What Shard can do in your place.

Shard doesn’t only suggest code. It has real Studio tools, uses them on the place you have open, and shows you every step it took.

Eleven tools, and you see every call.

The model picks the tools it needs. Each call gets its own row in the chat. Open a row to see what Shard sent, what came back and what changed.

Shard never edits its own plugin, and it doesn’t press Play. Its tools pause while you playtest.

Every tool in the docs

Under each name: how that call shows up in the chat.

Look around

ReadScriptsRead 3 scripts: Reads scripts, including edits you haven’t saved yet. Whole scripts, or only the part it needs.
InspectPathInspected 4 objects: Looks at an object: its children, properties, attributes and tags.
SearchCodeSearched scripts: Searches the text of every script, like Find in All Scripts.
SearchObjectsSearched the place: Finds objects by name, class, tag or attribute.
GetStudioStateChecked Studio: Checks your selection, the scripts you have open, and whether Studio is in edit or play mode.
GetOutputChecked Studio output: Reads the Output window: prints, warnings and errors.

Make changes

EditScriptsEdited 1 script: Changes one spot, several spots or a whole script, or creates a new Script, LocalScript or ModuleScript. It must read a script first, and it refuses to write over a newer version.
ApplyInstanceChangesCreated 3 objects: Creates objects or updates their properties, attributes and tags: numbers, text, booleans, Vector3, Color3, CFrame, UDim2 and Enums.
RunCodeRan code in Studio: Runs Luau in edit mode for jobs the other tools can’t do, and reports logs, return values and errors. It is isolated from the plugin, your login and your keys.

Learn and plan

BrowseWebLooked up Roblox docs: Searches the official Roblox documentation, runs a web search, or reads one web page. Pages are treated as information, never as instructions.
LoadSkillLoaded datastore guidance: Loads a built-in playbook for the kind of task at hand, like physics or UI.
Team toolsStarted agents: When you ask for agents, the lead gets tools to start them, check on them, message them and stop them.

Ten playbooks for Roblox work.

A skill is a step-by-step workflow for one kind of Roblox work, with links to the official docs. Shard loads the right one by itself, or you type / in the message box and pick one.

Skills only guide Shard. Loading one never changes your place by itself.

Skills in the docs

/physics: Impulses, movers and physical interactions. Bounce pads included.
/gameplay: Plan, build and check a gameplay feature from start to finish.
/ui: Interfaces that fit your game and work on every screen.
/client-server: Secure remotes, server authority and replicated state.
/datastore: Saved data with safe schemas, retries and server ownership.
/animation: Animation tracks with the right rig, ownership and lifecycle.
/debug: Reproduce a bug, trace the cause, make a focused fix.
/optimize: Find measured bottlenecks instead of guessing.
/refactor: Cleaner structure, same behavior.
/plugin: Build or debug your own Studio plugin.

A crew, when you ask for one.

Ask for a team and the lead Shard starts several agents at the same time, each with its own task. They message each other and report back.

Every agent step is a model request, so a crew of three uses about three times the messages of one Shard.

How agents work

2 agents working uiserver

In the chat, one row shows who is working. Click a name to open that agent’s own conversation.

Only when you ask: Say “use two agents” and the lead names them, splits the work and keeps coordinating.
One owner per script: An agent owns the scripts and objects it edits until it’s done. Only one change runs at a time across the whole crew.
Watch every agent: Open Chat details, then Subagents, to see what each one is doing and read its conversation.
One Stop for all: Stop ends the lead and every running agent.

A chat you can keep talking to.

Long build sessions need more than a text box. You can send while Shard works, stop it, come back to an old chat, and see what a chat has used.

Chatting in the docs

Queue and Steer: Send while Shard works. Your message waits in a queue, or press Steer to hand it to the running task before its next step.
Its own Stop button: Stop ends the run and every agent working for it. Send and Stop never share a button.
Context compaction: When a chat gets close to the model’s limit, Shard summarizes the older part and keeps the latest part word for word. Or type /compact.
Chat history: Chats save on your computer. Pin the ones you use most. Each chat remembers its own provider, model and reasoning level.
Usage details: See tokens, an estimated cost, prompt cache use and how full the context is, for every chat.
Six themes: Classic, Neon, Slate, Wood, Classic Light and Sandstone, in Settings, then Appearance.

You stay in charge of your place.

Shard changes real scripts and real instances, so every change has to be easy to see and easy to take back.

Undo in the docs

Shard: run codeone RunCode call
Shard: edit scriptsone script edit
Shard: change instancesone instance change

What Shard adds to Studio’s undo history. Each line is one Ctrl Z.

Studio undo: Each script edit, instance change and RunCode call is its own named undo step, so you can step back exactly as far as you want.
No stale overwrites: If a script changed since Shard read it, the edit is refused and Shard reads it again.
Sandboxed test code: RunCode runs in its own module, away from Shard’s plugin, your login and your API keys. It has a 120 second default time limit, and Shard can ask for more.
Context on request: Your selection and open scripts are not sent with every message. Shard checks only when it needs to.
Review what changed: After a run, a Review section lists each changed script with the lines added and removed.

One answer can contain several undo steps, so you may need to press Ctrl Z more than once.

Undo can’t take back things RunCode did outside your place, like a web request. There is no approval step before RunCode yet, so save a copy before big tasks.

Three ways to connect a model.

Shard doesn’t pick the model for you. Every request counts toward the same beta limits, whichever way you connect.

Models and providers

Your own API keyworks as a guest: OpenAI, Anthropic, Gemini, Grok (xAI), Groq, OpenRouter, Amazon Bedrock, Z.AI, Kimi (Moonshot) and Meta Model API. You pay the provider directly.
A connected accountneeds a Shard account: Use the ChatGPT plan you already have through Codex, or your Claude plan through Claude Code.
Shard-hosted modelsneeds a Shard account: No key and no subscription. Pick a model whose name starts with “Shard:” and pay with credits.

Try it on your own game.

Install Shard and ask for something small first. You can always press Ctrl Z.

Get Shard How to get started

Image attachmentsnot yet
Answers that appear word by wordnot yet
An approval step before test code runsnot yet

Shard is in public beta. These pieces are still missing. The changelog says what has landed.