Features

This page describes what MultiGPT version 1 is meant to do, as of version 0.4.0. Done: chatting with many providers and LM Studio including model management, attachments, projects, shared chats, the context menu in the chat, family accounts with billing accounts and budgets, comparison mode, MCP tools, MultiGPT as an MCP server, calculations, PDF sheets, your own documents with citations, web search, image generation and the runner (new in 0.4.0). Image editing, voice and the actual scratchpad are planned. Every section shows its status and milestone.

Detailed guides are in the wiki (German).

Login and accounts

Done Milestone 1

• Log in, log out and change your password.
• Login throttling: after five failed attempts, logging in with that user name from that address is blocked for 15 minutes.
• No self-registration. An administrator creates accounts – with mgpt-ctl createsuperuser, make user (with a choice of role) or in the admin area.
• Personal menu (since 0.3.5): the profile picture and name sit in the top right. A click opens settings, profile picture, my usage, API keys, family, administration, password and sign out – depending on your permissions. Without a profile picture your initials are shown; an uploaded picture is stored without metadata.
• The version is shown at the bottom of the sidebar, on the sign-in page and in the administration header (since 0.3.6).

Chatting with many providers

Done Milestone 3 / 4 / 5

• Chat list in the sidebar: new, rename, archive, delete, search. The sidebar can be hidden, answers use the full width, and the font size can be changed with A−/A+.
• The model can be chosen per message. Answers are streamed, a round button in the input field sends or stops, and an answer can be regenerated.
• Providers are connected via API key: OpenAI and all OpenAI-compatible services (e.g. OpenRouter and LM Studio), plus Anthropic and Google Gemini.
• In the admin area, “check connection now” tests a provider and names the cause of a problem, e.g. “connection refused”, “API key expired” or “check the base URL”. Unreachable cloud providers are greyed out in the chat.
• Your own messages can be edited as in ChatGPT, attachments included. Editing and “regenerate” create versions, and “‹ 1/2 ›” switches between them. Earlier versions are kept. Every message has a copy button.
• Markdown with code highlighting and a copy button. The libraries are bundled locally and the output is sanitised.
• Automatic chat titles, a system prompt per chat, export as Markdown.
• Input history (since 0.3.4): arrow up and down bring earlier messages back into the input field.
• Delete an answer with what follows (since 0.3.4): an answer can be deleted together with its question and everything after it. Costs stay booked.
• While the AI is still writing you can scroll freely: the view only follows while you are at the bottom; otherwise “↓ Neueste” (newest) takes you back to the end.
• With Anthropic, MultiGPT uses prompt caching for the system prompt, tools and history (since 0.3.4). This saves costs in longer chats.

Context menu and quoting in the chat

Done Milestone 13

Since 0.3.2: reuse selected text in answers, your own messages, comparison columns and source lists – with a right-click, via a bar below the selection (on phones too) or with Shift+F10.

• Quoting: the selected text goes into the input field as a quote, with its origin (the model, “Ich” for yourself or “Quelle [n]” for a source). Several quotes go into one message, one after the other.
• Quick actions: ask a follow-up, explain, simplify, translate, counter-arguments, find evidence and copy with a source reference.
• Nothing is sent on its own. Shift+right-click still opens the browser’s menu, and the custom menu can be switched off in the settings.
• Guide: wiki: Chat-Bedienung.

Attachments and image input

Done Milestone 5

• Attach images and documents with the paper clip, paste them from the clipboard or drag them into the input field. Images open large with a click (lightbox).
• Images only go to models with the “understands images” tick, across all three provider adapters. Documents (PDF, DOCX, TXT, MD) are passed on as text.
• Privacy: every image is re-encoded on the server. This drops EXIF data such as the location (GPS), camera data and embedded comments. The file type is checked by content, not by extension.
• Attachments are visible only in the chat they belong to and are served only after a permission check.

Formulas and SVG preview

Done Milestone 5

• Mathematical formulas are typeset with KaTeX. KaTeX lives on your own server; nothing is loaded from the internet.
• SVG code in answers gets a preview, with a switch to the code and “save as file”. The preview is just an image: scripts, links and external content in the SVG have no effect.

Projects

Done Milestone 5

Group chats into projects, as in ChatGPT or Claude – for example “Moving house” or “Tax return”.

• A project gives its chats instructions, a default model and preselected collections of your own documents.
• Projects appear in the sidebar above the chats; they can be expanded, pinned and archived. Search also finds project names.
• Only the person who created a project can see it.
• Guide: wiki: Projekte.

