# zidane/providers/local

> A provider for a local OpenAI-compatible server such as Ollama, llama.cpp or LM Studio.

Import: `import { … } from 'zidane/providers/local'`. Source: [src/providers/local.ts](https://github.com/Tahul/zidane/blob/main/src/providers/local.ts).

Use `local()` when the model runs on your machine or network. Pass `baseURL` and `defaultModel`, or set `LOCAL_LLM_BASE_URL` and `LOCAL_LLM_DEFAULT_MODEL`. A key is optional. Set `LOCAL_LLM_API_KEY` for a server that checks one. For a model that makes tool-call mistakes, spread `localReliabilityBehavior()` from `zidane` into `behavior`. It caps tool output and steers the model when it writes a tool call as text.

```ts
import { createAgent } from 'zidane'
import { local } from 'zidane/providers/local'

const provider = local({ baseURL: 'http://localhost:11434/v1', defaultModel: 'llama3.1:8b' })
const agent = createAgent({ provider })
const stats = await agent.run({ prompt: 'Say hello.' })
console.log(stats.turns)
```

## Functions

### `local`

Function.

Local OpenAI-compatible LLM provider.

Thin wrapper around [`openaiCompat`](https://zidane.si/docs/reference/providers-openai-compat.md#openaiCompat) for self-hosted runtimes that
speak the standard `POST /chat/completions` + SSE dialect: Ollama, vLLM,
LM Studio, Lemonade, llama.cpp's server, text-generation-webui, etc.

Caching and reasoning are left off — local runtimes typically strict-
validate the request schema and would 400 on the extra fields. Override
`capabilities` when pointing at a vision-capable deployment.

```ts
function local(params?: LocalParams): Provider
```

**Parameters**

- `params?` (`LocalParams`): Base URL, API key, default model, capabilities and extra headers.

**Returns** `Provider`

**Example**

```ts
import { createAgent } from 'zidane'
import { local } from 'zidane/providers/local'

const agent = createAgent({
  provider: local({ baseURL: 'http://localhost:11434/v1', defaultModel: 'llama3.1:8b' }),
})
await agent.run({ prompt: 'Say hello.' })
```

Source: [src/providers/local.ts:99](https://github.com/Tahul/zidane/blob/main/src/providers/local.ts#L99)

## Interfaces

### `LocalParams`

Interface.

Options for [`local`](https://zidane.si/docs/reference/providers-local.md#local).

```ts
interface LocalParams
```

**Properties**

- `apiKey?` (`string`): Optional bearer key. Most local servers don't authenticate; some (vLLM
  with `--api-key`, gateway proxies, llama.cpp with `--api-key`) do.
  Falls back to `LOCAL_LLM_API_KEY`, then to a placeholder string accepted
  by unauthenticated endpoints.
- `baseURL?` (`string`): Base URL of the local OpenAI-compatible server. `/chat/completions` is
  appended. Examples:

  - Ollama:    `http://localhost:11434/v1`
  - vLLM:      `http://localhost:8000/v1`
  - LM Studio: `http://localhost:1234/v1`
  - Lemonade:  `http://localhost:8000/api/v1`
  - llama.cpp: `http://localhost:8080/v1`

  Resolution order: explicit `params.baseURL` > `LOCAL_LLM_BASE_URL` env >
  throw.
- `capabilities?` (`ProviderCapabilities`): Provider capability flags. Defaults to text-only — typical OSS local
  deployment. Override when pointing at a vision-capable runtime (e.g.
  Ollama serving a multimodal model, vLLM with a VLM).
- `defaultModel?` (`string`): Default model id. Local runtimes have no fixed catalogue — whatever the
  user has loaded (e.g. `llama3.1:8b`, `qwen2.5-coder`, a HF repo path).
  Falls back to `LOCAL_LLM_DEFAULT_MODEL`.
- `extraHeaders?` (`Record<string, string>`): Extra HTTP headers sent on every request. Forwarded to the underlying
  [`openaiCompat`](https://zidane.si/docs/reference/providers-openai-compat.md#openaiCompat) client — use for auth/routing headers a self-hosted
  gateway in front of `baseURL` consumes.

Source: [src/providers/local.ts:5](https://github.com/Tahul/zidane/blob/main/src/providers/local.ts#L5)