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

# Custom Endpoints

> Use local or third-party OpenAI-compatible models with Kong

Kong works with any OpenAI API-compatible endpoint. Run analysis with local models (Ollama, vLLM), third-party providers (OpenRouter), or your own infrastructure.

## Supported platforms

| Platform       | Example URL                    | Notes                                    |
| -------------- | ------------------------------ | ---------------------------------------- |
| **Ollama**     | `http://localhost:11434/v1`    | Local models, no API key needed          |
| **vLLM**       | `http://localhost:8000/v1`     | High-performance local serving           |
| **OpenRouter** | `https://openrouter.ai/api/v1` | Multi-provider gateway, requires API key |
| **LocalAI**    | `http://localhost:8080/v1`     | Local models                             |
| **LM Studio**  | `http://localhost:1234/v1`     | Desktop app for local models             |

## Setup

### Option 1: Interactive setup

```bash theme={null}
kong setup
# Choose option 3: Custom endpoint
# Enter your endpoint URL, model name, and optional API key
```

### Option 2: CLI flags

```bash theme={null}
kong analyze ./binary \
  --provider custom \
  --base-url http://localhost:11434/v1 \
  --model mistral
```

## Limit overrides

Local models often have smaller context windows than cloud APIs. Use these flags to prevent prompt truncation:

| Flag                    | Default (cloud) | Suggested for local |
| ----------------------- | --------------- | ------------------- |
| `--max-prompt-chars`    | 400,000         | 50,000 - 150,000    |
| `--max-chunk-functions` | 120             | 20 - 40             |
| `--max-output-tokens`   | 16,384          | 4,096 - 8,192       |

```bash theme={null}
kong analyze ./binary \
  --provider custom \
  --base-url http://localhost:11434/v1 \
  --model mistral \
  --max-prompt-chars 100000 \
  --max-chunk-functions 30 \
  --max-output-tokens 4096
```

You can also set these during `kong setup` so they persist as defaults for your custom endpoint.

## Endpoint probing

Kong validates that the endpoint is reachable **before** starting analysis. If your local server is down, you'll get an error immediately instead of waiting for Ghidra to finish loading the binary:

```
Could not connect to LLM endpoint.
Ensure your server is running at http://localhost:11434/v1
```

## Cost tracking

Cost tracking is **disabled** for custom providers since Kong has no pricing data for arbitrary models. Token counts are still recorded and shown in the final stats:

```
Tokens: 1,234,567 in / 456,789 out / 1,691,356 total
  mistral: 234 calls, 1,234,567 in / 456,789 out
Cost tracking disabled for custom provider
```

## Further reading

* [LLM Providers](/configuration/llm-providers) — cloud provider setup
* [Setup Wizard](/getting-started/setup-wizard) — interactive configuration
* [CLI Reference: `kong analyze`](/reference/cli/analyze) — all flags