AI infrastructure tools
The bytes on the wire, the schema the API accepts, and what it costs.
The deterministic half of running a model in production: the SSE frames a streaming endpoint sends, the JSON Schema subset a strict tool call accepts, the difference between max_tokens and the context window, and what a cache write costs against a cache read. These tools take token counts as input rather than estimating them, because an approximate token count is wrong by an unpredictable margin on code and a confidently wrong number is worse than none.
Free, with no sign-up. Every one runs entirely in your browser: nothing you paste is uploaded, and there is no server to send it to.
LLM SSE Stream Decoder
Reassemble the message from the wire
Decode a raw LLM SSE stream from the Anthropic or OpenAI API, reassemble the message, and see exactly where a truncated response stopped.
LLM Tool Definition Converter
Anthropic, OpenAI and MCP, both ways
Convert an LLM tool definition between Anthropic, OpenAI and MCP formats, with the fields that have no equivalent named rather than dropped.
LLM JSON Schema Strict Mode Checker
What structured output will refuse
Check whether an LLM JSON schema survives strict structured output: recursion, minimum and maxLength are rejected, additionalProperties must be false.
LLM Chat Template Renderer
What the model actually sees
Render messages into the exact prompt string for ChatML, Llama 3, Mistral or Gemma, and catch the doubled BOS token that quietly degrades output.
LLM Rate Limit Header Decoder
Which bucket actually 429s
Decode Anthropic and OpenAI rate limit headers: which bucket is closest to empty, when it refills, and which of the two reset formats you are reading.
LiteLLM Proxy Config Generator
The prefix decides where it goes
Generate a LiteLLM config.yaml with the provider prefix that routes the request, keys as environment variable names, and a load-balanced deployment group.
vLLM Server Config Generator
Whether it will actually start
Build a vllm serve command and check the KV cache arithmetic: whether max-model-len fits, how many sequences really run, and where tensor parallelism breaks.
Ollama Modelfile Generator
num_ctx is not the model's context
Generate an Ollama Modelfile with num_ctx set deliberately, plus the server variables that decide the rest, including the one that divides context between requests.
Vector Database Sizing Calculator
The graph does not shrink
Size a vector collection: raw vectors, the HNSW or IVF index, payload, replicas and the build peak. Quantisation shrinks the vectors and not the graph.
LLM Agent Loop Cost Calculator
Turn 20 pays for turns 1 to 19
Cost a tool-calling agent run. Input tokens grow with the square of the turn count, which is what a cost-per-turn estimate misses, and caching divides it.
LLM Fine-Tuning Cost Calculator
Training is the cheap part
Price a fine-tune against a cached few-shot prompt: training, the per-request inference premium, hourly hosting, and the break-even between them.
RAG Chunking and Embedding Planner
Overlap costs more than it looks
Plan RAG chunking: how many chunks a corpus becomes, how much overlap adds to the tokens embedded, and whether the retrieved passages fit your context budget.
LLM Token Counter
Exact, in your browser, or not at all
Count tokens exactly with a real BPE vocabulary running in your browser. Nothing is uploaded, and characters divided by four is off by up to 69%.
LLM Inference Speed Calculator
Two bounds, and only one is the GPU
Estimate local LLM tokens per second: decode bounded by memory bandwidth, prefill bounded by compute, plus time to first token and the layer split when it does not fit.
LLM Quantization Comparison
Q4_K_M is not four bits
Compare every GGUF quantisation of a model: measured bits per weight, memory needed at your context, what fits, and what each one decodes at.
What LLM Can I Run on My GPU
Fitting is not the same as usable
Pick your GPU or Mac and see which local LLMs actually run on it, ranked, with estimated tokens per second and the quantisation that fits.
GPU VRAM Calculator for LLM Inference
Weights, KV cache and what actually fits
Work out the VRAM a local LLM needs: weights at any quantisation plus the KV cache at your context and batch size, from the real architecture.
LLM Context Window Planner
Where the budget runs out
Plan an LLM context window budget across system prompt, tools, history and output, and find where max_tokens truncates the answer instead of the input.
LLM Prompt Caching Planner
Where to put the breakpoint
Plan LLM prompt caching breakpoints and get the break-even read count: cache writes cost 1.25x at five minutes and 2x at one hour, reads cost 0.1x.
LLM API Cost Calculator
The number that matches the invoice
Calculate LLM API cost from your own token volumes and rates, including cache reads and writes, so the number matches the invoice rather than a guess.
LLM Model Cost Comparison
Two models, one workload, your rates
Compare LLM model cost side by side across Anthropic, OpenAI, Google and more, with your own rates overriding the table on either side.
LLM API Rate Limit Planner
Which limit you hit first
Plan around LLM API rate limits: what your request shape costs in tokens per minute, and which of RPM, input TPM or output TPM you hit first.
Reading on ai infrastructure
The problems these tools check for, argued out at length.