Tags, commands and modifiers

Put a tag in a system or user message, usually the character definition, and that asset applies to the request on top of whatever is on the key. Every public asset shows its eight-character code in Discovery.

Up to 16 tags count per request; a tag that repeats counts once. Codes are not case-sensitive, and spaces around the colon are fine ([preset: CODE]); with spaces, write the code in capitals so an ordinary bracketed word such as [Prompt: Continue] stays part of the story. A slug needs a dash and lower case. A tag that names nothing Composite knows is left alone if it is not a code, and dropped if it is one. Add turns=N to a tag and it applies only until N of your messages have followed it.

A request can run any number of presets and prompts. They apply in this order: what is attached to the key, then the key's Default tags, then profiles named in the request, then the other tags in message order. System text and prompts from every layer are joined in that order; where two presets set the same option (temperature, prefill) the later one wins. An asset named twice applies once. A profile named in a request brings only its assets and model pool: its memory layer and helper funding are never applied, and assets it points to that you cannot read are skipped. Add --show-info to a message to see the stack that ran.

A key or profile can also hold a model pool. Each request goes to one model from the pool, drawn at random or by weight (one model per line on the Keys page, model*3 to weight it). A pick the account can't use is skipped and the next draw is tried.

Written anywhere in the message you send. The flag is removed before the prompt is built, and only the newest message can switch one on. The status lines are added to the reply and are stripped again if your app sends them back in the history.

Tools are handled by Composite and never sent to your provider as tool definitions. Turn one on with a tag or shortcut in the character definition or your message.

Image links in your newest message (http or https, ending in .png, .jpg, .gif or .webp, up to 4) are sent to vision models as images; the model's provider fetches them, Composite never does. Other models see the link as text.

These make extra model calls. On a key with your own provider they run on your provider, so they cost Composite nothing; on a key you have set to use Composite's helper models, or on a Composite key, they bill your credits or plan. An account that can't pay for helper calls doesn't get them: the reply is simply the model's own.

Plugin nodes for these are listed in the plugin builder under "After the reply" and "Before the reply". A picked side-by-side answer is kept as preference data only if you turn on "Keep my side-by-side picks" in Account settings; it is deleted after 90 days, or when you turn the setting off.

Shorthand for Composite's official plugins and tools. Write them in square brackets and in capitals, like [SHORTREPLY]; the older angle-bracket spelling (

Shorthand for Composite's official plugins, written the way other roleplay proxies spell them so existing bots keep working.

Shorthand for Composite's official plugins, written the way other roleplay proxies spell them so existing bots keep working.

A plugin can define its own command. Type it in square brackets in your message, for example [recap], and any plugin on the key (or tagged into the bot) that listens for that name runs its branch. The command is removed from your message unless the plugin keeps it, and a message that is only the command is sent as-is so providers don't reject an empty turn. Which commands exist depends on the plugins you use; each plugin's Discovery card describes its own.

Written inside a lorebook entry and filled in each time the entry is inserted.

Added to a Composite catalog model id, such as minimax/minimax-m3:fast. The endpoint variations exist only for models with more than one usable endpoint; the Models page lists each model's variations. Reasoning effort suffixes also work with OpenAI-compatible provider keys: Composite removes the suffix and sends the effort separately. Native Anthropic provider keys do not translate these efforts into thinking controls.

Every error says where it came from. Frontends that only show the message get it in words: Provider error: means the model provider failed or refused, Composite error: means we did. Limits, walls and spending caps carry no prefix and name the next step instead. Upstream outages link to the status page.

Responses also carry x-composite-upstream-calls (how many calls went to the provider, retries and fallbacks included) and x-composite-served-model (the model that answered).