What the app does

Everything runs from a single installer. There is no server to set up and no MultipleChat account to create. Your chats, documents, the vector index and your API keys are stored in a folder on your own computer.

Core concepts

Explore the docs

Need help?

Email [email protected] and include your operating system and app version (shown at the top right of the app).

Five steps

1. Download and install. Get the .dmg for Mac or the .exe for Windows from the desktop app page. See macOS or Windows if you need help.
2. Pick a provider. On first launch, choose an AI provider. You can change it later in Settings → AI Providers → LLM.
3. Paste your key and choose a model. Cloud providers need an API key. Local runtimes (Ollama, LM Studio) need the runtime running with a downloaded model.
4. Create a workspace. Click New Workspace in the sidebar and name it after your project.
5. Send your first message. Optionally drag a PDF into the chat first, then ask about it.

What to try next

• Embed a folder of documents and ask specific questions: Chat with documents.
• Give the workspace a role and tone with a system prompt.
• Ask the model to research something live with @agent: AI agents.
• Press ⌘ / (Mac) or Ctrl / (Windows) to open the desktop assistant over any app.

System requirements

Current versions

The Mac and Windows releases are separate and features can differ between them. Use the 24-hour trial to check your workflow on your own computer. To see which version you have, look at the version number at the top right of the app.

Choose your platform

Installation problems

Local network access

By default the app's internal backend listens only on your own machine. Under Settings → Admin → General, Enable network discovery exposes the app's API (not its interface) to other devices on your local network. Restart the app fully after changing it.

Install

1. Download the .dmg installer from the desktop app page.
2. Open the .dmg and drag MultipleChat AI into your Applications folder.
3. Launch it from Applications. If macOS asks for confirmation the first time, choose Open.
4. Choose a provider, then start your trial or enter your license key.

If the app won't open

If macOS blocks the first launch, right-click (or Control-click) the app in Applications and choose Open once. This passes Gatekeeper's first-launch check; later launches work normally.

Permissions

• Microphone — only if you use voice input or the meeting assistant.
• Screen Recording — only if the desktop assistant should read the active window or a screen region.
• Accessibility — only for computer use, where the agent clicks and types for you.

Grant these in System Settings → Privacy & Security when prompted. You can revoke them at any time.

Install

1. Download the .exe installer from the desktop app page.
2. Run the installer and follow the prompts. Install for the current user (the default), not for all users.
3. If Microsoft Defender SmartScreen appears, choose More info → Run anyway.
4. Launch MultipleChat AI from the Start menu, choose a provider, then start your trial or enter your license key.

First start

The first launch can take a minute or two while the app prepares its data folder. The app keeps running in the system tray when you close its window — quit it from the tray icon.

Antivirus and firewall

Some security tools quarantine new desktop apps or block local network calls. If the app opens but never answers, allow it in your antivirus and make sure local connections to Ollama or LM Studio (on 127.0.0.1) are permitted.

How to update

Check your version

The version number is shown at the top right of the app. Mac and Windows have separate version numbers.

What's in an update

We ship updated builds regularly, merging improvements and security fixes from the upstream open-source project along with our own changes.

24-hour trial

The trial starts when you click Start trial, not when you install. Prepare an API key or a local model first so you can use the full window. Use the trial to check that the app works on your computer with your documents and providers: all license purchases are final.

Buy and activate

1. Buy a desktop license on the desktop app page ($250 USD, one-time).
2. Your license key arrives by email.
3. Open the app and enter the key under Settings → License. If the email doesn't arrive, check spam or contact support.

Online check-in

A licensed copy checks in online at least every 14 days. Between check-ins, local workflows keep working offline.

Good to know

• No recurring app subscription. Cloud providers bill your API usage directly.
• The desktop license does not include the MultipleChat web subscription, and vice versa.
• Lost your key? Email [email protected] from the address you bought with.
• One license is for one person and works on up to 3 of your computers at the same time, Mac or Windows.
• New or replacement computer? Email [email protected] from your purchase address and we reset your activations for free.
• All purchases are final; there are no refunds. See the License terms.

macOS

1. Quit the app and drag MultipleChat AI from Applications to the Trash.
2. To remove all data too, delete the app's folder in ~/Library/Application Support/.

Windows

1. Open Settings → Apps → Installed apps, find MultipleChat AI and choose Uninstall.
2. To remove all data too, delete the app's folder in %APPDATA%.

Reset without uninstalling

Quit the app and delete only the storage folder inside the app's data folder. On next launch the app starts as if freshly installed.

MCP tools you installed separately (for example with npx or uv) are not part of the app and stay on your system until you remove them.

Where to configure

Open Settings → AI Providers → LLM, select a provider, enter its details, pick a model and click Save. You can switch providers at any time; existing chats are kept.

Three levels of model settings

Provider types

Which key goes where

Use a key issued by the provider you select. A ChatGPT, Claude.ai or Gemini subscription login is not an API key, and an OpenRouter key won't work under the OpenAI or DeepSeek provider (or the other way round). API usage is billed by the provider, separately from your app license.

