The Assistant

Remix Studio includes a persistent, tool-using assistant for planning and operating the workspace in chat. It uses the same account-scoped tool definitions as the MCP server, while providing an in-app confirmation UI and conversation history.

Requirements and Model Selection

The assistant does not use a hidden global model credential. Create a provider in Remix Studio, then choose an assistant-capable text model in the composer or Assistant Settings.

The chat runtime currently supports provider records of these types:

• Google AI
• OpenAI
• Claude (Anthropic)
• Alibaba Cloud
• Kimi (Moonshot AI)
• MiniMax

Other provider families may generate project assets without being available as the assistant's chat model. A saved provider must belong to the user and contain a decryptable API key. API URL overrides are honored after the same URL-safety validation used elsewhere in the server.

The selected provider and model are stored on the conversation. The most recent selection is also remembered locally as the default for a new chat.

Workspace Context

The assistant can read account-scoped workspace data through tools rather than relying on a stale copy in the prompt. Read capabilities include:

• Libraries and paginated/searchable library items, or one item by ID.
• Storage totals and category breakdowns.
• Project workflows, album summaries, album items with their tags and filenames, or one album item by ID with its full prompt or generated text.
• Per-project job counts: drafts staged, queue depth, completed and failed runs, and album totals — and the queue monitor's counts across every project and provider.
• Usable provider/model pairings.
• Social accounts, campaigns, posts, post text, and post media metadata.

The composer can bind workspace context: type @ to mention a library, project, campaign or post, or paste a reference copied with Copy ID for AI from a library item, album item or post — library:<libraryId>/item:<itemId>, project:<projectId>/album-item:<itemId> or campaign:<campaignId>/post:<postId>. The assistant's instructions describe these shapes, so it reads the exact item instead of searching for it.

The composer also takes images: attach or paste them, and on Android pick a copied image, GIF or sticker in Gboard, in versions of Chrome that let a keyboard insert images. With a Google AI provider selected, the microphone control records audio in the browser and transcribes it through the configured Google provider before sending the resulting text.

Skills

Assistant skills are prompt templates stored in a dedicated text library. Manage them under Assistant Settings → Skills, then type / in the composer to search and insert one.

Skills help standardize recurring instructions; they do not add server permissions or bypass tool confirmation. Editing or deleting a skill is the same kind of library-item operation as editing other reusable text.

Tools and What They Can Change

The assistant can:

• Create and describe libraries.
• Create prompts individually or in batches.
• Update prompt text, titles, and tags, including batch metadata updates.
• Create projects with complete workflows — regular and Civitai; ComfyUI projects are set up in the app — and update existing projects.
• Tag album items and rename album files, one at a time or in bulk.
• Stage draft jobs from a project's workflow, and start all of them or a chosen number. Starting queues real generation and spends provider credits, so it is a confirmation-gated write like any other.
• Clear failed jobs from the queue.
• Export a project's album or a whole project bundle.
• Create and update campaigns and posts, attach owned media, and schedule posts.

What it changes shows up on the pages you have open without a reload — see Live Updates.

The file-upload tools are offered only to external MCP clients, since the in-app assistant has no way to send a file's bytes; attach media you already have in libraries or albums instead.

The tool catalog is visible under Assistant Settings → Tools, including each tool's input schema, category, and whether approval is required.

Confirmation and Approval Modes

Tools are categorized as read, write, or destructive.

• Read tools run without a mutation confirmation.
• Write tools pause and show a human-readable summary plus the exact arguments.
• Destructive tools also require explicit confirmation and cannot be placed on the auto-approved write list.

For a write tool, you can approve only the pending call or approve that tool for future calls in the current conversation. Conversation approvals are stored per chat and can be changed from the tool-approvals dialog. They do not become a global permission for other conversations.

Pending confirmations expire. Cancelling one records the cancellation and lets the conversation continue without applying the change.

WARNING

Approving a tool means later calls to that named write tool in the same conversation can execute without another prompt. Review the approved-tools list when a conversation's purpose changes.

Turns Survive a Refresh

A turn runs on the server, independent of the page that started it. Reload the page, lock your phone or lose the connection mid-run, and the conversation reattaches when you come back: it replays what you missed, puts the progress line on what the assistant is doing now, and follows the turn to its result. While it reconnects the chat shows Reconnecting to the assistant.... A pending confirmation is part of the conversation too, so the confirm/cancel card is there in any tab or browser.

Sending a message to a conversation that is already running a turn attaches to that turn instead of starting a second one.

Stop, which replaces Send while the assistant works, ends the turn at its next safe point: a model call in flight is abandoned at once, a tool that is already running finishes so its result is on record, and later tool calls in that step are recorded as cancelled. The transcript ends with Stopped at your request., and the conversation takes a new message straight away. Deleting a conversation also stops its running turn.

Turns are held by the server process, so a server restart still ends a turn in progress; the conversation then shows what was saved before the restart.

Resources

The right-hand sidebar lists what a conversation has changed: one compact row per library, project, campaign or post it created or modified, with the tool behind its latest change underneath (Create Prompt, Tag Album Items, Update Post Text), most recent first. Read-only lookups never add a row. Each row links into the app; hover it to drop the record into the composer as an @ mention, or to remove the row. The list is stored with the conversation, so it is there after a reload and in any browser.

Conversation History

Messages, tool calls, structured arguments/results, token counts, errors, and pending confirmations are persisted separately from project generation jobs. This keeps chat history independent of the generation queue.

The history screen can reopen previous conversations and archive chats you no longer need in the active list. Deleting a conversation removes its messages and confirmation records; it does not undo workspace changes the conversation already performed.

Assistant vs. MCP

	In-app assistant	External MCP client
Interface	Remix Studio chat UI	Claude, Codex, or another MCP client
Authentication	Logged-in Remix Studio session	OAuth 2.0 or personal access token
Tool definitions	Shared registry	Shared registry
Read tools	Execute directly	Execute directly
Write confirmation	In-app confirmation; eligible write tools can be approved per conversation	Two-call preview and hash-confirmation protocol
Conversation storage	Stored by Remix Studio	Managed by the external client

Both paths enforce user ownership in each handler. IDs supplied by a model do not grant access to another user's records.

Related

• MCP Support — authentication and the external confirmation protocol.
• Libraries & Prompts — reusable data and assistant skills.
• Campaigns — campaign/post tools available to the assistant.