ncode Desktop / Customize

Providers, models and web search

ncode works with the model providers you add. You need at least one, with your own API key: Anthropic, or any endpoint with an OpenAI-compatible API (OpenAI, OpenRouter, DeepSeek, a local Ollama or LM Studio, and many more).

Providers#

ncode talks to the model endpoints you add. It ships with no key of its own: you bring yours.

A provider has:

• a kind: Anthropic (the Messages API) or OpenAI-compatible (any endpoint that speaks the OpenAI chat API: OpenAI itself, OpenRouter, DeepSeek, or a local server such as Ollama or LM Studio);
• a base URL, for example https://api.anthropic.com or https://api.openai.com/v1;
• an API key (a local server may not need one);
• its models, typed in or fetched from the endpoint's model list, and a default model.

Models per role#

Different kinds of work can use different models. Each role has its own default, and a conversation can still pick its own chat and sub-agent model:

Reasoning effort#

Effort is how hard a model thinks before it answers. The built-in levels are:

A provider, and each of its models, can carry its own list of levels, so the levels you see are the ones your provider offers.

Pricing and budget#

You can enter a price per million input and output tokens (and cache reads and writes) for each model, so every run shows what it cost. A monthly budget is a spend target you watch; nothing is blocked when it runs out.

Adding a provider#

1. Open Settings → Providers & models and click Add provider.
2. Fill in Name, Kind, Base URL and API key (click Show to check what you pasted).
3. Optionally list Models (one per line) and a Default model; you can also fetch them in the next step.
4. For an Anthropic provider, Server-side fallback on refusal is on by default: when the API declines a request, it answers with its recommended substitute model instead of an empty reply.
5. Click Save provider.

For an OpenAI-compatible service, the base URL is the part before /chat/completions in its documentation.

Fetching models#

Click Fetch models on a provider's row, or Fetch all models from endpoints at the top, to ask each endpoint for its model list. It is also the quickest test of a key: if the list arrives, the key works. If it fails, the endpoint's error is shown, with your key removed from it.

Deleting a provider asks first; conversations that used it fall back to the default model.

Default models#

Settings → General → Defaults sets the model each role starts with: Chat model, Sub agent model, Default scheduled model, Default workflow model and Implementer model (consensus), plus a default effort for each. Click Save defaults. The research roles have their own pickers under Deep research.

A conversation can override its chat and sub-agent model from the composer's model chooser.

Effort levels per provider#

Click Efforts on a provider's row to edit which effort levels it offers and how each one reaches its API, for the whole provider or per model. Leave it alone to use the built-in levels for the provider's kind.

Prices, usage and budget#

• Settings → Pricing holds a price per million tokens for each model: input, output, cache read and cache write, plus its context window. With prices set, every run, agent and conversation shows what it cost.
• Usage history in the rail shows spend over the last 7, 30 or 90 days, run by run.
• Settings → Budget → Monthly budget (USD) is a target the usage bar fills against. Nothing stops when it is reached.

Web search#

Web search and page reading#

Agents search the web with web_search and read pages with web_fetch. Search needs at least one search engine with your own key; nothing is preconfigured.

• Engines answer searches: Tavily, Exa, Brave and Serper.
• Readers turn a web page into clean text: Jina and Firecrawl. Without a reader, pages are read directly.

Engines are tried in the order you arrange them. The first enabled engine that returns results answers; an engine that fails or finds nothing hands over to the next one. Each engine has its own key, an optional base URL and an on/off switch.

Reading a page on your own machine or network (localhost, private and link-local addresses) is allowed, but asks for approval unless the project is in Full access.

Setting up a search engine#

1. Get an API key from one of the engines (Tavily, Exa, Brave or Serper).
2. Open Settings → Deep research. The Search providers block lists every engine and reader.
3. On the engine's row, paste the key, switch it on and save.
4. Use the up and down arrows to set the order engines are tried in.

A reader (Jina or Firecrawl) is optional: agents read pages directly without one.