Model settings per workspace

Besides the model, each workspace's Chat Settings has a temperature (lower = more focused and repeatable, higher = more creative) and a chat history length — how many previous messages are sent with each new one (default 20). Local providers also let you set the model's context window and max tokens in the provider settings.

What leaves your computer

With a cloud provider, your prompt, the relevant document passages and the recent conversation go directly from your computer to that provider, under its privacy terms. There is no MultipleChat server in between. With a local model, nothing leaves your machine.

Connect a cloud provider

1. Create an API key in your provider's console. Add billing or credit if the provider requires it.
2. In Settings → AI Providers → LLM, choose the provider.
3. Paste the key. The model list loads once the key is accepted.
4. Pick a default model and click Save.

Supported providers

Keeping costs in check

• Set spending limits in your provider's console.
• Use a cheaper model as the system default and a stronger one only for the workspaces or agents that need it.
• Prefer embedding (RAG) over attaching very large files; attachments send the full text every turn.

At a glance

Connect OpenRouter

1. Create a key at openrouter.ai and add credit. Optionally set a spending limit on the key.
2. Open Settings → AI Providers → LLM and select OpenRouter.
3. Paste the key. In Chat Model Selection, models are grouped by organization (OpenAI, Anthropic, Qwen, Moonshot AI…). Pick one.
4. Click Save, create a workspace and send a short text message. Check usage in OpenRouter afterwards.

Finding model families

• Qwen — look under the Qwen organization. Pick a text-chat model; image models are not interchangeable with chat models.
• Kimi — listed under its developer, Moonshot AI.
• DeepSeek — available here too, or connect directly with the DeepSeek provider. Use the key that matches the provider you select.

If it doesn't connect

At a glance

Connect OpenAI

1. Create an API key on platform.openai.com and add billing.
2. Open Settings → AI Providers → LLM, select OpenAI and paste the key.
3. Choose a chat model available to your account and click Save.
4. Create a workspace and send a short message.

Also usable for

• Embeddings (Settings → AI Providers → Embedder → OpenAI).
• Text-to-speech, dictation and transcription.
• Image generation.

If it doesn't connect

At a glance

Connect Claude

1. Create a key in the Claude Console and add API credit.
2. Open Settings → AI Providers → LLM, select Anthropic and paste the key.
3. Choose a Claude model and click Save.
4. Create a workspace and send a short message.

Good to know

• A Claude.ai login or Pro subscription is not an API key, and your Claude.ai history isn't imported.
• Claude models are strong at agents and are the usual choice for computer use.
• Anthropic has no embedding model. Document indexing uses the built-in embedder (or another one you choose).

If it doesn't connect

At a glance

Connect Gemini

1. Create a key in Google AI Studio. Check its project, restrictions and quota.
2. Open Settings → AI Providers → LLM, select Gemini and paste the key.
3. Choose a supported Gemini chat model and click Save.
4. Test with a short text message before using files or images.

This doesn't sign you into the Gemini consumer app or import its history; a Google login can't replace the key.

If it doesn't connect

At a glance

Connect DeepSeek

1. Create a key on the DeepSeek platform and top up your balance.
2. Open Settings → AI Providers → LLM, select DeepSeek and paste the key.
3. Let Chat Model Selection load and choose a current model. Don't type an old model ID from a tutorial.
4. Click Save and send a short message.

If it doesn't connect

Ollama in short

1. Install Ollama from ollama.com and download a model, e.g. ollama pull llama3.1.
2. Make sure the server is running: ollama serve. (Running ollama run alone does not keep the server up.)
3. In Settings → AI Providers → LLM, choose Ollama. The app tries http://127.0.0.1:11434 automatically; enter the URL manually if it isn't found.
4. Select the model and click Save.

LM Studio in short

1. Install LM Studio, download a model and load it.
2. Start LM Studio's local server (Developer tab). The default address is http://localhost:1234/v1.
3. In Settings → AI Providers → LLM, choose LM Studio, confirm the base URL and pick the loaded model.

Choosing a model size

• 16 GB RAM handles most 7–8B models comfortably; larger models need more memory or a dedicated GPU.
• For agents, small local models often struggle with tool calling. A common setup is a local model for chat and a cloud model as the Agent LLM.
• Set the model's context window in the provider settings to match what the model actually supports.

MultipleChat AI is the chat app; Ollama or LM Studio runs the model. Keep the model runner open while you chat, and prepare it before starting your trial.

Set up Ollama

1. Install Ollama from ollama.com.
2. Download a model in a terminal, e.g. ollama pull llama3.1 or ollama pull qwen2.5. ollama list shows what you have.
3. Make sure the Ollama server is running (the Ollama app runs it, or start it with ollama serve).
4. In Settings → AI Providers → LLM, select Ollama. The app detects http://127.0.0.1:11434 automatically; otherwise enter the URL manually.
5. Choose the model, set its context window (see below) and click Save.

Context window

Set the token context window in the provider settings to what the model actually supports. Too small and the model forgets earlier messages or document passages; too large and it may run out of memory or slow down.

