Glossary: the words of this site, in plain language

Level ReadingEstimated time 10 minA word you don't know? Glossary

Every technical word used on this site, in alphabetical order, explained without jargon. If a tutorial loses you on a word, it is probably here.

Agent: A program driven by an AI model that does not only answer but acts: it reads files, runs commands, opens pull requests. See Giving code work to a local agent.
API: A door through which one program asks another for something, with a fixed format. "OpenAI-compatible API" means the format most AI tools already speak.
Backup: A copy you can restore. A copy you have never restored is not yet a backup. See Homelab 5.
Certificate (HTTPS): The digital ID card that lets a browser show the padlock. Let's Encrypt gives them for free. See Homelab 2.
CI (continuous integration): A machine that runs the tests automatically on every proposed change, and says green or red. See Homelab 8.
Compose (Docker Compose): A small text file that describes one or several containers and how to start them, so that one command brings everything up.
Container: A program packed with everything it needs, isolated from the rest of the machine. You start it, stop it, delete it without leaving traces. Docker is the usual tool.
Context: How much text a model can keep in mind at once, counted in tokens. More context needs more VRAM.
CSP (Content Security Policy): A header that tells the browser which scripts, styles and images a page may load. The best single defence against injected scripts. See the audit.
DDoS: Distributed denial of service: many machines flooding one site until it cannot answer. Small floods can be filtered at home; big ones only upstream. See resisting DDoS.
DNS: The internet's phone book: it turns a name like freedzhost.cloud into an address. See Homelab 4.
DNSSEC: Signatures on DNS answers, so a resolver can tell a real answer from a forged one.
GGUF: The file format of models for llama.cpp. One file holds the whole model, already compressed.
GPU / graphics card: The card that computes images, and AI models, very fast. Its own memory is the VRAM.
Homelab: Your own small data centre at home: one or a few machines running the services you would otherwise rent.
HSTS: A header that tells the browser to always use HTTPS for a site, even if someone types http://.
Hypervisor: Software that splits one physical machine into several virtual ones. Proxmox is one. See Homelab 1.
Idempotent: Running it twice changes nothing the second time. The property that makes a machine rebuildable. See Ansible.
KV cache: The memory a model keeps of what it has already read, so it does not recompute it at every new token. See the KV cache trap.
LLM (large language model): An AI model that reads and writes text. "Local" means it runs on your machine, not on someone's cloud.
MCP (Model Context Protocol): An open standard, published in 2024, for connecting tools and data sources to AI agents. One tool server can serve several agents.
Model (weights): The trained AI itself, a big file of numbers. "8B" means eight billion parameters: bigger is usually smarter and always heavier.
MoE (mixture of experts): A model made of many specialist sub-networks, of which only a few work on each token. It is why very large models can run at home. See the 177B on two cards.
OCR: Reading text in an image: turning a scan into words you can search. See Homelab 12.
Open source: Software whose code anyone can read, modify and share. "Fair-code" or "source available" is close, but with restrictions.
Port: A numbered door on a machine. A web site usually answers on 443. Each open port is a door to keep closed unless it is needed.
Prefill (prompt processing): The first phase of an answer: the model reads the whole prompt. Often the one you wait for with agents. See the KV cache trap.
Private network (VPN): An encrypted tunnel between your own devices, so your phone reaches home as if it were on the home Wi-Fi. See Homelab 6.
Proxy (reverse proxy): The single entrance of your homelab: it receives every visitor, holds the certificates, and passes each request to the right service. See Homelab 2.
Pull request (PR): A proposed change to code, waiting to be reviewed and merged. Nobody, human or agent, should change the main branch without one.
Quantisation (Q4, Q8…): Compressing a model by storing its numbers with fewer bits. Q4 is smaller and a little less precise than Q8.
RAG: Letting a model answer from your own documents: the relevant passages are found first, then given to the model with the question.
Rate limiting: Capping how many requests one address may send per second. Stops password guessing and small floods. See resisting DDoS.
Self-hosted: Run on your own machine instead of rented from a company.
Skill: A written procedure an agent loads only when a task needs it. See skills.
Snapshot: A frozen copy of a disk at one moment, taken in a second on copy-on-write file systems. See snapshots.
Speculative decoding: A small model guesses a few tokens ahead and the large one checks them in one pass: same text, faster. See speculative decoding.
Tailnet: Your private Tailscale network: your devices, linked wherever they are. See Tailscale.
TLS: The encryption behind HTTPS. Versions 1.2 and 1.3 only; older ones are broken.
Token: The unit a model reads and writes: roughly a short word or a piece of a word. Speeds are counted in tokens per second.
Vibe coding: Describing what you want, letting a model write the code, judging by the result. See vibe coding and digital art.
VM (virtual machine): A complete computer simulated inside a real one, with its own system. Many VMs can share one physical machine.
VRAM: The graphics card's own memory. A model runs fast only if it fits in it. See Hardware, honestly.
Wake-on-LAN: A network packet that switches a sleeping computer on. See sleep and wake.
Webhook: An address that triggers something when it is called. Sign it, or anyone who finds it can trigger it.
WireGuard: A small, fast encrypted tunnel, part of the Linux kernel since 2020. The base of Tailscale.

Found this useful? Tell us what to measure next.