Navigation
ELOWEN · Documentation

Last updated: October 27, 2026

Browse documentation · Load tools on demand
Skills, MCP and API

Load tools on demand

Keep a conversation with many tools small: Chetty describes only the tools it needs at first and fetches the rest when it reaches for one.

When tools are deferred

Deferred loading is on by default. It applies when a conversation has more than 10 tools without an explicit choice, and it counts MCP tools and plugin tools alike.

Then only the core tools (Bash, Read, Edit, Grep and Write) stay loaded in full and every other tool is loaded on demand. A smaller set stays available immediately, except tools that their plugin always defers, such as ListMcpResources and ReadMcpResource.

Admin: The switch and the threshold live in Settings → <assistant name> AI → Runtime, under Tool loading: Enable deferred tool loading and Automatic threshold (1–100, default 10). Manage tools sets a tool or a whole plugin to immediate or on demand. Changes apply to new and recreated conversations; existing conversations keep the tool set they started with.

Find a tool

Without code mode or provider-hosted search, the assistant looks tools up with a word match over tool names, descriptions and parameters. It is not a semantic search, so English terms work best with English tool descriptions.

A free-text query accepts an optional limit, which defaults to 8:

ToolSearch({"query":"create github issue","limit":8})

Load an exact tool

For a known name, select: loads exactly the tools listed. It reports which names were loaded, refused or unknown, and the search limit does not cut that list short:

ToolSearch({"query":"select:mcp__github__create_issue"})

The answer lists the loaded tool names, and their full definitions are available from the next model call.

What deferral does not change

Deferral only shrinks the prompt. It never changes server permissions, account grants or tool policy: a tool that is unavailable to the account stays unavailable whether it was loaded early or on demand.

In Code mode there is no separate search step. Deferred tools stay callable from the script, and the model filters the ALL_TOOLS list by name and description before it reads a full declaration. With a provider's own tool search, some Claude models and GPT-5.4 or newer, the core tools are sent in full and the provider finds the rest.

See MCP Integration for server tools, and Code mode for running tools from a script.