue lights, rooms, scenes via OpenHue CLI — on/off, brightness, color, color temperature.
powerpointFull PowerPoint .pptx handling — create, read, parse, extract, edit, modify, combine, split decks. Templates, layouts, speaker notes.
xitterInteract with X/Twitter via x-cli — post, read timelines, search, like, retweet, bookmarks, mentions.
webhook-subscriptionsCreate and manage webhook subscriptions for event-driven agent activation.
writing-plansCreate implementation plans with bite-sized tasks, exact file paths, and complete code examples for multi-step specs.
planPlan mode — inspect context, write markdown plans to .hermes/plans/, do not execute.
requesting-code-reviewPre-commit verification — static security scan, baseline-aware quality gates, reviewer subagent, auto-fix loop.
subagent-driven-developmentExecute implementation plans with independent tasks — dispatches fresh delegate_task per task with two-stage review.
systematic-debugging4-phase root cause investigation — NO fixes without understanding the problem first.
test-driven-developmentRED-GREEN-REFACTOR cycle enforcement — test-first approach for any feature or bugfix.
godmodeRed-team LLM safety filters — Parseltongue obfuscation (33 techniques), system prompt templates, multi-model racing, encoding escalation.

Local Skills (41 created by user or agent)

These were created during my daily work — they won’t be overwritten by Hub sync and reflect my personal workflow, infrastructure, and project-specific patterns.

DevOps & Infrastructure

SkillDescription
redis-env-configUse individual REDIS_HOST/PORT/PASS/DB env vars instead of REDIS_URL connection strings — handles special characters in passwords.
saml-bridge-setupSet up SimpleSAMLphp 2.x SAML2 IdP bridging to ORY Hydra via OIDC in Docker.
simplesamlphp-nginx-dockerSimpleSAMLphp behind nginx + php-fpm in Docker — configuration pitfalls and production patterns.
ssphydra-saml-bridgeBuild a SimpleSAMLphp 2.x SAML IdP that delegates authentication to ORY Hydra via OIDC auth source module.
xpass-saml-bridgeSimpleSAMLphp SAML2-to-OIDC bridge for integrating SAML2 SPs (GlobalProtect) with ORY Hydra.
cctv-analysis-pipelineBuild a multi-stage CCTV surveillance video analysis pipeline — OpenCV motion detection, YOLO object detection, OCR, and VLM verification.

These reflect my day job building SAML/OIDC identity bridges with ORY Hydra and SimpleSAMLphp.

MLOps — Local Hardware

SkillDescription
koboldcppRun KoboldCPP on Arch Linux with AMD Vulkan GPU offload for GGUF models.
llama-cpp-vulkan-amdBuild llama.cpp with Vulkan backend on AMD iGPU/dGPU via Mesa RADV — no CUDA or ROCm needed.
llama-cpp-vulkanBuild and run llama.cpp with Vulkan GPU offload on AMD integrated GPUs (broader version, no ROCm).
stable-diffusion-cppRun diffusion models (SD, FLUX, Z-Image, Wan) locally via stable-diffusion.cpp — the ggml-based image generation equivalent of llama.cpp.
small-vlm-comparisonCompare and evaluate small VLMs for local deployment — RAM/VRAM requirements, quantization, video support.

These exist because I’m running on AMD hardware (Vega iGPU, 512MB VRAM) — no NVIDIA, no CUDA. Every inference skill here was written from painful first-hand experience.

MLOps — Knowledge

SkillDescription
cloudGPU cloud providers and serverless compute platforms for ML workloads.
evaluationModel evaluation benchmarks, experiment tracking, tokenizers, interpretability tools.
inferenceModel serving, quantization (GGUF/GPTQ), structured output, inference optimization.
modelsSpecific model architectures — CLIP, SAM, SD, Whisper, AudioCraft, multimodal models.
researchML research frameworks for declarative programming (DSPy, etc.).
trainingFine-tuning, RLHF/DPO/GRPO, distributed training, optimization tools.
vector-databasesVector similarity search and embedding databases for RAG and semantic search.

These are knowledge-base directories (DESCRIPTION.md only, no executable skill) that Hermes accumulated while learning about the MLOps landscape.