Local embeddings with Ollama

To keep document indexing on a model you choose, pull an embedding model (e.g. ollama pull nomic-embed-text) and select it in Settings → AI Providers → Embedder → Ollama. Re-embed your documents after switching.

Ollama on another computer

Point the base URL at that machine's IP, e.g. http://192.168.1.20:11434. On that machine, make Ollama listen on the network (set OLLAMA_HOST=0.0.0.0) and allow port 11434 through its firewall. Requests then leave your computer, but stay on your network.

If it doesn't connect

Set up LM Studio

1. Install LM Studio and download a model from its Discover tab.
2. Open the Developer tab, load the model and start the server. The default address is http://localhost:1234/v1.
3. In Settings → AI Providers → LLM, select LM Studio, confirm the base URL and pick the loaded model.
4. Set the context window to match the length you loaded the model with, then Save.

Embeddings with LM Studio

Load an embedding model in LM Studio and choose Settings → AI Providers → Embedder → LM Studio.

If it doesn't connect

Setup

1. In Settings → AI Providers → LLM, choose Generic OpenAI.
2. Enter the Base URL (usually ending in /v1) and the API key if the server requires one.
3. Type the exact model name the server expects.
4. Set the token context window and max tokens to the model's real limits, then click Save.

Common uses

• Team gateways such as LiteLLM that centralize keys and budgets.
• Self-hosted runtimes like vLLM, llama.cpp server or LocalAI.
• Corporate Azure or private deployments exposed through an OpenAI-style proxy.

Team gateway (e.g. LiteLLM)

Companies often run one gateway that holds the provider keys and hands each employee a personal “virtual” key with a budget. Everyone sets Generic OpenAI with the gateway URL and their own key: spend is controlled centrally, access can be revoked per person, and nobody needs raw provider keys.

Your messages

• Copy the text.
• Edit and resend. Everything below the edited message is removed and the chat continues from there.
• Speak — read aloud with text-to-speech.

Model replies

• Copy, or Edit the stored reply without resending (useful for correcting a fact the model should build on).
• Regenerate for a new answer to the same prompt.
• Thumbs up / down to rate answers. Ratings are stored locally with the chat log.
• Citations under a reply show which document passages were used.

The prompt box

Long conversations

When a conversation outgrows the model's context window, older messages are trimmed automatically. For a fresh start in the same workspace, open a new thread or type /reset.

Why workspaces

Documents and conversations in one workspace are invisible to the others. That keeps context clean and answers focused. A good pattern is one workspace per project or client — "Contracts", "Thesis", "Support macros" — rather than one giant chat.

Create a workspace

1. Click New Workspace in the sidebar and give it a name.
2. Open the workspace's gear icon to adjust its settings.

Workspace settings

Threads

Each workspace can hold several threads. Threads share the workspace's embedded documents and settings but keep separate chat histories — handy for exploring two directions on the same material. Attachments added in a thread stay in that thread.

The three modes

Change the mode

Open the workspace gear icon → Chat Settings → Chat mode, then save.

When you see an “@” in the prompt box

If an @ appears, start your message with @agent to use tools. If it doesn't appear, the workspace is already in Agent mode and every message can use tools.

“No relevant information found” in Query mode

1. Set Vector Database → Document similarity threshold to No restriction.
2. Rephrase using the document's own words (“initializing the application” rather than “starting the app”).
3. Make sure the document is actually embedded in this workspace.
4. For more explanation around the facts, switch to Chat mode.

Set a system prompt

Open the workspace gear icon → Chat Settings → System Prompt, write your instructions and save.

You are a senior contracts lawyer reviewing documents for a Swiss SME.
Answer in plain English, cite the clause number for every claim,
and end with a short "Risks" list. Today is {date}.

Variables

Variables in curly braces are replaced when the message is sent. Valid variables are highlighted in blue once you stop typing; unknown ones are left as plain text.

Ready-made prompts to copy

Tips

• Give a role, the audience and the output format. Short and specific beats long and vague.
• One workspace per role keeps prompts simple — a “Legal” workspace and a “Marketing” workspace instead of one prompt that does everything.
• If the model ignores the prompt, shorten it, put the most important rule first, or try a stronger model.

Custom variables

In Settings → Tools → System Prompt Variables, click Add Variable to define your own static values — a company name, a style guide link, a signature — and reuse them across workspaces.

Use a slash command

Type / in the prompt box (or click the slash icon) to see the list, then pick one. Its saved text is inserted into your message.

Create your own

1. Open the slash command menu and choose Add new.
2. Set the command (e.g. /tldr), the text it expands to, and a short description.
3. Save. It's available in every workspace.

Built-in commands

Start fresh

• New thread — click + New thread under the workspace. Same documents and settings, empty history.
• /reset — clears the history of the current chat so the next message starts clean.
• Edit an earlier message — resends from that point and removes everything after it.

Rename or delete a thread

Hover a thread in the sidebar and open its menu to Rename or Delete it. Deleting a workspace deletes all its threads and removes its documents from that workspace.