Sharing chats

Done Milestone 5

• Share a chat with individual accounts or a group. Each share has RWUD permissions: read (always), write (join the conversation), update (edit messages, rename) and delete.
• Recipients find shared chats under “Mit mir geteilt” (shared with me) and can continue them as their own copy.
• Revoking takes effect at once, even for an answer that is still running.
• Costs are borne by whoever triggers an answer, not by the chat’s owner. Which models someone may choose depends on their own role and budget.
• Notices (since 0.3.6): if a recipient deletes or archives the chat or deletes an answer, the owner gets a notice in the personal menu – with the name and chat title, without content.
• Guide: wiki: Chats teilen.

Local models with LM Studio

Done Milestone 4

LM Studio runs on a PC in the home network and provides an OpenAI-compatible server. MultiGPT connects to it as a provider but does not start or wake it.

• A status indicator in the header shows whether LM Studio is reachable. The browser checks every 30 seconds while the tab is visible.
• When LM Studio is off, local models are greyed out. When it is on, the models that LM Studio actually reports are offered, along with their capabilities.
• If LM Studio goes away in the middle of an answer, the text received so far is kept and the cause is shown – with a dedicated hint when the model’s context is too small.
• Local models cost € 0 and stay usable even when the budget is used up.
• In real use in the home network: chat (e.g. gpt-oss-20b), embeddings (nomic-embed-text) and OCR (olmOCR) run through LM Studio.

Model management (since 0.3.2, for LM Studio 0.4 or later): if you like, MultiGPT itself decides which models are loaded in LM Studio.

• Before a chat, a comparison, embeddings, OCR or figure descriptions, MultiGPT loads the right model with its context window and unloads others to make room – never beyond a memory budget for the graphics card, keeping a reserve for the display and system.
• Pinned models and models in use stay loaded, and indexing does not push out the chat model. If something does not fit, the chat says “not enough graphics memory” and an indexing run waits.
• Indexing models stay loaded only for the run (since 0.3.4). Models unused for 30 minutes are unloaded by the worker.
• In the admin area you can choose the context window per model, with the expected memory shown next to it; a separate page lists the loaded models. On the console: mgpt-ctl lms.
• If several MultiGPT installations share one LM Studio, each unloads only the models it loaded itself (since 0.3.6).
• Guide: wiki: Anbieter und Modelle.

Models and capabilities

Done Milestone 4 / 4a

The capability matrix in the admin area lists every provider’s models with their capabilities, editable in place:

• Main type: chat, image generation, embedding, speech recognition, speech output or music.
• Ticks: tools, understands images, edits images.
• MCP: none, all or selected MCP servers (see below).

New models get their capabilities automatically: LM Studio reports them itself, otherwise MultiGPT infers them from the model ID. For existing models, sudo mgpt-ctl guess_capabilities --apply catches up. Guide: wiki: Anbieter und Modelle.

Roles, groups and budgets

Done Milestone 2 / 6

Four default roles, adjustable in the admin area:

• Permissions are checked on the server, for every page and before every provider call – not just by hiding things in the interface. There are separate permissions for image generation, calculations and PDF documents, among others.
• Privacy: chats are private. Administrators do not see other people’s chats either, only usage figures. The one exception: teenager accounts have an option “Einsicht in Chats erlaubt” (chat insight allowed), off by default. When it is on, administrators can read that account’s chats but not write in them, and the member permanently sees a notice.
• Groups: collections and individual chats can be shared with groups.
• A “Familie” (family) page (/familie/) for administrators: create and lock accounts, assign roles, reset passwords, set budgets, switch chat insight, manage groups and view everyone’s usage.

Billing accounts and costs

Done Milestone 6

Every provider bills through a billing account:

• monetary in EUR or USD (cloud providers), tokens only (local models such as LM Studio) or flat rate (subscription; requests and tokens are only counted).
• Prices with validity dates: per model from a given date, with separate prices for cache, long context and reasoning tokens. Image prices per size and quality. MultiGPT ships no prices; the administrator maintains them.
• Exchange rates for USD accounts; every answer is recorded as a booking with tokens, amount and rate. Since 0.3.6 the worker fetches the ECB reference rate USD→EUR automatically on working days (can be switched off), and EUR accounts can be switched to USD; existing bookings stay unchanged.
• Budgets per account, role and person (person before role): in euros, or as a quota for token accounts. From 80 % a warning appears; from 100 % the account’s models are blocked until the next month. Free models stay usable.
• Mein Verbrauch (my usage, /verbrauch/) shows everyone their tokens and costs per account and model, their budget status and the last few months.
• Guide: wiki: Kosten und Budgets.

