What you can do in Max Harness
Everything on this page is implemented in the current release. Where a feature depends on the provider, the model or an external service, that is stated next to it.
Chat
Streaming replies
Answers appear as they are generated. Stop a reply at any time and keep what was written so far.
Edit and regenerate
Change an earlier message and continue from there, or regenerate an answer with the same or a different model.
Markdown and code
Tables, lists and fenced code render properly. Code blocks have syntax highlighting, a copy button and a download button.
History and search
Conversations are saved to your account. Search covers titles and message text, and the command palette opens with Ctrl or Cmd + K.
Per-reply model labels
Each answer shows the model and provider connection that produced it, even after you switch models in the same thread.
Models and comparison
Model switching
Pick any enabled model from any connected provider in the model menu. The choice applies from the next message.
Side-by-side comparison
Send one prompt to two, three or four models at once. Choose the answer the conversation should continue from.
Generation settings
Set temperature, top-p, output limit, reasoning effort and a system prompt per conversation. Unset values are left to the provider, because some models reject them.
Usage and cost estimates
Input and output token counts are shown when the provider reports them. Cost is calculated only from a published or user-entered price and is labelled as an estimate.
Organisation
Projects
Folders for related chats with shared instructions that are added to every conversation in the project.
Custom assistants
Reusable personas with their own instructions, preferred provider and model, and default settings.
Pinned chats
Keep important conversations at the top of the sidebar.
Memory
Optional and off by default. Add facts yourself or approve a model's request to save one. View, edit, pause or delete any item.
Files
Document uploads
PDF, DOCX, Markdown, CSV, JSON and common source files. Text is extracted on your server so any text model can read it. Claude and Gemini connections receive PDFs natively.
Image uploads
PNG, JPEG, WebP and GIF for models with image input. If the selected model is known not to accept images, the upload is refused with an explanation.
Generated files
With the file tool enabled, a model can create Markdown, CSV, JSON, HTML and code files for download. Images returned by image-capable models are saved to the conversation.
Limits
Scanned PDFs without a text layer cannot be read by text-only models. Audio and video are not supported. The default size limit is 20 MB per file.
Tools and agents
Web search
Works when a search service is configured: Brave Search, Tavily or a SearXNG instance. Set by the administrator for everyone or by each user with their own key.
Page reading
Fetches a public web page and passes its text to the model. Requests to private network addresses are blocked.
Multi-step tasks
A model can chain up to eight tool steps in one turn. Each call and its result are shown in the conversation.
Approval before action
Every tool can be set to Off, Ask or Allow. Tools that change something, such as saving to memory or creating a file, always ask.
No server access
Agents can only use the tools listed here. There is no shell, no code execution on the server and no access to other users' data.
Accounts and security
Sign-in options
Email and password, plus Google and GitHub once the administrator enters OAuth credentials.
Two-factor authentication
Time-based codes from any authenticator app, with one-time recovery codes.
Session management
See where you are signed in and sign out other devices.
Encrypted keys
Provider keys are encrypted in the database and used only on the server. The browser sees the last four characters at most.
Export and deletion
Download your data as a ZIP, remove individual keys, clear all conversations or delete the account.
Using it day to day
Keyboard shortcuts
New chat, search, focus the composer, toggle the sidebar and stop generation without reaching for the mouse.
Light and dark themes
Follow the system setting or choose one. Both themes are complete across the site and the workspace.
Installable web app
Add Max Harness to a phone home screen or install it as a desktop app from Chrome, Edge or Safari. A connection is still needed to talk to models.
Responsive layout
The workspace adapts to phones and tablets with a slide-over sidebar and a composer that stays within reach.
What Max Harness does not do
Knowing the edges saves time.
It does not include model access
You need an API key from at least one provider, or a local Ollama server the web server can reach. A chat subscription such as ChatGPT Plus or Claude Pro is not an API key.
It does not run code or control a computer
Agents are limited to web search, page reading, memory and file creation. That is a deliberate safety choice for a shared web host.
It does not generate voice or video
The workspace is built around text chat with image and document input. Image output is saved when a chat model returns one.
It does not guarantee prices
Cost figures are estimates from list prices and can differ from your invoice, for example with cached tokens or negotiated rates.