How much history the model sees

Each workspace's Chat Settings → Chat History sets how many previous messages are sent with every new one (default 20). Higher values give the model more memory of the conversation but use more tokens. When a conversation outgrows the model's context window, the oldest messages are dropped automatically.

Export chats

1. Open Settings → Workspace Chats. It lists every message with its workspace and time.
2. Click Export and choose a format.

You can also delete individual log entries there. Importing ChatGPT or Claude.ai history is not supported. To keep a single answer, use the copy button under it.

Turn it on

Personalization is off by default.

1. Open a workspace chat and click the settings icon at the top right.
2. Open the Memories panel and switch on Enable Personalization.

Workspace vs global memories

Add a one-sentence fact with +. Use a card's menu to edit, delete, or move it between workspace and global.

Automatic extraction

When personalization is on, the app can propose memories from your chats on its own. It runs in the background every few hours once a workspace has been idle for a while and has enough new conversation, keeps only confident facts, and merges duplicates. It needs a model with tool calling. You can switch automatic extraction off and add memories manually instead.

UI preferences

Open Settings → Customization → UI Preferences.

• Theme — Light, Dark or System.
• Color palettes — pick a built-in palette or create your own from a main and accent color with live preview. The palette also applies to the desktop assistant window.
• Display language.
• Your name — used in the home-screen greeting. Leave empty to hide the greeting.
• Default window — open the main window, the assistant, or nothing (start quietly in the background).

Chat preferences

Under Settings → Chat you can set the default font size and whether dictated speech is sent automatically.

Attach or embed?

If attachments exceed the context window you can Cancel, Embed them instead (recommended), or Continue anyway, which trims text and loses information.

Embed documents into a workspace

1. Open a workspace and click the upload / documents icon.
2. Drag in files — PDF, Word, text, Markdown, CSV, code and more — or point it at a folder.
3. Move the files into the workspace and click Save and Embed. Indexing runs on your machine.
4. Ask questions. Open the citations under a reply to see which passages were used.

Not sure what you can upload, or want to import a website or GitHub repo? See File types & sources.

Pinning

Click the pushpin next to an embedded document to send its full text with every message, bypassing search. Use it for a document that must always be in context — and only when it fits the model's window.

Tips for better answers

• Prefer text-based PDFs; run OCR on scans first.
• Ask specific questions (“What does §7.2 say about termination?”) rather than “summarize everything”.
• Search finds passages that resemble your question; it doesn't read the whole library. For tasks across every file (“count all invoices over €1,000”), use an agent.
• Documents that aren't embedded in the workspace are invisible to the model.

Supported file types

There's no fixed size limit for embedding. For attachments, the model's context window is the practical limit.

Scanned PDFs

Scans have no text layer, so results depend on OCR quality. For best answers, run OCR on the PDF yourself first (e.g. in Acrobat or Preview's export) and upload the searchable version.

Websites, GitHub, YouTube and more

In a workspace's document manager, open Data Connectors to import:

• Bulk link scraper — one or more web pages, optionally following links.
• GitHub repository — public repos directly; private repos need a personal access token.
• YouTube transcript — only works for videos that have a transcript.
• Confluence and other connectors, depending on your version.

Imported content lands in your documents like an upload; move it into the workspace and click Save and Embed. Some sites block scraping. For pages behind a login, use the private browser tool.

Folders

You can upload a folder, but only its top level is read. Upload subfolders separately, or give an agent access to the folder through an MCP filesystem server.

Keeping documents up to date

Embedded documents are snapshots. If you edit the original, re-upload it — or enable the Live document sync feature preview and click the eye icon next to an embedded file. The app then checks local files about every 10 minutes while it's open. Individual files only; moved files show as “Not Found”.

Remove or delete a document

• Remove from a workspace — in the workspace's document manager, remove it from the workspace list. The file stays in your documents and can be re-added quickly.
• Delete completely — delete it from My Documents. It's removed from every workspace and its cached data is cleared.

Open the workspace gear icon → Vector Database.

Troubleshooting order

1. Lower the similarity threshold.
2. Increase snippets or turn on accuracy-optimized search.
3. Pin the document as a last resort.

Embedding model

The built-in embedder runs locally with no setup. In Settings → AI Providers → Embedder you can switch to Ollama, LM Studio, LocalAI, OpenAI, Azure OpenAI or Cohere.

Text splitting & chunking

In Settings → AI Providers → Text Splitter & Chunking:

• Text chunk size — default 1,000 characters (about 250 tokens), capped at the embedder's maximum shown under the field.
• Text chunk overlap — default 20 characters; must be smaller than the chunk size.

The splitter breaks text at paragraphs first, then lines, then spaces. Smaller chunks suit dense reference material; larger chunks suit narrative text. Change one setting at a time. New settings apply only to documents embedded afterwards — remove and re-embed existing files to apply them.

Vector database

Your index is stored locally in LanceDB by default — nothing to configure. In Settings → AI Providers → Vector Database you can switch to Chroma, Milvus, Pinecone, Qdrant, Weaviate, Zilliz, AstraDB or pgvector.