Comparison mode

Done Milestone 6

Ask two or three models the same question at the same time and read the answers side by side.

• A switch in the input field opens the choice of two or three models.
• The answers appear as columns and are stored in the chat tree as versions of the same message. Later, “‹ 1/3 ›” switches between them.
• Picking one answer makes its model the chat’s default model.
• MCP tools are off in comparison mode, so no confirmations pop up in several columns at once.

Tools via MCP

Done Milestone 4a

MultiGPT is a client for the Model Context Protocol. Models that support tools can call functions from connected MCP servers.

• Only administrators add MCP servers – local (stdio) or via Streamable HTTP. Credentials are stored encrypted.
• Import: configurations in the {"mcpServers": {…}} format (as used by Claude Desktop, Cursor or n8n) can be pasted in. Entries with placeholders instead of real credentials are created disabled.
• Online status: every server shows “online”, “offline” with the cause, “unchecked” or “disabled”; the worker checks regularly.
• Classification: a tool table lists all of a server’s tools. Tools that change something or send data out only run after a confirmation in the chat. New, unclassified tools require confirmation.
• Access per model: the administrator decides which models may use MCP (none, all, selected servers). This protects against models that are more easily tricked into unwanted calls. It is checked on the server, in addition to the role’s permission.
• Every call appears in the chat as an expandable line with arguments, result and duration. At most 10 tool rounds per answer, with a time limit per call.
• Tools on demand (since 0.3.2): with many tools – n8n with more than 50, for example – the model gets only a core set and looks up more with find_tools. MultiGPT also estimates the request before sending, drops older messages if needed and reports a context that is too small with figures.
• Administrators see MCP servers going offline live in the status bar (since 0.3.5).
• Guide: wiki: MCP.

MultiGPT as an MCP server

Done Milestone 15

Since 0.3.2 MultiGPT is an MCP server itself: at https://<hostname>/mcp/, other programs such as n8n, Claude Desktop or your own agents control MultiGPT with an account’s API key.

• Tools: ask models (ask), upload, search and read documents, start, follow, pause and cancel indexing, read in directory sources, fetch generated files and call tools such as create_pdf or run_python directly. Results come back directly as text, image or PDF (since 0.3.3).
• Reading chats (since 0.3.6): with the “read chats” permission, a key lists and reads the account’s own chats and chats shared with it – never anyone else’s, not even for administrators. Chats created by ask go into an “API” project by default.
• Permissions: a key can never do more than its account. Every member creates keys under settings → API keys, as far as the role allows (on for administrators and adults). Keys are stored only as a hash, can be revoked and optionally expire.
• Every call is recorded in the audit log without content. For older clients there is also the HTTP+SSE transport at /mcp/sse.
• On the console: mgpt-ctl apikey and mgpt-ctl index status --follow.
• Guides: wiki: API-Keys and wiki: n8n.

Calculations in a sandbox

Done Milestone 4a

Language models easily miscalculate. Models with tool support therefore get the built-in run_python tool and calculate exactly.

• numpy (numerics, statistics), scipy (since 0.3.6), sympy (equations, derivatives, integrals, exact fractions), mpmath (high precision) and the standard library.
• Charts made with matplotlib appear as PNG or SVG in the answer; SVG is sanitised on the server.
• Code and output can be expanded in the chat.
• Security: the code is treated as untrusted and runs in a bubblewrap sandbox – no network, no server files, no keys or environment, with limits on time, memory and processes. Without a working sandbox the tool is not offered at all.
• A separate role permission “Berechnungen ausführen” (run calculations; on for administrators, adults, teenagers). Guide: wiki: Berechnungen.

Worksheets and documents as PDF

Done Milestone 4a

Since 0.3.1: ask for a sheet and you get a print-ready PDF (the create_pdf tool) that opens in the chat with “Ansehen” (view).

