Resources
Everything under Resources, in one place: calculators and comparisons, explainers by layer, and directories of the things you plug into a model.
Each directory is read from the source it points at (a repository, a page, a SKILL.md) and says what was left out and why.
- MCP servers
- 8,106
- Agent skills
- 3,022
- Open-source projects
- 468
- Explainers
- 113
Everything under Resources
What each one can reach, and which credential it wants. Installing one hands a model a key to something real.
Instructions a model loads rather than a process it connects to: reaching nothing, and needing no credential.
The ten repositories that gained the most stars each day, with every one that has ever made the board.
Projects you can self-host instead of a paid tool, each with whether it is alive, what its licence lets a business do, and how to run it.
Explainers
See all 113 in the blog →Context window length and what it costs
A large context window is a limit, not a charge; the charge is every token you actually place in it, billed as input on every call that carries it.
Edge inference: when it makes sense
Run a model on the device or near it when data must stay local, the network is unreliable, or round trips are too slow. The price is a much smaller model.
How to use the LLM cost calculator
Pick a model or let the calculator shortlist by context and volume, set monthly tokens and cached share, choose a tier, and read every seller ranked by cost.
Inference engines: what the category is for
An inference engine is the software that loads model weights onto a GPU and turns requests into tokens efficiently. You need one the moment you self-host.
Latency or throughput: which to tune for
Tune for latency when a person is waiting on the answer and for throughput when nobody is. Batching is the dial, and it moves cost and speed opposite ways.
LLM costs people forget until the invoice
System prompts, retries, chat history, evals, embeddings and tool calls all bill tokens, and a budget built from one call per user action misses most of them.
What we would write, and what each one needs
Published in advance because a reader deciding whether to trust this directory learns more from what we refuse to assert than from what we do.
Whether GPU prices are converging
NOT WRITTENA price series long enough that a movement means something — roughly twelve weeks. We hold a handful of capture days, so we can say how far apart sellers are today but not whether the gap is closing. The half that needed no history is written.
Which models are fastest per pound
NOT WRITTENBenchmark rows, of which we hold none. Doing this properly means running the benchmarks ourselves — a different business from reading published pages, with its own hardware bill. The alternative is to cite whoever did measure it rather than restate their work as ours.
What open weights actually save you
NOT WRITTENA flag that can say “closed”. Today the open-source column is set when a repository is found and never set false, so a closed product and an unchecked one are the same value — we could not name a single closed-weight product with confidence. It is the three-state rule the rest of the directory follows, applied to a column that has only two of them.
What GPU do I need — use-case pages
Each page states the rule it selects on — never a ranking to trust. A part with no memory figure on record fails the rule rather than slipping through it.
Frequently asked
Are MCP servers safe to install?
MCP server vs agent skill — which do I need?
Elsewhere on the site
The pages that already explain themselves: methodology, the price spread and how to cite us. Last updated 16 September 2026.