Cloud databases store your document passages with that vendor; the default keeps them on your computer.

Start an agent

Start a message with @agent (or set the workspace to Agent mode). The model plans the steps and picks tools on its own.

@agent find the three most recent EU AI Act summaries online, compare them
with the draft PDF in this workspace, and save the differences as a file.

Type /exit to end the agent session.

Pick the right model

Agents depend on the model producing valid tool calls. Set the Agent LLM in the workspace's agent configuration to a capable tool-calling model — most current frontier models from OpenAI, Anthropic and Google qualify. Small local models are the most common cause of agents that do nothing.

Configure skills

Enable, disable and configure skills in Settings → Agent Skills, then click Save. Inside a chat, the Tools button toggles skills for that conversation. See Built-in skills.

Approving risky actions

Skills that change things on your computer or send data out (for example writing files or calling external services) can ask for your approval before they run. Review the request in the chat and approve or reject it; unanswered requests are rejected after about two minutes.

Intelligent tool selection

With many skills and MCP servers enabled, every message carries a long tool list. In Settings → Agent Skills, open the settings icon and enable Intelligent Tool Selection so only the tools relevant to each request are sent — fewer tokens and fewer wrong tool picks.

Some skills work immediately; others need setup (a database connection, allowed folders). Click a skill in Settings → Agent Skills for its options. Beyond these, add tools through MCP servers or build no-code agent flows.

Where MCP servers are configured

MCP servers are listed in a JSON file in the app's data folder at storage/plugins/anythingllm_mcp_servers.json. The file is created the first time you open Settings → Agent Skills, and starts empty. Only MCP tools are supported (not resources, prompts or sampling).

Add a server

1. Install the server's runtime yourself (Node for npx, uv/uvx for Python servers). The app doesn't install them, and the command must be on your PATH or given as a full path.
2. Add the server to the mcpServers object in the JSON file.
3. In Agent Skills, click Refresh. Servers reload without restarting the app.
4. Click the server to check its status and the tools it exposes, then use @agent.

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/me/Projects"]
    },
    "remote-tools": {
      "type": "streamable",
      "url": "https://mcp.example.com/mcp",
      "headers": { "Authorization": "Bearer YOUR_TOKEN" }
    },
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "YOUR_TOKEN" },
      "anythingllm": { "autoStart": false }
    }
  }
}

Transports & options

Manage servers

• Servers start when you open Agent Skills or first use @agent, then keep running in the background. Many servers slow startup.
• The gear icon on a server lets you Start, Stop or Delete it. Delete removes it from the file and stops the process.
• Errors appear on the server's card; check the logs for details.

Turn it on

1. Open Settings → Agent Skills and enable Web Search (web browsing). Enable Web Scraping too if you want the agent to read full pages from a URL.
2. Under the skill, open Search Provider. A free engine that needs no key is used by default.
3. Click Save, then ask with @agent, e.g. “@agent search the web for this week's ECB rate decision and summarize it”.

Search providers

The exact list depends on your app version. Keyless engines are rate-limited and sometimes return thin results; a keyed provider is more reliable for daily use.

If search doesn't work

• Make sure you're in an agent session (@agent or Agent mode) — plain chat doesn't search.
• Check the provider key, or switch providers if the free engine is rate-limiting you.
• Ask concrete questions with dates, names or sites; vague questions get vague results.

Set it up

1. Open Settings → Tools → Browser Tool and click Open Private Browser.
2. Log in to the sites you want the agent to read. Sessions are kept across restarts until they expire.
3. Ask @agent to read or summarize a page from that site.

What it can and can't do

• Reads page text only — no images or video, and it doesn't click or fill in forms.
• The browser is isolated from your normal browser; logins are stored only on your computer.
• Some services detect automated access and may block accounts. Check the site's terms.
• Clear Browser Data removes all stored sessions.

What a flow is

A flow is a chain of blocks the agent can run as a single skill — for example “scrape this page, summarize it with these instructions, save it to a file”. Once saved, the agent calls it when your request matches its description, and capable models can chain several flows.

Create a flow

1. Go to Settings → Agent Skills and click Create Flow.
2. Fill in Flow Information: a clear name and a description of when to use it — the agent reads this.
3. Define inputs in Flow Variables.
4. Use Add Block to insert steps between the start and Flow Complete.
5. Click Save. New flows are enabled automatically.

Available blocks

Edit or delete a flow from the gear icon next to it in Agent Skills.

Create a job

1. Open Settings → Scheduled Jobs and create a new job.
2. Give it a name and the prompt the agent should run.
3. Set the schedule with the visual builder, or type a cron expression.
4. Choose which tools it may use: skills, flows and MCP servers.

You can also enable the Create scheduled jobs skill and simply tell @agent what to run and when; it returns an editable job card.

Runs and results

• The list shows each job's next run; you can enable, disable, edit, delete or Run now.
• Each run records its status (queued, running, completed, failed, timed out), the agent's steps and tool calls, any generated files, and the final answer.
• Continue in thread opens a result in a dedicated “Scheduled Jobs” workspace so you can follow up.

