AI Agent

Studio includes an optional AI agent panel that can help you create mods through natural conversation. Ask it to write template patches, find assets, diagnose compile errors, or explain how MENACE's systems work.

Opt-in

AI features are disabled by default. Nothing is sent to any external service until you explicitly enable the feature and start a conversation. See Enabling AI features below.

How it works

The agent panel connects to an external AI agent (like Claude Code, GitHub Copilot, OpenCode, etc) via the Agent Client Protocol (ACP). Studio provides the agent with specialised tools for searching templates, inspecting game data, reading and writing project files, and compiling mods. The agent uses these tools to carry out your requests.

You chat with the agent in natural language. It reads your project, calls Jiangyu tools as needed, and proposes changes for your approval.

Enabling AI features

AI features are off by default. To enable them:

1. Open Settings.
2. Under the AI section, toggle Enable AI features on.
3. The agent panel becomes available in the pane workspace.

Installing an agent

Before you can chat, you need an agent installed:

1. Open the agent panel (command palette → "Open Agent Panel", or split a pane and pick "Agent").
2. Click the agent dropdown in the panel header.
3. Click Add more agents… to open the registry browser.
4. Browse available agents, then click Install on the one you want.

Currently supported agents are distributed via bunx. Studio bundles its own Bun runtime, so no system Node.js is required.

API key

Most agents need an API key for their LLM provider. The agent will prompt you for authentication on first use. Jiangyu Studio does not store your API key or any account credentials. The agent process manages its own credentials.

Using the agent

Starting a conversation

Once an agent is installed and selected, type a message in the input at the bottom of the panel and press Enter. The agent connects, creates a session, and begins responding.

What you can ask

Some examples:

• "Buff the rocket launcher's damage by 25%"
• "What templates exist for modular vehicle weapons?"
• "Write a KDL patch that doubles the carbine's range"
• "Compile my mod and explain any errors"
• "What's the vanilla Vitality on squad_leader.darby?"
• "Search for audio assets related to gunfire"
• "Clone mod_weapon.medium.rocket_launcher into a heavy variant with double damage"

The agent has access to all of Jiangyu's domain tools. It can search and inspect templates, read game values, write KDL files, trigger compiles, and explain results.

Slash commands

Type / in the input to see available agent commands. These vary by agent but typically include things like /compact, and model configuration.

Permissions

When the agent wants to write or modify files, it asks for permission. You'll see an inline prompt with the proposed action and three choices:

• Allow once: approve this single action.
• Allow always: approve this and future similar actions for the session.
• Reject: deny this action so the agent can try something else.

Read-only actions (searching, inspecting, listing files) never require permission.

Tool calls

As the agent works, you'll see collapsible tool call blocks in the chat. These show what the agent is doing (searching templates, reading files, compiling, etc.). Click to expand and see full details.

Diffs

When the agent edits a file, the change is shown as an inline diff with added and removed lines highlighted.

Sessions

Conversations are saved per-project. Click the history icon in the panel header to see previous sessions and resume them. Session titles are generated by the agent based on your conversation.

Mode and config

Some agents offer configuration options (model selection, reasoning level, etc.). These appear as dropdowns in the panel header or a config popover. Changes take effect immediately for the current session.

Privacy

Agent conversations are processed by the selected agent's LLM provider. Your project files, template data, and game asset metadata may be sent to the provider when the agent uses tools. No data is sent until you actively use the agent panel.

Studio does not collect telemetry or send data to Jiangyu's developers.