> ## Documentation Index
> Fetch the complete documentation index at: https://docs.powabase.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Bring your own LLM

> Point agents, copilots, and indexing jobs at any LiteLLM-supported model (including OpenRouter models like DeepSeek) using the correct provider key and model-string format.

Powabase resolves every model through [LiteLLM](https://docs.litellm.ai/), so the `model` string you pass to an agent, an orchestration, or an indexing job is a **LiteLLM model ID**. Two things have to be right: the registered provider key, and the model string's format for that provider.

<Note>
  **Prerequisites:**

  * Authentication configured (see the [Connect & authenticate](/guides/auth-connection) guide)
  * An API key for the upstream provider you want to use (OpenAI, Anthropic, Google, or OpenRouter)
</Note>

## Model-string format

The prefix tells LiteLLM which provider to route to. Bare IDs route to first-party OpenAI/Anthropic; everything else is prefixed.

| Provider        | Format                     | Example                                |
| --------------- | -------------------------- | -------------------------------------- |
| OpenAI          | bare ID                    | `gpt-4o`, `gpt-5.4-mini`, `o4-mini`    |
| Anthropic       | bare ID                    | `claude-sonnet-4-6`                    |
| Google (Gemini) | `gemini/<model>`           | `gemini/gemini-2.5-pro`                |
| OpenRouter      | `openrouter/<org>/<model>` | `openrouter/qwen/qwen3-235b-a22b-2507` |

<Warning>
  **OpenRouter slugs must match LiteLLM's cost-map keys, not OpenRouter's own slugs.** LiteLLM's `openrouter/...` identifiers occasionally differ from the slugs shown on OpenRouter's website or `/api/v1/models`. When a model errors with a routing or cost-lookup failure, check the slug first: confirm it exists as a key in LiteLLM's [model cost map](https://github.com/BerriAI/litellm/blob/main/litellm/model_prices_and_context_window_backup.json) (search for `openrouter/`).
</Warning>

<Tip>
  The `model` field on an **agent** passes straight through to LiteLLM and is **not** restricted to the model picker shown in Studio. The curated dropdown only gates the Studio UI and the `AGENT_DEFAULT_MODEL` setting; via the API you can set any LiteLLM-resolvable model string. It must support **function calling** if the agent has tools or a knowledge base, since a non-tool model fails on its first tool call.
</Tip>

## Which keys do you need?

Four providers accept a bring-your-own-key (BYOK) credential: `openai`, `anthropic`, `google`, `openrouter`. Register one with the [AI Provider Keys API](/api-reference/ai-provider-keys), then point your model string at it.

If a provider is also **AI-on-us** on your pod (the platform has a server-side key for it), you can skip BYOK and pay in credits instead. Check which providers are covered:

```bash cURL theme={null}
curl '{BASE_URL}/api/ai-provider-keys/platform_supported' -H "apikey: {API_KEY}" -H "Authorization: Bearer {API_KEY}"
# {"providers": ["openai", "anthropic"]}
```

OpenRouter is rarely AI-on-us, so for OpenRouter models you almost always need your own key. Without a usable key, agent runs fail with `402 provider_key_decrypt_failed`. The [billing model](/concepts/billing-model#byok-provider-keys-and-ai-on-us) covers how BYOK and credits interact.

## Worked example: DeepSeek via OpenRouter

<Steps>
  <Step title="Register your OpenRouter key">
    <CodeGroup>
      ```python Python theme={null}
      requests.post(
          f"{BASE_URL}/api/ai-provider-keys",
          headers=headers,
          json={"provider": "openrouter", "api_key": "sk-or-..."},
      )
      ```

      ```bash cURL theme={null}
      curl -X POST '{BASE_URL}/api/ai-provider-keys' \
        -H "apikey: {API_KEY}" \
        -H "Authorization: Bearer {API_KEY}" \
        -H "Content-Type: application/json" \
        -d '{"provider": "openrouter", "api_key": "sk-or-..."}'
      ```
    </CodeGroup>

    The platform validates the key against the provider before storing it. A hard rejection (bad credential) returns `400` and stores nothing.
  </Step>

  <Step title="Create an agent pointed at the DeepSeek model">
    Use the `openrouter/<org>/<model>` format. DeepSeek's OpenRouter org is `deepseek`:

    <CodeGroup>
      ```python Python theme={null}
      agent = requests.post(
          f"{BASE_URL}/api/agents",
          headers=headers,
          json={
              "name": "DeepSeek Agent",
              "model": "openrouter/deepseek/deepseek-chat",
              "system_prompt": "You are a helpful assistant.",
          },
      ).json()
      ```

      ```bash cURL theme={null}
      curl -X POST '{BASE_URL}/api/agents' \
        -H "apikey: {API_KEY}" \
        -H "Authorization: Bearer {API_KEY}" \
        -H "Content-Type: application/json" \
        -d '{"name": "DeepSeek Agent", "model": "openrouter/deepseek/deepseek-chat"}'
      ```
    </CodeGroup>

    <Note>
      `openrouter/deepseek/deepseek-chat` follows the documented format, but confirm the exact slug against LiteLLM's cost map (see the warning above). DeepSeek publishes several models (chat vs. reasoner, for instance), and the LiteLLM key is what the platform routes on. If you give the agent tools, pick a DeepSeek model that supports function calling.
    </Note>
  </Step>

  <Step title="Run it">
    Run the agent as usual. The platform decrypts your OpenRouter key at call time and routes the request through LiteLLM.

    ```bash cURL theme={null}
    curl -X POST '{BASE_URL}/api/agents/{agent_id}/run' \
      -H "apikey: {API_KEY}" \
      -H "Authorization: Bearer {API_KEY}" \
      -H "Content-Type: application/json" \
      -d '{"message": "Hello!"}'
    ```

    A `402 provider_key_decrypt_failed` means the OpenRouter key is missing or unreadable; re-register it in step 1.
  </Step>
</Steps>

## Related

<CardGroup cols={2}>
  <Card title="AI provider keys reference" icon="key" href="/api-reference/ai-provider-keys">
    Store, validate, rotate, and inspect per-project provider credentials.
  </Card>

  <Card title="Build an agent" icon="robot" href="/guides/build-agent">
    Create an agent, assign tools and a knowledge base, then stream a conversation.
  </Card>

  <Card title="Billing model" icon="credit-card" href="/concepts/billing-model">
    How AI-on-us credits and BYOK keys interact, and the 402/503 error shapes.
  </Card>

  <Card title="Settings reference" icon="sliders" href="/api-reference/settings">
    Defaults like `AGENT_DEFAULT_MODEL` and the curated model choices.
  </Card>
</CardGroup>