Open it

Press ⌘ / on macOS or Ctrl / on Windows from anywhere. The shortcut can be changed in its settings.

What it can do

• Use the active window, or a screen region you select, as context for your question.
• Everything the main chat can: drag in documents, attach images, use @agent with skills and MCP tools, run slash commands.
• Use its own system prompt and embedded documents.

Settings

Open the three-dot menu in the assistant → Desktop Assistant Settings, or Settings → Desktop Assistant in the main app.

• Open thread behavior — start a new thread each time (default) or continue the previous one.

All assistant threads are saved to the Assistant Chat workspace, so you can pick them up in the main window.

Requirements

• An agent model that supports computer use, such as Anthropic's Claude.
• On macOS, Accessibility and Screen Recording permissions for the app.

Stay in control

• Pause or abort at any time from the app, by closing it, or with ⌘ K / Ctrl K.
• Give it only tasks you would give a screen-sharing assistant. Watch the first runs.
• Avoid tasks involving payments, passwords or irreversible deletions.

Text-to-speech

Settings → Voice & Speech → Text-to-speech. Click the speaker icon under any reply to hear it.

Voice input

Configure dictation in Settings → Voice & Speech: system native (default), OpenAI, Deepgram, Groq, or an OpenAI-compatible endpoint. Dictated text is sent automatically; turn off Auto-submit speech input in Settings → Chat to review it first.

File transcription

Audio and video you upload are transcribed before embedding. Settings → Transcription uses a built-in local Whisper model by default (downloaded on first use), or OpenAI / an OpenAI-compatible service.

How it works

• Captures audio from your computer and transcribes it with a local Whisper model.
• Produces a summary and action items after the meeting.
• Transcripts can be added to a workspace so you can chat with them later.

Before your first meeting

• Allow microphone access (and system audio capture where your OS asks).
• Tell participants you're recording, and follow the consent rules that apply where you and they are.

Ask about an image

Drag, paste or attach an image in the prompt box and ask about it — a screenshot, a chart, a photo of a whiteboard. This needs a vision-capable chat model in that workspace (most current GPT, Claude and Gemini models; for local use, a vision model such as Llama 3.2 Vision or Qwen-VL in Ollama or LM Studio).

Set up image generation

1. Open Settings → AI Providers → Image Generation.
2. Choose a provider (e.g. OpenAI or OpenRouter, or a local option such as Ollama where available), add the key and pick an image model.
3. Save.

Generate an image

In any chat, type /img followed by your prompt:

/img a minimal flat illustration of a desk lamp, warm light, beige background

To refine a result, open the image's ⋯ menu → Edit and describe the change (not every provider supports editing).

Prompt tips

• Name the subject, style, composition and mood: “isometric 3D, soft shadows, top-down, cheerful”.
• Say what to leave out: “no text, no people”.
• Image generation is billed by the provider per image and is usually pricier than text.

Set it up

1. Open the workspace's Chat Settings and choose Model Router as the LLM provider.
2. Add rules. Each rule picks a model when it matches.
3. Set a fallback model for messages no rule matches, and save.

Rule types

• Keywords — e.g. “code”, “contract”.
• Message length — long prompts to a long-context model.
• Time of day.
• Has an image — send to a vision model.
• Let a model decide — a small model classifies the request first.

Typical setup: a local model for everyday questions and a cloud model for coding and long documents.

Create an API key

Open Settings → Tools → Developer API and generate a key. Keep it secret; delete keys you no longer use.

Endpoints

The app serves interactive API documentation (Swagger) at /api/docs on its local address. From there you can try every endpoint: workspaces, documents, embedding, threads and chat.

Calling from another device

By default the API is reachable only from your computer. To call it from another device on your network, enable network discovery (see Installation overview) and restart the app. Never expose it to the internet.

Data folder

The quickest way there: click the version number at the top right of the app and choose Open storage.

What's inside storage

What leaves your computer

• Messages go directly from your computer to the provider you configured, under that provider's terms. There is no MultipleChat server in between.
• We never see, collect, store or train on your chats or documents.
• Telemetry is disabled in our builds; the app sends no usage analytics.
• The only call to MultipleChat is the periodic license check-in.

Back up or move to a new computer

Quit the app and copy the whole data folder. On the new computer, install the app, quit it, and replace its data folder with your copy. Then activate your license on the new computer. If all 3 activations are in use, email [email protected] from your purchase address and we reset them.

“Invalid API key”

Re-paste the key without spaces, and check in the provider's console that it's active and funded where credit is required.

“Could not respond to message” or “fetch failed” in chat

The app couldn't reach the model or it took too long. Check that Ollama or LM Studio is running, that the model fits your memory, and that the context window isn't set higher than the model supports. For cloud models, check the key, credit and your connection. The logs show the real error.

Answers ignore my documents

• Confirm the files are embedded in this workspace.
• Lower the similarity threshold and ask a more specific question (see Retrieval settings).
• For questions across every file, use an agent rather than plain RAG.

