Free AI & LLM Developer Tools

Tokens are the unit that everything is priced and limited in

A model does not see words. It sees tokens, and the count depends on which tokeniser the model uses, so the same paragraph can differ noticeably between providers. English prose runs around 1.33 tokens per word; source code is closer to 2.4 and JSON closer to 2.7, which is why a page of structured data costs roughly twice a page of prose. The token counter runs the real BPE vocabulary rather than estimating, words to tokens gives the quick ratio, the context window calculator checks a document against each model's window with room reserved for the reply, and the API cost calculator and model comparison price the same workload across providers. The prompt token optimiser shows what the padding in a system prompt costs when it is sent a million times.

Getting data into and out of a model

Retrieval and fine-tuning are mostly format work. The RAG chunker splits documents with configurable overlap and reports the token count of every chunk, the embedding similarity calculator compares vectors by cosine, dot product or distance, and the fine-tuning dataset converter moves a dataset between the OpenAI, Anthropic, Gemini, ShareGPT, Alpaca and ChatML shapes. The JSONL validator catches the structural faults that make an upload fail hours later: a bad line, a broken role alternation, a row over the limit. On the wire side, the tool schema converter translates a function definition between provider shapes, MCP and plain JSON Schema, and the stream parser reconstructs a message, its tool calls and its usage from a raw SSE dump. The prompt template generator expands one template over a CSV or JSONL file, and the chat export formatter turns a ChatGPT or Claude export into readable Markdown.

Self-hosting and sampling

Before downloading weights, work out whether they fit. The GGUF size calculator uses effective bits per weight rather than the nominal quantisation level, and the VRAM calculator adds the KV cache, which grows with context length and concurrency and is what usually causes the out-of-memory error. The temperature and top-p visualiser shows what those sampling numbers do to the distribution, which is more useful than adjusting them by feel.

None of these tools calls a model. They are tokenisation, arithmetic, parsing and format conversion, which is why there is no API key field and why your prompts and datasets stay in the browser.