• Handwriting rulings for learning to write: rulings 0 to 4 with the “little house” guide, lined, squared 5 and 10 mm, dimensionally accurate in millimetres.
• Maths: arithmetic tasks with an answer sheet, digit boxes, times tables, number lines, clock faces. The server generates tasks and answers, not the model – so the answers are always right.
• Science and general worksheets with gap texts, tick boxes and pictures, plus letters, invitations and tables.
• When printing choose “actual size” or 100 % so the rulings stay accurate.
• Security: the PDF is rendered in a separate process without network or file access; only your own attachments may be embedded as images.
• A separate role permission “Dokumente erzeugen (PDF)” (create documents; on for administrators, adults, teenagers). Guide: wiki: Dokumente erzeugen.

Ask your own documents (RAG)

Done Milestone 7

• Create collections and upload documents (PDF, DOCX, TXT, MD and image files, up to 25 MB per file), private or shared with groups – read-only or with write access.
• A background service (worker) reads the text, recognises scanned pages via OCR, splits the text into passages and computes embeddings. Running jobs can be cancelled.
• Several documents at once (since 0.3.5, default 2) in separate worker processes. Runs can be paused and resumed (since 0.3.6), and uploads are processed before large runs.
• In the chat you pick collections. Search uses PostgreSQL and pgvector, optionally combined with German full-text search.
• The documents used are listed below the answer with their locators. A PDF can be viewed in the browser and jumps to the right page.
• Figures: optionally, a vision model describes images and charts in PDFs and Word documents. The description is searched as well and can be cited like text.
• Document tools: models with tool support can list documents (list_documents), fetch details and the table of contents (document_info) and read passages verbatim (read_document). They only read, and they only see collections you may read yourself.
• Access control: passages from other people’s private collections never end up in a query; access is checked inside the search query itself. Administrators see only metadata in the RAG overview, never content.
• Fully local: embeddings with nomic-embed-text and OCR with olmOCR run through LM Studio in the home network, with Tesseract on the server as an optional fallback. So no document content leaves the house.
• Directory sources: an administrator can fill a collection from a folder on the server or NAS that is read in again periodically. Only folders below the directories allowed in RAG_SOURCE_ROOTS are accepted.
• Guide: wiki: RAG.

Citations

Done Milestone 7

• Every document source carries its locator: section heading, page and paragraph, e.g. “Abschn. 7.5.3, S. 12, Abs. 3”.
• Citation style per account: DIN ISO 690, APA 7, Harvard, Chicago (author-date) or MLA 9. “Copy citation” and “copy bibliography” are also available as BibTeX.
• The server formats citations by fixed rules, not the model.
• Bibliographic data can be maintained per document: books, edited volumes, chapters, articles, reports and standards with number and edition (e.g. “DIN EN ISO 9001:2015-11”). A DOI lookup via Crossref is prepared and off by default.
• Guide: wiki: Zitieren.

Web search and page fetching

Done Milestone 8

• A “web search” switch in the input field. Results go to the model as numbered sources, and the links are listed below the answer. This also works with local models.
• The search backend is a self-hosted SearXNG in the home network.
• Fetching pages and crawling sites: models with tool support read single pages (fetch_url) or several pages of a website (crawl_site). URLs in your question are read directly. robots.txt is respected.
• Protection: addresses in the home network are blocked, including via redirects (SSRF protection). Fetched pages are treated as untrusted and reach the model only as marked source material, never as instructions. URLs do not appear in the log.
• Invented links: links in an answer that appear in no source, no tool result and not in the question get a ⚠. The “Belege prüfen” (check sources) button prepares a follow-up with web search. Base rules for all models urge them not to make things up.

Create and edit images

Partially done Milestone 9

Done (0.3.0): image generation with OpenAI (gpt-image) or Google Gemini.

• Chat models with tool support call the generate_image tool themselves instead of “drawing” images as SVG code.
• An “image” mode in the input field with format (square, portrait, landscape) and quality.
• If the chosen model has no tool support, MultiGPT offers “generate with image model” before sending – never without your consent.
• Images are stored without metadata, and the cost per image is booked. The description only goes to the image model’s provider.
• Guide: wiki: Bilder.

Planned:

• Mark an area of an image and have it changed (inpainting), or create variations.
• Classic editing (crop, resize, rotate, convert, add text, collage) via a bundled MCP server – also by instruction, such as “make the image square”. The original is kept.

Voice

Planned Milestone 10

• Microphone button: the recording becomes text that you can correct before sending.
• Speaker button on every answer, optionally reading answers aloud automatically.
• Requires HTTPS in the home network, because browsers only allow the microphone over HTTPS.