Agent does nothing or calls the wrong tool

This is almost always the model. Agents need valid, schema-matching tool calls.

• Set the Agent LLM to the strongest tool-calling model you can — a cloud model for agents and a local one for chat works well.
• Prefer general instruct models; larger or higher-quantization local models are more reliable.
• Disable skills and MCP servers you don't need, or turn on Intelligent Tool Selection.
• Type /reset and ask again.

Can't connect to Ollama

• Open http://127.0.0.1:11434 in a browser; it should say Ollama is running.
• Start the server with ollama serve.
• For Ollama on another machine, use that machine's IP and allow port 11434 through its firewall.

“Fetch failed” or “documents failed to add” when embedding

The built-in embedder downloads its model the first time you embed. If a firewall or proxy blocks huggingface.co or cdn.anythingllm.com, embedding fails.

1. Allow those hosts (see Network), restart the app and test with a small .txt file in a new workspace.
2. If it still fails, quit the app, delete the models/Xenova folder inside the data folder and try again so it re-downloads.
3. Windows: install the latest Microsoft Visual C++ Redistributable (x64).
4. The default vector database needs a CPU with AVX2. Some virtual machines hide it.

“Document processor unavailable”

The app's internal document service can't be reached on your computer. A VPN, firewall or security tool blocking local connections, or another program using port 8888, are the usual causes. Allow the app, free the port, restart, and check the logs. Long audio or video transcriptions can also keep it busy for a while.

MCP server won't start

On macOS, apps opened from Finder don't inherit your terminal's PATH, so npx or uvx may not be found (“ENOENT”, “Connection closed”). Use full paths from which npx in the config, then click Refresh. On Windows, use the full path from where npx.

App won't open after download (macOS)

Right-click the app → Open once to pass Gatekeeper's first-launch check.

Blank or white window on start (Windows)

• The first start can take a minute or two — give it time.
• Install for the current user only; installs for all users aren't supported.
• Run the app from PowerShell to see the error (see logs).
• Last resort: back up your data folder, uninstall, delete %APPDATA%'s app folder, and reinstall.

Update fails on Windows (“app cannot be closed”)

The app keeps running in the system tray. Quit it from the tray icon, then click Retry in the installer. If an older install for all users exists, uninstall it first; your data is kept.

The model shows long “thinking” text

Reasoning models (DeepSeek-R1, Qwen3 and others) produce it themselves. Use a non-reasoning variant or turn thinking off in the provider if possible.

App opens but won't accept input

Another program may be using the app's internal port (3001) — for example a Docker container. Stop it, or move it to another port, then restart the app.

Collect logs

• Click the version number at the top right → Open logs.
• For live output, start the app from Terminal (macOS) or PowerShell (Windows) by running its executable directly; logs print to the console.

What the app connects to

Proxy

The app has no proxy setting of its own. On networks that require a proxy, use a system-level proxy or VPN, and ask IT to allow the hosts above. Blocked downloads usually show up as “fetch failed” when embedding a document.

Local connections must be allowed

The app runs small internal services on your own computer. Some VPNs, firewalls and security tools block these, causing “Document processor unavailable” or a window that won't respond. Allow the app in your firewall, and make sure no other program uses ports 3001 or 8888 (for example a Docker container).

Buying & setup

Which computers are supported?

Macs with Apple Silicon (M1 or newer) on macOS 13 Ventura or later, and 64-bit Windows 10 or 11. Intel Macs and Linux aren't supported by the current downloads. The Mac app is version 2.0.0 and the Windows app is version 1.0; features can differ between them.

What does “bring your own API key” mean?

You create an API key with a provider (OpenAI, Anthropic, Google, OpenRouter, DeepSeek…) and paste it into Settings → AI Providers → LLM. The app talks to that provider directly with your key, and the provider bills your usage. See AI providers.

Does my ChatGPT Plus, Claude Pro or Gemini subscription work?

No. Consumer subscriptions don't include API access. Create an API key with the provider (pay per use), or use a local model.

How much does it cost?

A 24-hour trial, then a one-time $250 license. Cloud API usage is billed separately by your provider; local models cost nothing per token. See Trial & license.

Does the desktop license include the MultipleChat web subscription?

No. The desktop app is a separate, single-user product. Web-only features such as the Studios and AI Collaboration are part of the web subscription.

How do I activate my license?

The key is emailed to the address you used at checkout. Enter it under Settings → License. If it doesn't arrive, check spam or email [email protected].

Can several people share one install?

No. It's a single-user app; each person installs it on their own computer.

On how many computers can I use my license?

On up to 3 of your own computers at the same time, in any mix of Mac and Windows, at home or at work. The license is for you personally, not for colleagues.

How do I move my license to a new computer?

Install the app on the new computer and enter your key. If all 3 activations are already used, email [email protected] from the address you bought with, and we reset your activations for free. To bring your chats and documents along, see Back up or move to a new computer.

Can I get a refund?

No. All license purchases are final and we don't give refunds for any reason. Please use the free 24-hour trial to check the app on your own computer before you buy. Details are in the License terms and section 36 of our Refund Policy.

