Models and keys

Agent V does not ship its own model. You connect a provider, and the agent sends its requests there with your key. Everything on this page lives in Settings, “Providers”: “Model providers and their keys. Keys are encrypted on this device and sent only to their provider.”

Providers

Provider	Key needed
OpenAI	Yes
Anthropic	Yes
Gemini	Yes
Ollama	No, unless you use Ollama Cloud
DeepSeek	Yes
Groq	Yes
OpenRouter	Yes
xAI	Yes
Mistral	Yes
Custom OpenAI-compatible	Only for public hosts
OpenCode Go	Yes
Amazon Bedrock	An API key or AWS access keys
Google Vertex AI	A service-account key or this computer’s gcloud login
Azure OpenAI	Yes (added as a custom endpoint)

The “Model” group at the top of the Providers section sets the “Provider for new tasks”, the “Model” and the “Effort”. The provider menu lists only providers that have a key or need none. The refresh button reloads the model list. Each provider row also has “Use for new tasks”, and the one in use is marked “In use”. A workspace can run its own model: Settings, “General”, “Workspaces” (“Override gives a workspace its own model and agent settings.”). The brief on the “New task” page also has a model and effort menu; see Writing a task.

Adding a key

1. Open Settings, “Providers”, and find the provider under “API keys”, or under “Custom endpoints” for your own servers.
2. Click “Add key”, or “Manage” when the row has a key or needs none.
3. Paste the key into the “Paste a {provider} API key” field and click “Save key”.

“Get a key” links to the provider’s key page until a key is saved. “Clear” removes a saved key; for Bedrock and Vertex AI the button is “Sign out”. Each row says where it stands: “Key saved”, “Local · no key needed”, “No key”, or “Can’t save keys”. Dictation’s cloud engines (Settings, “Voice”) use the OpenAI or OpenRouter key saved here.

Where keys are stored

Keys are encrypted with your operating system’s secure storage and written to secrets.json in the app’s data folder. A key is sent only to the provider it belongs to.

If your system has no secure storage, keys cannot be saved at all: the field reads “Secure storage unavailable” and the page says “OS secure storage is unavailable on this system, so API keys cannot be saved.” On Linux, Agent V also refuses the insecure basic_text backend. Set up a real password store (for example --password-store=gnome-libsecret) to save keys there.

Custom OpenAI-compatible endpoints

With no endpoints added, “Custom endpoints” reads “Any OpenAI-compatible server: vLLM, llama.cpp, LM Studio, a hosted gateway”.

• Click “Add endpoint”, pick “OpenAI-compatible” or “Azure OpenAI”, give it a name (“Name, such as Home GPU”) and a base URL, then click “Add”. For a common host, “Start from…” fills in both: Together AI, Fireworks, Cerebras, DeepInfra and Moonshot (Kimi), or LM Studio, vLLM and llama.cpp on this machine. Paste the host’s key on the endpoint’s row once it is added.
• You can save up to 20 endpoints. Each one appears as its own provider. “Remove” deletes one, but not while it is the provider for new tasks.
• The base URL should contain /v1, or a vendor’s mount such as /v1/openai. If it has no /v1, Agent V adds it.
• The built-in “Custom OpenAI-compatible” row starts at http://127.0.0.1:8080/v1. An added endpoint needs its own URL.
• “Public hosts need a key; loopback and a private LAN can go without.”

When you switch to an endpoint, Agent V asks it which models it serves and picks one of them, so a task never starts on a model left over from another provider. The model picker marks a model the server doesn’t list as “Not reported by this server”.

Model limits

Agent V sizes every prompt to the model’s context window, and asks for no more output than the model allows. Hosted providers report both. A local or custom server often doesn’t, and a guessed window that is too large means the server cuts the prompt from the front without saying so.

• What Agent V reads by itself: llama.cpp’s /props, LM Studio’s model list, and, for Ollama, the context the running model was loaded with (OLLAMA_CONTEXT_LENGTH, or the model’s num_ctx), which is often far below the model’s maximum.
• Setting them yourself: the rows for Ollama, Custom OpenAI-compatible and each added endpoint list “Model limits”. Each model shows its context, and where that came from: “from the server”, “reported”, “guessed” or “set by you”. “Set” takes a context window and, optionally, a max output, as a number of tokens such as 32768 or 32k. “Reset” goes back to what the server says.
• Reachability: a local provider’s row says “Local · 3 models” when the server answers, or “Not reachable” with “Check again” when it doesn’t.

On Groq, xAI, Mistral, DeepSeek and OpenRouter, a request asks for up to 16k output tokens, or 32k on the step after a reply was cut off. Left to themselves some of these hosts stop early (DeepSeek’s default is 4k), and OpenRouter holds credit against the model’s full maximum. If a host refuses the limit, the request is sent once more without one. A max output you set yourself is sent as it is.

A custom endpoint’s model has a cost when the server reports one, when its id matches a model in the models.dev catalog, or when you give it a price in Settings, “Agent”, “Cost”, “Prices”. Otherwise its tasks show tokens only. See Usage and cost.

