Configuration

All settings live in the add-in's Settings panel (gear icon) and are stored locally in your browser via localStorage. No telemetry.

The panel has two tabs:

• Connection: provider, endpoint, key, model, advanced options, test
• Behavior: custom instructions, Human in the Loop

Custom Instructions

A free-form text box in the Behavior tab. Whatever you write is injected into the agent's system prompt on every turn, so it applies persistently to the whole conversation — not just the next message.

• They are stored locally with your other settings (localStorage) and survive reloads.
• They are sent to the provider as part of the system prompt, alongside the built-in rules. Your custom instructions take precedence over any conflicting built-in rules.
• Editing them mid-conversation takes effect from the next message: providers with content-keyed caches (Anthropic, OpenAI, OpenCode Zen) adapt automatically, and Gemini invalidates its context cache when the system prompt changes.
• Clear the field and press Apply to remove them.

Example:

Always write in British English. Use a formal tone. Never change the document's heading structure.

Human in the Loop

When enabled, every document-modifying tool call pauses for your approval before it runs — useful for anything you can't easily undo. See Features for details.

Presets

A preset is a one-click bundle of providerId + endpoint + default model + advanced configuration. The Preset selector appears at the top of the Connection tab.

Preset	Provider	Endpoint	Default model	Key required
OpenAI	OpenAI-compatible	https://api.openai.com/v1	gpt-5.6-luna	Yes
Anthropic Claude	Anthropic	https://api.anthropic.com/v1	claude-haiku-4-5	Yes
Google Gemini	Gemini	https://generativelanguage.googleapis.com/v1beta	gemini-3.5-flash-lite	Yes
Ollama	OpenAI-compatible	http://localhost:11434/v1	llama3.1	No
OpenRouter	OpenAI-compatible	https://openrouter.ai/api/v1	openai/gpt-4o	Yes
OpenCode Zen¹	OpenAI-compatible	https://opencode.ai/zen/go/v1	deepseek-v4-flash	Yes
Custom	OpenAI-compatible / Anthropic / Gemini	Custom	Custom	Optional

¹ OpenCode Zen only appears on self-hosted / dev deployments. See here.

Selecting a preset fills in the endpoint and model, clears the API key, and sets the proxy flag from the preset (on for OpenCode Zen, off for the rest). When the preset is Custom, an extra Provider dropdown appears so you can pick the provider type (OpenAI Compatible, Google Gemini, Anthropic Claude) before configuring the endpoint.

Endpoint URL

The base URL of the provider, without trailing slash. Examples:

• https://opencode.ai/zen/go/v1
• https://api.openai.com/v1
• http://localhost:11434/v1

The add-in appends /responses, /chat/completions, /messages, or the Gemini method path automatically, depending on provider and endpoint flavor.

API key

Only shown when the selected preset requires one. Stored in browser localStorage, sent directly to the provider in request headers. Never sent anywhere else.

Model

The Model dropdown automatically loads available models from the provider's /models endpoint (when a key is present and the provider supports it).

This request respects the proxy setting: with "Proxy API requests through this server" on, the /models call also goes through the proxy, so providers without browser CORS load their model list.

It could happen that the /models endpoint isn't reachable, or the provider needs no API key (e.g. Ollama) If for any reason the model list can not be obtained, a free-text input appears so you can type the model name.

For the OpenRouter preset, a Free models only filter narrows the list to :free models.

Advanced

Collapsed by default. Contains:

Max Tokens

The maximum number of tokens the model may produce per response (default 4096). Applies to both reasoning and answer combined. If a reply is cut off because it reached this limit, the add-in shows a truncation notice:

"The model's reply was cut off because it reached the output token limit. Increase Max Tokens in Settings and try again."

If this happens, raise this value and try again.

Use old /chat/completions endpoint

Available for OpenAI-compatible providers (not Gemini/Anthropic).

• Off (default): uses the Responses API (/responses).
• On: uses the legacy /chat/completions endpoint.

Enable this only if your endpoint doesn't support /responses (for example, some Ollama setups or older gateways).

Reasoning Effort

Controls how much the model thinks before answering. It's a free-form text field; blank (default) means don't send anything and use the provider's default.

Supported values are provider/model-specific. Setting an effort a model doesn't support will make the request fail. You can use Test Connection to check whether your provider accepts the value before applying it.

Prompt Caching

The available caching settings depend on the provider you're using.

For Anthropic, you can set the Cache TTL: 5m (default) refreshes for free within active bursts, while 1h survives longer pauses but cache writes cost 2×.

For Gemini, you can toggle Enable Prompt Caching to turn the context cache on or off, and set Cache rebuild size (tokens) to rebuild the cache when the conversation grows past that many tokens (default 2000, min 1024). Caching only happens when the prompt is large enough to be worthwhile.

For OpenAI, OpenRouter and OpenCode Zen, caching happens automatically, so no settings are needed.

Proxy API requests through this server

Available in Advanced on self-hosted / dev deployments (the build flag VITE_PROXY_ENABLED controls it; GitHub Pages builds ship without it).

When enabled, provider calls are rewritten to a same-origin /proxy/<encoded-baseUrl>/<path> endpoint instead of calling the provider directly. The serving process forwards the request server-to-server.

Use it for providers that don't allow browser-origin requests, most notably OpenCode Zen.

Connection test

Test Connection sends a tiny request (maxTokens: 10) to the configured endpoint and reports latency or the exact error. Useful for validating keys and endpoints before starting work.

Apply / Clean

Settings are drafts until you press Apply. Clean discards your edits and reloads the stored configuration.