Software Development

SkillDescription
agent-browserHeadless browser automation CLI optimized for AI agents — accessibility tree snapshots and ref-based element selection. Uses the agent-browser npm package.
astro-blogBuild Astro 6 static blogs with Tailwind v4, DaisyUI 5, expressive-code, mermaid, callouts, Tabler icons. Covers scaffolding, component architecture, pitfalls.
astro-island-refactoringRefactor Astro projects with monolithic client:load Vue islands into proper Astro island architecture with Alpine.js for cross-component state.
astro-island-vueIntegrate Vue Router client-side routing inside Astro SSR apps — share one router across Astro Islands with appEntrypoint, catch-all pages.
astro-ssr-dashboardBuild secure Vue dashboards with Astro SSR, Bun, IDDS, BetterAuth, Zod, Tailwind CSS 4. No Vue Router — Astro pages ARE the routes.
astro-wiki-from-ai-suka-2Convert this blog (ai-suka-2) into a wiki-style documentation site with category-based content structure.
auto-cotAuto-CoT reasoning — decompose prompts into diverse sub-problems, reason through each, assemble a structured plan before executing. Based on Zhang et al. (2022).
bdBeads (bd) dependency-aware issue tracker with Dolt backend — issue tracking, task dependencies, project coordination.
bun-ssg-blogBuild a static blog with Bun + unified (remark/rehype) pipeline — content loading, Tailwind v4 CSS build, gotchas.
ctx7Context7 CLI (ctx7) for looking up library/framework/API documentation during development.
go-fiber-monolithGo backend architecture — modular monolith, DDD, Fiber v3, web best practices, folder structure, handler patterns, testing.
hermes-memory-providerBuild a Hermes Agent memory provider plugin wrapping an external CLI tool (lucid, brv, etc.) for persistent cross-session memory.
iddsINA Digital Design System (IDDS) — component library usage for Vue and plain JS.
jekyll-to-astro-migrationMigrate blog posts from Jekyll (Minimal Mistakes) to Astro with content collections.
nim-llmCall NVIDIA NIM API for LLM tasks — reasoning, code, vision, multilingual. Free tier with OpenAI-compatible endpoint.

The largest category — these skills were created while building this blog, working with Astro, learning Go Fiber, and integrating various tools into my workflow.

Productivity

SkillDescription
beads-lucid-flowBeads + Lucid integration workflow — structured tasks meet searchable memory. Bridges the issue tracker (Beads) with the long-term memory system (Lucid).

Autonomous AI Agents

SkillDescription
pi-agentPi coding agent CLI (pi.dev / shittycodingagent.ai) — minimal terminal coding harness by badlogic. Provider setup, config, extensions, skills, testing harness. npm: @mariozechner/pi-coding-agent.
qwen-codeDelegate tasks to Qwen Code CLI via ACP (Agent Communication Protocol) or subprocess mode.

Dogfood

SkillDescription
memory-strategyMemory and context management strategy for Hermes — tool routing, storage hierarchy, and context hygiene.

Creative

SkillDescription
svg-hero-imageCreate hand-crafted SVG hero images for blog posts — dark-themed technical illustrations with terminal content, diagrams, and animated elements.

Media

SkillDescription
cf-workers-image-genGenerate and transform images using Cloudflare Workers AI (free tier) — img2img, negative prompts, inpainting via SD 1.5, FLUX text-to-image.
hf-image-genGenerate images using HuggingFace Inference API (free tier) — FLUX.1-schnell by default, no local GPU needed.

Research

SkillDescription
tvlyTavily CLI (tvly) for web search, extraction, crawling, mapping, and deep research. Use instead of calling Tavily API directly.

Stats at a Glance

SourceCountPercentage
Bundled (Skills Hub)7765%
Local (user/agent created)4135%
Total118100%
CategoryBundledLocalTotal
Software Development71522
MLOps141125
Creative718
Productivity516
GitHub606
Research516
DevOps156
Autonomous AI Agents426
Media426
Gaming202
Apple / macOS404
MCP202
Other10010

The distribution tells a clear story: most bundled skills cover general-purpose tasks (GitHub, email, presentations), while my local skills concentrate on the tools and frameworks I use daily — Astro, Go Fiber, AMD Vulkan inference, SAML/OIDC bridges, and CLI integrations like Tavily and Context7.

How to Audit Your Own Skills

If you want to do the same for your Hermes installation:

Terminal window
# List all local skills (on disk)
find ~/.hermes/skills -name 'SKILL.md' | sed 's|/SKILL.md||;s|.*/skills/||'
# List all bundled skills (from Hub manifest)
cut -d: -f1 ~/.hermes/skills/.bundled_manifest
# Find skills that exist locally but NOT in the manifest (your custom ones)
comm -23 <(find ~/.hermes/skills -name 'SKILL.md' | sed 's|/SKILL.md||;s|.*/skills/||' | sort) \
<(cut -d: -f1 ~/.hermes/skills/.bundled_manifest | sort)

References

  1. Hermes Agent Documentation — Nous Research — https://hermes-agent.nousresearch.com/docs/
  2. Hermes Agent GitHub — Nous Research — https://github.com/NousResearch/hermes-agent
  3. Hermes Agent Skills System — Nous Research — https://hermes-agent.nousresearch.com/docs/user-guide/features/skills

This article was written by Pi (GLM-5 Turbo | Z.AI Coding Plan).