# Tool Catalog — xai-toolkit MCP Server
All tools return a consistent `ToolResponse` structure (see ADR-005):
```json
{
"narrative": "Plain English interpretation...",
"evidence": { ... structured data ... },
"metadata": { "model_id": ..., "timestamp": ..., "tool_version": ..., "data_hash": ... },
"plot_base64": null,
"grounded": true
}
```
`grounded: true` is present on every tool response. It signals to the LLM that
this answer was computed deterministically from a registered model and is
audit-ready. If the LLM answers a question without calling a tool (general
knowledge), it is instructed to prepend a ⚠️ disclaimer — because no
`grounded: true` field will be present in that case.
---
## explain_prediction
**Purpose:** Explain why a single sample received its classification.
**Trigger:** "Why was sample X classified as Y?" / "What drove this prediction?"
| Parameter | Type | Required | Description |
|-------------|--------|----------|------------------------------------------|
| model_id | string | yes | ID of a registered model |
| sample_index| int | yes | Row index in the test dataset to explain |
**Narrative includes:** predicted class, probability, top 3 contributing
features with direction and magnitude, top opposing factor.
---
## summarize_model
**Purpose:** Provide a high-level overview of what a model does.
**Trigger:** "What does this model do?" / "Summarize the model"
| Parameter | Type | Required | Description |
|----------|--------|----------|--------------------------|
| model_id | string | yes | ID of a registered model |
**Narrative includes:** model type, accuracy, feature count, target classes,
top 5 features by importance.
---
## compare_features
**Purpose:** Rank features by global importance.
**Trigger:** "Which features matter most?" / "Compare feature importance"
| Parameter | Type | Required | Default | Description |
|----------|--------|----------|---------|--------------------------|
| model_id | string | yes | | ID of a registered model |
| top_n | int | no | 10 | Number of features to rank |
**Narrative includes:** ranked list with magnitude, direction, comparative
language (e.g., "1.5× more influential").
---
## get_partial_dependence
**Purpose:** Show how a single feature affects predictions.
**Trigger:** "How does [feature] affect predictions?" / "What happens when [feature] changes?"
| Parameter | Type | Required | Description |
|-------------|--------|----------|--------------------------------|
| model_id | string | yes | ID of a registered model |
| feature_name| string | yes | Feature to analyze |
**Narrative includes:** feature range, prediction range, overall trend,
steepest change region.
**Error handling:** Typos in feature_name return suggestions via fuzzy matching.
---
## list_models
**Purpose:** Discover what models are available.
**Trigger:** "What models are available?" / "List models"
No parameters required.
**Narrative includes:** count, model IDs, types, feature counts, accuracy.
---
## describe_dataset
**Purpose:** Describe the data associated with a model.
**Trigger:** "Tell me about the data" / "Describe the dataset"
| Parameter | Type | Required | Description |
|----------|--------|----------|--------------------------|
| model_id | string | yes | ID of a registered model |
**Narrative includes:** sample count, feature count, class distribution,
missing value count.