Models

Can I use it offline or without an API key?

Yes. Connect Ollama or LM Studio and keep the built-in embedder and default vector database, and chats and document search run on your computer. You need the internet for first-time downloads, web search, cloud models, and the license check-in at least every 14 days.

Will it use my GPU?

Cloud models run on the provider's hardware, so no GPU is needed. Local models run in Ollama or LM Studio, which use your GPU automatically — Metal on Apple Silicon, CUDA on NVIDIA. If replies are slow, the model is probably too large for your memory; try a smaller one. 16 GB RAM or more is recommended for local models.

Can I use different models in different workspaces — or only for agents?

Yes. Set a workspace model in the workspace's Chat Settings (⌘/Ctrl Shift L) and an agent model in its Agent Configuration. The model router can even pick a model per message. The embedder is one setting for the whole app.

How do I use a GGUF model I downloaded?

Load it in LM Studio (drop it into its models folder), or import it into Ollama with a Modelfile (ollama create mymodel -f Modelfile). Then select it in the app as usual.

How do I hide the “thinking” of reasoning models (DeepSeek-R1, Qwen3)?

The thinking output comes from the model itself; the app has no switch for it. Use a non-reasoning variant of the model, or turn thinking off in the provider if it offers that.

Can I send images to the model?

Yes, with a vision-capable model. Drag or paste the image into the prompt box. See Images.

Can it create images?

Yes. Configure Settings → AI Providers → Image Generation, then type /img and your prompt.

Documents

Why doesn't the AI use my document?

Uploading only converts a file; the model sees it once it's embedded in this workspace or attached to the chat. With embedding, only the 4–6 most relevant passages are sent per question. Try a lower similarity threshold, a more specific question, accuracy-optimized search, or pinning. See Retrieval settings.

Why can't it count or compare across all my files?

Document search sends a handful of matching passages, not the whole library. For questions across every file, use an agent or attach the specific files.

Which file types can I upload, and how big?

PDF, Word, PowerPoint, Excel, text, Markdown, CSV, HTML, EPUB, images (OCR) and audio/video (transcribed), among others. No fixed size limit for embedding. Full list: File types & sources.

My scanned PDF gives no answers.

Scans have no text layer. Run OCR first and upload the searchable PDF.

Can I add websites, GitHub repos or YouTube videos?

Yes, through Data Connectors in the document manager. See File types & sources.

Can I upload a folder with subfolders?

Only the top level of a folder is read. Upload subfolders separately.

I edited my document. Does the AI see the change?

Not automatically. Re-upload it, or enable the Live document sync preview.

How do I remove a document?

Remove it from the workspace in the document manager, or delete it from My Documents to remove it everywhere. See Remove or delete a document.

I changed the embedder and now search is broken.

Vectors from different embedders don't mix. Remove and re-embed your documents (or create a new workspace). With Ollama as embedder, choose an embedding model such as nomic-embed-text, not a chat model.

Agents

Do I have to type @agent?

Only if you see an @ in the prompt box. In Agent chat mode, every message can use tools.

The agent doesn't use tools or makes things up.

Almost always the model. Use a strong tool-calling model as the agent model, disable skills you don't need, and type /reset before asking again. See Troubleshooting.

My MCP server fails to start on macOS.

Apps opened from Finder don't get your terminal's PATH. Use the full path to npx, uvx or node (find it with which npx) in the MCP config. See MCP integration.

Can the agent read pages I'm logged into?

Yes, with the private browser tool.

Data & privacy

Where are my chats and API keys stored?

On your computer, in the app's data folder: chats, documents, the vector index and keys. See Where your data lives.

Does MultipleChat see my chats?

No. With a cloud model, your prompt and relevant document passages go straight to that provider under its terms. Nothing passes through a MultipleChat server, and telemetry is disabled.

How do I back up or move to a new computer?

Quit the app and copy the whole data folder. On the new computer, install, quit, and replace its data folder with your copy.

Can I move the data folder to another drive?

Not supported — moved or linked folders can break updates and the vector database. To save space, keep large local models on another drive in Ollama or LM Studio.

How do I export my chats?

Settings → Workspace Chats → Export (CSV or JSON). See Chat history & export.

Can it remember things about me?

Yes, if you turn on Memories. Don't store secrets there — memories are sent to your model.

The app

How do I update?

Install the new version over the old one; your data is kept. On Windows, quit the app from the system tray first. See Updating.

Can I make the text bigger?

Use the font size control in the prompt box, or the usual zoom shortcuts ⌘/Ctrl + and −. Theme and language are in Appearance.

Does it work behind a corporate proxy?

The app has no proxy setting of its own; use a system proxy or VPN and allow the hosts listed in Network, proxy & firewall.

The app uses a lot of CPU or RAM.

That's nearly always a local model. Pick one that fits your memory, close other heavy apps, or use a cloud model. Very long threads can slow the window — start a new thread.

Contact support

Email [email protected] with your operating system, app version (top right of the app) and, if possible, the relevant lines from the logs.