Scratchpad

Planned Milestone 13

• Collect material from chats, your own documents, the web, tools, images and notes, each with its origin and citation data.
• Turn it into a fully referenced document: the model writes only from the entries, MultiGPT sets in-text citations and the bibliography, and verbatim quotes are checked.
• As first steps, the context menu in the chat has been available since 0.3.2 (see above) and, since 0.4.0, permanent storage of files from tool calls (see “New in 0.4.0”); your own prompt templates and “add to scratchpad” follow.

Runner

Done Milestone 14

Since version 0.4.0. The runner is a workspace for models – not on the server, but as a container on the computer the browser runs on. Inside it, a model works like a coding tool, but locked in. A guide is in the wiki under Runner.

• How it works: the model reads, searches and edits files (read, write, edit, glob, grep), runs commands with bash, tests and fixes – over many rounds until the task is done. The work image has Python, git, Node and other standard tools.
• Locked in: the container runs with rootless Podman and by default has no network (can be enabled per role), no home directory and only a volatile in-memory workspace. Only a fixed image is started, with fixed limits for CPU, memory and processes.
• Work stored on the server: every change is committed; the work is kept as a git bundle with its full history in MultiGPT’s database, is backed up with it and can be downloaded. Nothing is left on the computer after the session. For now the whole state is transferred each time; uploading only the changes (incremental bundles) is still missing.
• In the browser: a terminal (directly over a WebSocket, also for programs that wait for input), a preview for HTML pages in a sandboxed frame and the model’s task list. Results can be brought into the chat as an attachment on request.
• Web server and ports: with serve the model starts the lighttpd web server for static files. In the container it listens on 8080 (http) and 8443 (https) and is reachable on the computer at 24080 and 24443. Note: by default the runner binds these ports to 0.0.0.0, i.e. the whole home network and without any login. The setting MGPT_EXPOSE_BIND=127.0.0.1 restricts them to the computer itself.
• Installing packages: with the “package sources” role permission the model can install Debian packages (apt_install) as well as pip, npm, cargo and GitHub packages – only through a proxy with a fixed allow list of package sources, never to the home network or arbitrary sites.
• Scratchpad: a place for intermediate results, either in the container (volatile) or in a folder on the computer. Only what the model explicitly publishes with publish as an attachment reaches MultiGPT.
• Devices on the home network: via ssh or telnet, but only to devices the administrator approves, and with a confirmation before every command. The model picks devices by name only; the server manages the credentials. With setup_host, a model can, after confirmation, set up an approved computer as another runner.
• MultiGPT stays the central point: every access goes through the server, which checks permissions, confirmations, quota and the audit log. The runner connects outgoing only, over TLS, and is paired via a token (revocable, with an expiry date). The server signs its frames, and the runner’s tools are listed in a signed catalog; the runner accepts nothing that does not match.
• The agent team: agents such as n8n or Claude also reach the runner only through a MultiGPT API key with the runner.use permission.
• “Runner runtime” time quota: container minutes per role or person and month, a warning from 80 %, a stop at 100 % – the work is kept.
• The runner agent: a small Rust program multigpt-runner (static binaries for amd64 and arm64) that you download from MultiGPT and run as a service of your own user. Requirement: Linux with systemd and rootless Podman; on Windows and macOS via WSL2 or a Linux VM.
• Still open: Docker is not supported, only Podman. Container logs on the “Runner” page are missing, and tool runs do not yet appear live in the chat terminal.

New in 0.4.0

Done

Besides the runner (see above), 0.4.0 brings these changes since 0.3.6:

• “LM Studio” interface: a provider type of its own where you only enter the address of the LM Studio server. Chat runs over /v1, management over /api/v1. Model management runs through it only.
• Tokens per second for each answer: the header of every answer shows the speed and the time to the first token, measured by MultiGPT itself – for all providers.
• Swapping when graphics memory is full: if another model is taking up the space, the chat offers to unload it and load the requested one – only after confirmation.
• Reasoning effort per model: MultiGPT reads the available levels from LM Studio, and a default can be set per model.
• Memory estimates for Gemma 4 and Qwen3.8 based on measured needs; chat models without a setting are loaded with “parallel 1”.
• Scratchpad on the server: files from tool calls (including SVG, sanitised) are stored permanently, appear on the answer and in the “Scratchpad” menu.
• Reconnect MCP servers: after an outage, an MCP server can be reconnected right in the chat.

Deliberately not in version 1

• No starting or waking LM Studio from within the app.
• No real-time voice conversation.
• No image models running on the server itself.
• No separation of several households in one installation.