Documentation
Everything you need tostart working in Azemax.
This guide covers the hosted public beta: creating an account, opening a workspace, working with the AI assistant, connecting a provider key, and running code.
Azemax is a browser workspace. It does not install a local toolchain or attach to a repository on your computer. If you are looking for the product overview first, read the features page.
Quick start
The hosted product has no installer. You need a current desktop browser and an internet connection, because sign-in, workspace sync, AI requests, and code execution all use online services.
- Open the workspace. If you are signed out, Azemax shows the authentication surface instead of the editor.
- Create an account with email and password, or continue with Google. Email and password accounts must verify their address before the workspace unlocks. Google sign-in arrives with a verified address.
- After sign-in, create a project or open an existing one from the workspace. A project contains the file tree, open tabs, and conversation history for that workspace.
- Add or open a file in the explorer, then edit it in the Monaco editor. Use the assistant panel when you want to ask a question or request a change.
No guest mode
The workspace
The interface has four working areas: the file explorer, the editor and tab strip, the assistant panel, and the bottom panel for output and diagnostics. You can resize the major panels by dragging their dividers.
Files and tabs
The explorer supports creating files and folders, renaming, deleting, and moving items. Destructive operations are validated against the current tree, and the delete confirmation can be disabled in settings if you prefer. Open files appear as tabs; a modified file is marked until it is saved.
Editor behavior
The editor is Monaco, with syntax highlighting, multi-cursor editing, bracket matching, folding, and a minimap that can be turned off. Editor settings such as font size, tab size, word wrap, and auto-save live under Settings → Editor.
Diagnostics
The problems panel collects diagnostics reported by the editor tooling and lists the file and location for each entry. Selecting an item jumps to the relevant line. Diagnostics are a development aid, not a guarantee that a program is correct.
Layout
The bottom panel hosts execution output and diagnostics. You can collapse it or switch between its tabs. If you work on a smaller screen, the layout adapts, but the hosted product is designed primarily for desktop browsers.
Using the assistant
Open the assistant from the activity bar or the panel toggle. The composer has three modes, and the mode changes what the assistant is allowed to do.
| Mode | Behavior | Use it when |
|---|---|---|
| Agent | Can inspect the workspace, read and search files, check diagnostics, and create, modify, rename, or delete files. | You want a multi-file change carried out in the workspace. |
| Ask | Answers questions about the code and workspace without writing changes. | You want an explanation or a second opinion before editing. |
| Edit | Applies a focused code change, typically to the file or selection you are working in. | You know what should change and want it applied directly. |
Agent and Edit modes write to the workspace as part of the task. The assistant reports the paths it touched, and changes are validated against the current tree before they are written. Some inline editor actions show a preview for review; the main assistant flow does not require an Accept/Reject step.
Review every change
Giving the assistant useful context
The assistant receives the active file, relevant open files, the file tree, selected text, and results from workspace tools. When a request is vague, name the file, folder, function, or error you mean. For large projects, ask the assistant to search for the relevant files instead of assuming it has the entire tree in context.
Stopping a response
A streaming response can be stopped from the composer while it is running. Stopping a response does not roll back file changes that were already written, so check the workspace afterwards.
Provider configuration
Azemax has two AI modes. Azemax Default routes requests through the Azemax backend and uses the service’s model selection. Bring Your Own Key sends requests from your browser directly to a provider you configure.
Switch modes
Open Settings → AI → Provider and choose the mode. The AI Providers category is where credentials and model selection live.
Connect a built-in provider
- Open Settings → AI Providers.
- Select a provider, then paste an API key created in that provider’s console.
- Choose a model from the starter list, or let Azemax fetch the provider’s catalog when the provider supports model discovery.
- Save the configuration, then send a short test message from the assistant.
| Provider | Credential | Notes |
|---|---|---|
| OpenAI | OpenAI API key | Starter models are selectable in settings; model discovery is used when available. |
| Anthropic | Anthropic API key | Uses the Anthropic messages API through the provider adapter. |
| Google Gemini | Google AI Studio key | A Google AI Studio key is separate from a Google account sign-in. |
| OpenRouter | OpenRouter key | Model availability and free-tier limits are controlled by OpenRouter. |
| Groq | Groq key | Fast inference for open models; the live catalog may differ from this starter list. |
| NVIDIA NIM | NVIDIA API key | Use the model identifier exposed by the NIM endpoint you have access to. |
| Custom OpenAI-compatible | Endpoint API key | Requires an HTTPS base URL and an OpenAI-compatible chat completions API. Tool calling is required for agent features. |
The exact model catalog is controlled by the provider and can change without an Azemax release. If a model you expect is missing, enter its identifier manually or refresh the catalog.
