Glossary
New to AI coding? These are the words the docs use, in plain terms.
Models and prompts
Model
The AI that reads your prompt and writes the reply, such as Claude Opus, Claude Sonnet or GPT-5.6. Larger models handle harder work, but cost more per prompt and use up a plan’s allowance faster. Cursor, Claude Code and Codex each let you pick one.
Prompt
What you type to the AI: a question, a task or a reply such as “continue”. OptiFlume looks at each prompt before it is sent.
Effort
How long a model thinks before it answers (also called reasoning effort or thinking level). High effort helps with hard problems but takes longer and uses more tokens; low effort is quicker for routine work. In Claude Code and Codex, OptiFlume can change the effort as well as the model.
Token
The unit models read and write in, about three quarters of a word. Prices, usage limits and context windows are all counted in tokens.
Model switch
Moving a prompt to a different model, or a different effort, from the one you had selected. OptiFlume switches one prompt at a time, shows it where you work and says why; your saved default model stays as it is. See Routing and overrides.
Override
Keeping your own model for a prompt: start it with ~ in Claude Code and Codex, or ! in Cursor, and OptiFlume leaves it alone.
Agent
An AI that does more than answer: it reads files, runs commands and edits code on its own, step by step, until the task is done. Cursor, Claude Code and Codex all work this way. The Guardian watches what an agent is about to do.
Context and memory
Context
Everything the model sees along with your prompt: the conversation so far, the files it has read and the output of the commands it ran. The model knows only what is in its context.
Context window
The most context a model can take at once, counted in tokens. When a conversation gets close to it, your tool has to make room.
Compaction
How your tool makes room in a long conversation: it replaces the older turns with a summary. The conversation keeps going, but details can be lost, such as a decision made an hour ago or an approach that already failed.
Working memory
OptiFlume’s short notes on a session: what you are trying to do, the decisions made and the approaches that already failed. They go to the model after a compaction, and to any model OptiFlume switches to, so it picks up where you were. Kept on your Mac only. See Working memory and session watchdog.
Prompt cache
A short-lived copy of the recent conversation that the model provider keeps, so the next request can re-read it at a lower price. It expires after a few minutes without a request, and each model has its own. See Does a switch mid-session cost more?
Session
One conversation in your tool, from its first prompt until you start a new one.
Plans and limits
Usage limits (allowance)
Claude and ChatGPT subscriptions include a set amount of use that resets on a schedule, instead of a bill per token. Larger models and higher effort use it up faster. On those plans, what OptiFlume saves you is allowance, not dollars.
Per-token pricing
Paying for each token a model reads and writes, at that model’s price. Cursor counts usage this way, at each model’s per-token price, which is why OptiFlume can show a dollar saving in Cursor.
How OptiFlume behaves
Fails open
If any part of OptiFlume fails or takes too long, your prompt goes through unchanged on the model you selected. OptiFlume can make a prompt cheaper or better suited, but it can never stop your work.
Suggest
OptiFlume names a better-suited model and your prompt goes ahead on yours, without a switch. This is what the free tier does in every tool, and a mode you can choose in Cursor, Claude Code and Codex. See Modes and Stop/Keep.
Local models
Two small AI models that OptiFlume runs on your Mac, about 1 GB together: one helps judge each prompt, the other writes working-memory notes. They never send anything off your Mac. See The local models.
Hook
A small add-on your tool runs at set moments, such as when you send a prompt. Codex asks you to approve OptiFlume’s hooks once; they show the reason for each switch and run the Guardian.
Guardian
OptiFlume’s safety layer: it asks before secret files, keys and personal data reach the model, and before deletes and destructive commands. Free on every plan. See Guardian.
Session watchdog
Part of the Guardian that notices when a session is going in circles, such as the same fix failing again and again, and says so.