Azure OpenAI

In “Add endpoint”, pick “Azure OpenAI” and paste the resource endpoint from the Azure portal, like https://my-resource.openai.azure.com. A deployment URL works too; Agent V keeps only the resource and uses Azure’s v1 API (/openai/v1), which needs no api-version. The key goes in an api-key header. Pick models by their deployment names.

Extra headers

An endpoint you added has an extra headers box, one Name: value per line and up to 16, for gateways that route on a header or want an organisation id. Headers the request owns (Host, Content-Type, Content-Length and the like) can’t be set. The headers are stored in settings, not the key vault, so put secrets in the API key.

Tool calls written as text

Some local models write a tool call as text in their own template syntax (<tool_call>…, [TOOL_CALLS]…, <|python_tag|>…) instead of as a real tool call. Agent V recognises these, <function=…> and a reply that is only a JSON call, and runs the call, but only for tools the model was offered. JSON shown inside a longer explanation is left as a reply.

Amazon Bedrock

Open the “Amazon Bedrock” row and set the “AWS region”. Tasks call bedrock-runtime.<region>.amazonaws.com, and the model list comes from bedrock.<region>.amazonaws.com. Then sign in one of two ways:

• “API key”: a Bedrock API key, sent as a bearer token.
• “Access keys”: an access key ID, a secret access key and, for temporary keys, a session token. Requests are signed with AWS Signature Version 4.

Every model goes through Bedrock’s Converse API. The model list shows your account’s system inference profiles (the global. and us. ids newer Claude models require) and the on-demand text models. Claude gets thinking and prompt caching, Amazon Nova gets prompt caching, and other models get what Converse supports.

Google Vertex AI

Open the “Google Vertex AI” row and set the “Google Cloud project” and the “Vertex location”: global, us, eu, or a region like us-east5. Then sign in:

• “Key file”: paste a service-account key (the whole JSON file).
• “gcloud login”: use the login gcloud auth application-default login wrote on this computer. “Check this computer” shows what it found. Then click “Use this login”. Nothing is copied; the file is read when a task starts.

Gemini models go through streamGenerateContent, Claude models through streamRawPredict, and the open and partner models Vertex serves as a service (Llama, DeepSeek, Qwen, gpt-oss and others) through Vertex’s OpenAI-compatible Chat Completions endpoint, named publisher/model, for example meta/llama-3.3-70b-instruct-maas. Mistral and AI21 models use routes of their own that Agent V does not speak yet. Vertex has no endpoint that lists the models a project can use, so the list comes from the models.dev registry, after Agent V checks your sign-in works.

Bedrock, Vertex AI and Azure models are priced from the models.dev catalog when their id matches one there, or from a price you set in Settings, “Agent”, “Cost”, “Prices”; otherwise their tasks show tokens but no cost, and the spend limits can’t count them.

Claude through OpenRouter

For Claude models through OpenRouter, or through a custom endpoint pointed at openrouter.ai, Agent V marks where the provider may cache the prompt.

Helper and utility models

Settings, “Agent” has two more model choices. Both start as “Same as the task”.

Setting	What it runs
“Helper model”	Instances. “Same as the task” runs each instance on the model its task was started with.
“Utility model”	Compaction summaries, commit messages and Structure. A cheaper model is enough for these. Compaction still decides what to keep by the task’s model. If this provider has no key, the task’s model is used.

Pick a provider, then one of its models. Choosing a provider alone saves nothing.

Ollama: local or cloud

Ollama is the default provider. It points at http://127.0.0.1:11434, the local Ollama daemon, and needs no key. Open the Ollama row to point it at another host.

If you save an Ollama API key, the provider moves to Ollama Cloud at https://ollama.com. In the app’s words: “Saving an API key moves this to Ollama Cloud (https://ollama.com); a local host never needs one.”

OpenCode Go

“OpenCode Go is a $10/month subscription. Subscribe, then paste the API key from its console.” The “Subscribe” link opens the OpenCode Go page. Its model list is not fixed in the app: Agent V reads the current list from the public models.dev registry.

Behind a proxy

Settings, “General”, “Network” sets the proxy for everything the app sends: provider calls, MCP servers, updates, downloads, the agent’s browser, and the git, npm and other commands the agent runs.

“Proxy”	What it does
“System”	Uses HTTPS_PROXY or HTTP_PROXY, with NO_PROXY, when set. Otherwise it uses the operating system’s proxy.
“Manual”	Uses the “Proxy address” for everything except the hosts in “Skip the proxy for”. This computer is always skipped.
“None”	Connects directly.

The line under the choice shows the proxy in use and where it came from. A manual proxy that asks you to sign in takes a “User name” and a “Password” under the address; the password goes to the key vault, never settings.json. The app’s own traffic and the agent’s browser sign in with them. Commands the agent runs get the proxy address without the password, so a command cannot read it out of its environment; to give them one, set HTTPS_PROXY=http://user:password@host:port before starting the app. A SOCKS system proxy reaches the agent’s browser and updates, but not provider calls. With a PAC file, provider calls follow the proxy it picks for https://api.openai.com.

If the provider does not connect

The first-run setup checks the provider by reading its model list, and shows the reason if that fails. Home also lists a missing key under “Needs you”: “Tasks can’t start until a key is saved for it”, with an “Add key” button that opens this page.