AI and LLM developer tools

Three of these tools share a single question: what is this text going to cost me, in tokens and in money? Token Counter is the blunt instrument — paste text, pick a model, get a token count alongside character and word counts. Context Window Calculator takes the same count and puts it against a model's advertised window, so you can see that your 40-page contract is at 82% of a 128k window before you discover it by way of a truncation error. AI Cost Calculator is pure arithmetic on top of published per-million prices: input tokens, output tokens, number of requests, total.

It is worth being precise about how the counting works, because token counts are the kind of number people quietly trust. The counter uses gpt-tokenizer, an implementation of OpenAI's byte-pair encoding. For OpenAI models that is the real tokenizer. For Claude and Llama it applies a small fixed multiplier to the GPT count — a calibration, not a second tokenizer. That is accurate enough to plan a budget or catch a context overflow, and it is not accurate enough to reconcile an invoice. Use it as a forecast; use your provider's usage dashboard as the ledger.

Model Comparison is the reference table the other three read from: a sortable, filterable view of context window, input price, output price, tier and multimodal support across the models this site tracks. It is a hand-maintained table, not a live feed. Providers change prices and ship new models faster than any static table can follow, so treat it as a shortlist-builder and confirm the number on the provider's own pricing page before it goes into a spreadsheet someone will act on.

Prompt work splits neatly into composition and serialisation, and there is a separate tool for each. System Prompt Builder is about composition: it gives you labelled fields for role, tone, constraints, output format and examples, and assembles them into a system prompt as plain text, as an XML-tagged block, or as a JSON message object. The value is not the string concatenation — it is that the fields remind you which parts of a system prompt you routinely forget to write down.

Prompt Formatter is about serialisation: you have a system, user and assistant turn, and you need them in ChatML, in Llama 3's header-token format, in Claude's XML convention, in an OpenAI-style message array, or as plain labelled text. This matters mostly when you are working against a raw completions endpoint or a local runtime, where getting a special token wrong produces a model that ignores your instructions rather than an error message.

Prompt Library is the place the good ones end up. It stores prompts with titles and tags in your browser's local storage, with search across title and body, and JSON export and import so a library can move between machines or into version control. Because it is local storage, it is per-browser and per-profile: clearing site data clears the library. Export is the backup.

The second half of this collection generates files whose only reader is another program, where a small syntax error costs you twenty minutes. JSON Schema Builder turns a list of named, typed, described properties into a tool definition, and wraps it in OpenAI's function-calling envelope, Anthropic's input_schema envelope, or leaves it as bare JSON Schema. It handles one level: a flat object with typed properties and a required list. Nested objects and array item schemas are not modelled, so a genuinely recursive tool definition needs hand-finishing after export.

MCP Config Generator writes the mcpServers block that Model Context Protocol clients read, for both stdio servers (command, arguments, environment) and SSE servers (URL, environment). One caveat worth knowing before you paste: arguments are split on whitespace, so a single argument that contains a space needs editing by hand afterwards.

CLAUDE.md Generator and AI Rules Generator solve the same problem for two different ecosystems — a project instruction file that a coding agent reads on every session. The first produces a CLAUDE.md with sections for overview, stack, key commands, conventions, and explicit do/don't lists. The second targets Cursor, Windsurf and GitHub Copilot, emitting .cursorrules, .windsurfrules or copilot-instructions.md with the right heading conventions for each. Neither reads your repository; they are structured prompts for you, turning a blank file into a filled-in form.

Skill / Agent Builder covers the newer shape of the same idea: a named capability with trigger phrases, a description and an instruction body, emitted as Markdown with YAML frontmatter, as YAML, or as JSON. And AI Instruction Diff closes the loop — paste two versions of a system prompt or rules file and get a line-level diff, which is how you find the one sentence someone added that changed the model's behaviour.

This collection is about tooling for people who build with hosted models. It is deliberately separate from the on-device AI tools, which are the opposite arrangement: models that download into your browser and run there, no API key and no provider involved. The two overlap in exactly one place worth knowing about — PII Detector & Redactor runs a named-entity model locally, which makes it a reasonable pre-processing step before text goes to a hosted API.

If your interest in browser-based tooling is mainly that nothing gets uploaded, the privacy-first collection sets out what that claim does and does not mean on this site.