Every request to the model carries the definitions of the tools it may call. Aiden used to send all of its built-in tools on every request, which was the biggest fixed cost in the context meter (Built-in tools in the Context tab). Two changes keep that cost small.
Skills are listed once, in <available_skills> in the system prompt, and loaded through a single skill tool that takes the skill's name. Before, every skill was its own tool (skill_<key>) and was also listed in the prompt, so each installed skill cost tokens twice.
- The tool loads only skills the model may invoke in that chat. A workspace skill is unavailable under No access, and a skill that turned off its model invocation is never listed.
- Turning Skills off in Settings applies from the next
skillcall, as before. - A Bot's
skilltool loads only the skills granted to that Bot. The requested name is checked against the Bot's current grants when the call is made, so taking a skill away stops it at once. - The per-skill key (
skillToolKey) is still each skill's internal identity for grants and slash commands; it is no longer a tool name. The tool matches a listed name first and a key only when no name matches.
The CLI keeps Pi's own skill loading and its skill_aiden_app tool; CLI-hosted Bots keep their own per-skill tools.
A chat starts with its everyday tools: files, commands, web search and page reading, questions, to-dos, memory, the skill tool, codemode and tool_search, the browser gateway and Bot cards. It also gets load_tools, whose description lists every other built-in tool in one line. The model loads what it needs by name, and the loaded tools are declared from its next step. The list lives in ON_DEMAND_TOOL_NAMES in main/services/on-demand-tools.ts: subagents, the advisor, scheduling, visuals and artifacts, images, classification, background jobs, pull-request following, Aiden's app tools, past-chat recall, Computer Use, form filling, devices, Bot routines and image inspection.
- Nothing is granted by loading. Every tool is still built, admitted, approved and run exactly as before; only the moment its definition is sent changes.
load_toolsrefuses names the chat was not given. - Once a chat has loaded or called a tool, later requests in that chat declare it from the start, in the order the chat first loaded it, so the tool list stays the same and its definitions stay in the provider's prompt cache.
- A direct call to a tool that was not loaded yet fails as "not found", and the tool is declared on the next step so the retry works.
- The system prompt still describes held-back tools such as
aiden_helpandsubagent;load_toolstells the model to load a tool before its first use. - Workspace chats use this only when codemode is on (it is by default). Codemode scripts can still call every callable tool without loading it.
- Bots use the same list. Their tools stay installed, and a request simply leaves out the definitions of tools the Bot has not loaded. Bot computers (
computer,computer_shell) stay declared from the start. - The context meter counts only what is declared. Held-back tools are not shown as a deferred connection.
Measured on 2026-10-28 with the end-to-end LM Studio fixture (a fresh workspace chat with Full access, the two built-in skills, codemode on), counting a quarter token per character of each tool definition as the context meter does:
| Before | After | |
|---|---|---|
| Tools declared | 45 | 22 |
| Built-in tools in the Context tab | 9.7K | 3.6K |
Each installed skill now costs one prompt line instead of a tool definition plus a line. Computer Use (about 1.8K) and the device tools (about 0.8K) are held back too when those features are on, so a Mac with them enabled saves more. A Full Access Bot with subagents, routines, image sharing and Computer Use declares about 3.7K fewer tokens per request.