The AI Assistant lets you analyze risk data through natural language. It is off until you supply a provider and, for a cloud provider, an API key. MTC Skopos never proxies these calls: your client talks directly to the provider you configure, with your key.
Provider setup
| Provider type | Typical models | Notes |
|---|---|---|
| Anthropic | Claude Opus 5.5, Sonnet 5.5, Haiku 5.5, Fable 5.1 | The default |
| OpenAI Compatible | GPT-class models, and any OpenAI-compatible endpoint | Set a base URL for anything other than the public OpenAI API |
| Azure OpenAI | Azure-hosted deployments | Needs a base URL, API version, and auth header |
| Local | A model on your machine or network: LM Studio, Ollama, llama.cpp, vLLM | No API key needed. See Local provider |
- Go to Settings > AI Assistant
- Select your Provider Type
- Enter your API Key. It is optional for the Local provider
- Choose a Model. The shipped default is
claude-sonnet-5-5 - Set a Base URL if you use a self-hosted, proxy, or Azure endpoint
Azure OpenAI
Azure deployments need three extra fields:
| Field | Default |
|---|---|
| Base URL | none, required |
| API Version | 2024-10-21 |
| Auth Header | api-key |
Local provider
Use the Local provider for a model that runs on your own machine or network. The API key is optional, and the base URL defaults to http://localhost:1234.
| Server | Base URL |
|---|---|
| LM Studio | http://localhost:1234 |
| Ollama | http://localhost:11434 |
| llama.cpp | http://localhost:8080 |
The Local provider adds three settings:
| Setting | What it does |
|---|---|
| Context window | The context length the model is loaded with on your server. LM Studio and vLLM report it, so it is filled automatically. For Ollama or llama.cpp, enter the value configured on the server |
| Reasoning switch | How your server turns reasoning on, off, or sets its depth: reasoning_effort (gpt-oss, Ollama), chat_template_kwargs (Qwen3 on vLLM, llama.cpp, SGLang) or enable_thinking (Qwen API). Keep Not sent to use the server default |
| Reasoning budget field | Caps the reasoning tokens according to the effort, and always leaves room for the answer. On a local server, reasoning and answer share the output token limit, so uncapped reasoning can leave no answer. Set it only if your server supports the field |
Model settings
| Setting | Default | What it does |
|---|---|---|
| Effort | Model default | How much the model reasons before it answers. Higher effort gives a more thorough analysis, but is slower and uses more tokens. Low or Medium is usually sufficient for summaries. The levels are Low, Medium, High, Extra high and Max. If the model does not support the level you select, the next lower level is used. With the Local provider, Off disables reasoning when the server has a reasoning switch |
| Temperature | Model default | Response variability. Select Custom to set a value from 0.0 (near-deterministic) to 1.0 (creative). Some models do not accept a temperature, for example the current Claude models. For these models, the value is not sent |
| Max Output Tokens | 32,000, or the model maximum if lower | Ceiling on the length of a single response |
| Response Language | English | Language the model answers in |
| Context Window | Set from the model, 1,000,000 for the current Claude models | Total conversation budget before compaction starts. When the provider reports the model limits, these values are used |
Two further settings govern long conversations: a reserve of 30,000 tokens is held back for the response, and 40,000 tokens of recent turns are kept verbatim when older turns are compacted. You rarely need to change either.
Tool limits
These caps decide how much data any single tool call may return. They exist because analysis results are frequently far larger than a context window.
| Setting | Default | Ceiling |
|---|---|---|
| Batch size | 10 | 25 |
| Search results | 10 | 100 |
| Entities per risk | 20 | 100 |
| Overview top N | 10 | 100 |
| Aggregation groups | 20 | 100 |
| Detail items | 50 | 100 |
| Flat recommendations | 30 | 100 |
| Max characters per tool result | 24,000 | none |
The ceiling is a hard cap in the application. Asking the model for "all users" does not lift it. A tool that has more rows tells the model, and the model can request the next page. A tool result longer than the character limit is cut, and the model is told to narrow its query or page through the results. Use a lower character limit with a local model that has a small context window. Where completeness matters, scope the analysis narrowly first and ask the assistant about the smaller result. See Scoping Your Analysis.
Lower values produce faster, cheaper answers. Raise them when you need breadth and accept the token cost.