Use Gala from AI tools

There are four ways to put Gala in reach of an AI assistant or agent. They are not alternatives to one another: the plugin runs the MCP server for you, and the agent guides and llms.txt are what an agent reads to write code instead of calling tools.

Way	What it is	Use it when
The omni-tool plugin for Claude Code	A Claude Code plugin from the gala-chain marketplace: it starts the Gala MCP server and adds a teaching agent, three slash commands, a skill and a hook. Two smaller plugins, launchpad and swap, run the same server with fewer tools.	You use Claude Code and want one install that works and teaches. See the omni-tool plugin.
The MCP server @gala-chain/gala-mcp-server	The server on its own, over stdio, for any MCP client: Claude Code, Claude Desktop, Cursor or your own. 324 tools and 286 prompts in toolsets you select.	You use another client, you want to choose toolsets and environment yourself, or you build your own agent. See MCP server setup.
The per-package agent guides	One short, task-oriented page per package at /<package>/agents.md: which package for which job, recipes, invariants with the test that enforces each, failures with their fix.	An agent writes application code against an SDK instead of calling tools. Start at the agent guide of this package.
llms.txt and the machine-readable files	/llms.txt indexes every page as raw markdown, /llms-full.txt and /<package>/llms-full.txt carry whole packages, and /mcp/tools.json lists every tool with its parameters.	You feed documentation to a model or build your own retrieval.

Which one to choose

• You use Claude Code and want to start fast: install the plugin. It configures the server with the prod network and no signing key, so the first session can only read.
• You use Claude Desktop, Cursor or another MCP client: configure the server directly. The plugin is Claude Code only.
• You want a small tool surface: select toolsets (core, launchpad, gswap), or install the launchpad or swap plugin. A smaller set costs the model less context and gives it fewer ways to be wrong.
• You want the model to write code, not call tools: point it at the agent guide of the SDK and at llms-full.txt. Calling a tool is a signed action in your wallet; generating code is not.
• You are unsure whether to give an agent a key: do not. Every read tool works without one. Read security and keys before you add a key.

What is behind each way

All of them end in the same three SDKs: @gala-chain/gala-core-sdk (wallets, signing, the GALA balance, bridging), @gala-chain/launchpad-sdk (the Launchpad backend) and @gala-chain/gswap-sdk (the GSwap DEX). The server wraps them as tools; the guides document them as code. An answer you get from a tool and the code an agent writes from the guide go through the same signing and the same address rules.

Where to look things up

• Every tool and prompt, with parameters, per toolset: the reference pages of this package, or /mcp/tools.json for machines.
• Which docs answer a question, from inside a session: the gala_launchpad_search_gala_docs tool, which works offline too.
• What went wrong: troubleshooting.