$ capsul glossary
Glossary
The words that decide what a coding agent costs, defined in one sentence each and then explained.
- Token
- A token is the unit a language model reads and bills in: a common word, a word fragment, or a punctuation mark, roughly three to four characters of English on average.
- Context window
- The context window is the maximum number of tokens a model can consider at once, covering the instructions, the conversation, and everything sent alongside the question.
- Input tokens and output tokens
- Input tokens are everything sent to the model on a request; output tokens are what it writes back, and in agent work the input side dominates the bill by an order of magnitude or more.
- Token budget
- A token budget is a hard ceiling on how much context is allowed into a request, enforced before the request is sent rather than negotiated with the model.
- Prompt caching
- Prompt caching lets a provider reuse the processed form of a repeated prefix across requests, billing those tokens at a fraction of the fresh rate.
- Cache hit rate
- The cache hit rate is the share of input tokens on a request that were served from a prompt cache rather than processed fresh.
- Context rot
- Context rot is the degradation in a model's answers as its context grows, through dilution of the relevant material, contradictions between old and new content, and accumulated dead ends.
- Agentic loop
- An agentic loop is the cycle in which a model calls a tool, reads the result, and decides what to do next, repeating until it produces an answer.
- Tool call
- A tool call is a structured request from a model to run something outside itself, such as reading a file or executing a command, whose result is fed back into the context.
- Repository map
- A repository map is a compressed outline of a codebase, listing its files and the symbols they define, so a model can locate things without reading everything.
- Rate limit and usage limit
- A rate limit caps how much you may send in a window of time, measured on tokens processed rather than on the number of messages sent.
- CLI coding agent
- A CLI coding agent is a command-line program that lets a model read, run and edit things in a working directory, under whatever approval rules you set.
- CLAUDE.md
- CLAUDE.md is a Markdown file of standing instructions that Claude Code loads at the start of every session, which makes its contents part of every request the session sends.
- AGENTS.md
- AGENTS.md is a Markdown file of standing project instructions addressed to AI coding agents in general rather than to one product, which an agent loads into its context before it starts work.
- Context compaction
- Context compaction replaces an agent's conversation history with a summary written by the model, so that a long session fits back inside the context window at the price of detail.
- Subagent
- A subagent is a separate agent instance, with its own context window, system prompt and tool access, to which a coding agent hands a task and from which it receives only a summary of the result.
- Model Context Protocol (MCP)
- The Model Context Protocol (MCP) is an open standard for connecting AI applications to external tools and data, in which a server exposes tools, resources or prompts that any compatible client, such as a coding agent, can use.
- Tool definition
- A tool definition is the name, description and input schema that describe a tool to a model, counted as input when a client includes it in a request.
- Cache write
- A cache write is the processing and storing of a prompt prefix for reuse, billed above the normal input price in exchange for much cheaper reads on the requests that follow.
- Cache TTL
- A cache TTL (time to live) is how long a cached prompt prefix survives without being used, after which the next request has to process it and write it to the cache again.
- Cache-weighted input tokens
- Cache-weighted input tokens measure a request's input with each token counted in proportion to its price relative to fresh input: fresh tokens in full, cache reads at a tenth and cache writes at their premium.
- API-equivalent cost
- API-equivalent cost is what a session's token usage would cost at a provider's list API prices, a yardstick for comparing models and workloads when the work actually ran on a subscription that does not bill per token.
- Cost per task
- Cost per task is the total spent to complete one unit of work from start to finish, every request, retry and delegated call included, as opposed to the price of a token or of a single request.
- System prompt
- A system prompt is the block of instructions sent ahead of the conversation in every request to a model, setting the role the model plays and the rules it follows.
- Tokenizer
- A tokenizer is the component that cuts text into the tokens a model reads and bills, which is why the same text can come to a different number of tokens on two different models.
- Extended thinking
- Extended thinking is the reasoning a model generates as tokens before it answers, billed as output tokens whether or not that reasoning is shown to you.