AI connection and models

Configure the existing Axe connection and select chat or voice-input models.

Open AI Settings

Open Cyber AI chat from View, then click its AI settings gear. The dialog identifies whether a credential is configured and its source. Configured does not prove a successful provider request.

The human particle editor remains usable without an AI connection. Help > Cyber AI: ask anything activates contextual inspection; it does not configure a provider.

Credential and model fields

Enter your API key in the password field. Leaving it blank keeps the current customer key. Remember key using OS encryption persists the key when encrypted storage is available; otherwise uncheck Remember for a session-only key.

Chat model is required: the Axe agent runs on it and OpenRouter rejects a request without one. Voice input model is the model used to transcribe your microphone; only models that accept audio input can be picked, and leaving it blank turns voice input off. Reply voice picks the operating-system text-to-speech voice used to read Axe's replies aloud (the same setting as Preferences > Local speech); Preview speaks a sample.

Save submits the fields and closes on success. Save refuses a blank chat model and a voice model that does not accept audio, and says why. A failed Save displays an error and preserves the draft for correction or retry. Cancel closes without submitting the current draft.

Provider

OpenRouter is the cloud provider: your API key, pay per token. Anthropic (Claude direct) talks to Anthropic's own API through the official SDK. Click Log in with Anthropic: your browser opens Anthropic's login (the ant CLI's OAuth flow; the CLI is fetched once into your CyberStorm profile if it is not installed), you pick your organization and workspace, and CyberStorm is connected. No key to copy. Or enter an Anthropic API key (console.anthropic.com), or set ANTHROPIC_API_KEY; a key wins over the login. Usage bills that organization at Anthropic's published prices. The model list shows every Claude model you can use (Fable, Opus, Sonnet, Haiku) with its price per million tokens, so you pick by cost. Fable is the most capable and the most expensive, Opus 5.5 / Opus 5 far cheaper, Sonnet 5 cheaper still; every Claude model calls tools and reads images; voice input is not offered on this provider. Log out removes the stored login. Local server is your own OpenAI-compatible server on your machine or LAN: Ollama (http://host:11434/v1), LM Studio (http://host:1234/v1), llama.cpp server, vLLM, or anything that serves /v1/models and /v1/chat/completions. Enter its base URL ending in /v1; a key is only needed if that server asks for one. Axe needs a model that supports tool calling (function calling); a local server does not report that, so the list marks tools as unknown and you find out by trying: a model without tool calling answers in text and Axe cannot read or edit anything.

The chat model and the voice input model are separate picks; voice input needs a model that accepts audio and never bills the chat model.

Models list and recommendations

The model list loads from the provider when the dialog opens and sits under the fields. Three recommended picks come first: Best (highest fit), Balanced (best fit at or below the median cost) and Budget (cheapest with tools), each with its reason line and a Use for chat button. They are chosen from the provider's own capability and price data (fit score first, then cost), not from a CyberStorm test run. A new profile gets the Balanced pick pre-filled as the chat model; Save keeps it. If you never open AI Settings, the first message you send picks the Balanced model itself, saves it and tells you which one it chose.

Each row shows the model name and ID, Fit out of 7 (tool calling 4, image input 2, context 128k+ 1 / 32k+ 0.5), price per million tokens in and out (USD, as the provider reports; $0 on a local server), context length, and what it accepts: tools, image, audio. Recommended first by default; sort by cheapest, name or largest context; the search box narrows by name or ID.

Fits Axe (tool calling) is on by default and hides models that cannot call tools, which Axe needs for every edit and read. Turn it off to see everything. Free-tier models (":free" or $0 on a paid provider) and batch-only variants (":batch", served only through OpenRouter's Batch API and answered with HTTP 404 on chat) stay listed and flagged in Accepts but are never a recommended pick or the default. Chat on a row sets it as the chat model; Voice is enabled only on rows that accept audio input and sets the voice input model. The chosen rows are highlighted. Save is still required.

Reload models fetches the list again; a failed request displays the error and can be retried.

Forget and history

Forget customer key immediately removes the customer credential and saves the currently entered model IDs. Cancel does not undo this completed action. A configured development credential, when present, is separate.

If removal fails, the dialog keeps the draft and displays the error; retry after resolving it. Action history opens the saved proposal records. Inspect displays exact paths and old/new values. Revert action is available for committed records and creates an undoable edit in the original document. It rejects missing targets and later conflicting values; open the original file or resolve the conflict before retrying. Undo latest edit runs shared editor Undo, which may undo a different recent edit. Close returns to AI Settings.