Connect a custom OpenAI-compatible endpoint
Select the Custom OpenAI-compatible provider and supply:
- An HTTPS base URL that accepts OpenAI-style chat completion requests.
- An API key, if your endpoint requires one.
- The model identifier your endpoint expects.
Agent features that use tools require the endpoint to support tool or function calling. Endpoints that depend on client certificates, a VPN, or a browser extension cannot be reached from the hosted page.
How keys are stored
By default, a key is held only in memory for the current tab and is lost when the tab closes. If you enable Remember on this device, Azemax encrypts the key with AES-GCM using a key derived from a passphrase you choose. Only the ciphertext is written to IndexedDB; the passphrase and derived key are not stored.
Key hygiene
Remove stored keys
The AI Providers page includes an action to remove stored provider API keys from the device. Clearing site data in the browser also removes the encrypted vault, and you can revoke the key in the provider console independently.
Running code
The Run action sends the active file to the hosted execution service and opens the output panel. The panel shows standard output, standard error, compiler messages, and the exit status for the run.
The hosted product uses an isolated Wandbox sandbox. It does not execute arbitrary code on the application server, and it does not expose a general-purpose PTY shell to hosted users. Local development builds of the source project can enable a local shell, but that capability is not part of the hosted beta.
| Language | Notes |
|---|---|
| Python | Script execution |
| JavaScript / TypeScript | Node-based execution |
| C / C++ | Compile and run |
| Java | Compile and run; the public class should match the file name |
| Ruby, PHP, Perl | Script execution |
| Rust, Go, Swift, Scala | Compile and run |
| Lua, Bash | Script execution |
| C# | Compile and run |
| R | Script execution |
The service allows roughly 60 seconds of execution time and 30 seconds of compile time per run. These limits keep the sandbox available to other users. Versions are supplied by the sandbox provider and can change; pin behavior in your own environment if exact versions matter.
Not a production runner
A program that reads from standard input may wait for a line before continuing. If execution appears stalled, stop it and check whether the program is blocked on input or an infinite loop.
Saving and sync
Edits are kept in the open editor buffer until they are saved. Auto-save can be enabled under Settings → Editor → Saving & Files. The tab strip marks unsaved files so you can see what still needs to be written.
When you are signed in, projects, settings, and assistant conversations are associated with your account and can be restored on another device. Access is scoped to the authenticated user. Workspace sync uses Firebase; it is not a peer-to-peer or local-only sync mechanism.
There is no public project-sharing link and no multiplayer editing in the beta. Export or copy files you need to keep outside the workspace, and do not treat Azemax as your only backup.
Troubleshooting
The workspace will not load after sign-in
Email and password accounts must verify their address before product access is enabled. Check the verification email, including spam. Google accounts skip this step. If the page still fails, sign out, reload, and sign in again; then check that your browser allows storage for the site.
The first request is slow
The API runs on a hosted service that may sleep when idle. The client sends a lightweight health request when the workspace starts, but the first real request after a long idle period can take longer while the backend wakes. Retry once the connection is established.
Bring Your Own Key fails
- Confirm the key belongs to the provider you selected and has not been revoked.
- Check that the provider account has access to the model you chose.
- For a custom endpoint, confirm the base URL is HTTPS and implements OpenAI-style chat completions.
- If the browser blocks the request, check whether an extension, proxy, or corporate network is intercepting it.
A run fails immediately
Check the output panel for a compiler message before assuming the sandbox failed. Unsupported languages, missing file names for Java classes, and syntax errors are reported as compile or runtime errors. If the service itself is unavailable, retry; the sandbox is a third-party dependency.
The assistant ignores a file
Large files and long conversations can exceed a model’s context window. Reference the file explicitly, close unrelated tabs, or start a new conversation for a separate task. Ask mode is useful for narrowing down what the assistant actually sees.
Something else is wrong
Report it through the contact page. Include the browser, the action you took, the expected result, and any error text. Do not include API keys, passwords, or private source code in a support message.
Privacy and data
Data handling depends on the features you use. Signed-in workspaces and settings sync through Firebase. AI prompts and the context you include go to your selected provider in BYOK mode, or to the Azemax backend in default mode. Code you run goes to the hosted sandbox.
Do not paste secrets, regulated data, or code you are not allowed to share into a hosted tool. Read the Privacy Policy for the data categories, providers, retention practices, and deletion process, and the Terms of Service for acceptable use and execution risks.
Ready to try the workflow?
Sign in and open a project. No install and no payment card.