AI Tools

AI tools let Claude interact with your extension programmatically. When you add tools, Claude can read data, make changes, and help users work with your custom file types.

Why Add AI Tools?

Without tools, Claude can only:

• Read the raw file content

• Suggest edits to the raw content

With tools, Claude can:

• Query structured data ("What columns are in this spreadsheet?")

• Make targeted changes ("Add a row with these values")

• Perform complex operations ("Sort by the date column")

• Understand your data model ("What entities are defined?")

Tool Definition Structure

Tools are defined in your extension's entry point:

// src/index.ts
import type { ExtensionAITool } from '@nimbalyst/extension-sdk';

export const aiTools: ExtensionAITool[] = [
  {
    name: 'my_tool_name',
    description: 'What this tool does - Claude reads this to decide when to use it',
    inputSchema: {
      type: 'object',
      properties: {
        param1: {
          type: 'string',
          description: 'Description of param1',
        },
        param2: {
          type: 'number',
          description: 'Description of param2',
        },
      },
      required: ['param1'],
    },
    handler: async (args, context) => {
      // Implement tool logic
      return { result: 'success' };
    },
  },
];

Registering Tools in the Manifest

Add tools to your manifest.json:

{
  "permissions": {
    "ai": true
  },
  "contributions": {
    "aiTools": [
      "myext.get_data",
      "myext.update_data"
    ]
  }
}

Tool Handler Context

The handler receives a context object with useful information:

interface AIToolContext {
  // Path to the current workspace (if any)
  workspacePath?: string;

  // Path to the active file (if any)
  activeFilePath?: string;

  // Access host services such as filesystem, UI, and AI helpers
  extensionContext: ExtensionContext;
}

Example: Spreadsheet Tools

Here's a complete example for a CSV/spreadsheet editor:

import type {
  AIToolContext,
  ExtensionAITool,
  ExtensionToolResult,
} from '@nimbalyst/extension-sdk';

async function loadActiveFile(context: AIToolContext): Promise<{
  filePath: string;
  content: string;
} | ExtensionToolResult> {
  if (!context.activeFilePath) {
    return { success: false, error: 'No active file is open.' };
  }

  try {
    const content = await context.extensionContext.services.filesystem.readFile(context.activeFilePath);
    return {
      filePath: context.activeFilePath,
      content,
    };
  } catch (error) {
    return {
      success: false,
      error: `Failed to read active file: ${error instanceof Error ? error.message : String(error)}`,
    };
  }
}

// Helper to parse CSV
function parseCSV(content: string): string[][] {
  return content.split('\n').map(row => row.split(','));
}

export const aiTools: ExtensionAITool[] = [
  {
    name: 'csv.get_schema',
    description: 'Get the column names and row count of the current CSV file',
    scope: 'global',
    inputSchema: {
      type: 'object',
      properties: {},
    },
    handler: async (_args, context) => {
      const loaded = await loadActiveFile(context);
      if ('success' in loaded) {
        return loaded;
      }

      const rows = parseCSV(loaded.content);
      const headers = rows[0] || [];

      return {
        success: true,
        data: {
          columns: headers,
          rowCount: rows.length - 1,
          filePath: loaded.filePath,
        },
      };
    },
  },

  {
    name: 'csv.get_rows',
    description: 'Get rows from the CSV file. Returns data as objects with column names as keys.',
    inputSchema: {
      type: 'object',
      properties: {
        startRow: {
          type: 'number',
          description: 'Starting row index (0-based, excluding header)',
        },
        count: {
          type: 'number',
          description: 'Number of rows to return (default: 10)',
        },
      },
    },
    handler: async (args, context) => {
      const loaded = await loadActiveFile(context);
      if ('success' in loaded) {
        return loaded;
      }

      const rows = parseCSV(loaded.content);
      const headers = rows[0] || [];
      const dataRows = rows.slice(1);

      const start = typeof args.startRow === 'number' ? args.startRow : 0;
      const count = typeof args.count === 'number' ? args.count : 10;
      const selectedRows = dataRows.slice(start, start + count);

      return {
        success: true,
        data: {
          rows: selectedRows.map(row => {
            const obj: Record<string, string> = {};
            headers.forEach((h, i) => {
              obj[h] = row[i] || '';
            });
            return obj;
          }),
          totalRows: dataRows.length,
        },
      };
    },
  },

  {
    name: 'csv.add_row',
    description: 'Add a new row to the CSV file',
    inputSchema: {
      type: 'object',
      properties: {
        data: {
          type: 'object',
          description: 'Object with column names as keys and cell values',
        },
      },
      required: ['data'],
    },
    handler: async (args, context) => {
      const loaded = await loadActiveFile(context);
      if ('success' in loaded) {
        return loaded;
      }

      const rows = parseCSV(loaded.content);
      const headers = rows[0] || [];

      // Build new row from data object
      const values = (args.data as Record<string, string>) || {};
      const newRow = headers.map(h => values[h] || '');
      rows.push(newRow);

      const nextContent = rows.map(r => r.join(',')).join('\n');
      await context.extensionContext.services.filesystem.writeFile(
        loaded.filePath,
        nextContent
      );

      return {
        success: true,
        message: `Added a row to ${loaded.filePath}.`,
        data: {
          rowIndex: rows.length - 1,
        },
      };
    },
  },
];

Updating File Content

When a tool needs to modify a file, write through the filesystem service:

handler: async (args, context) => {
  // ... modify data ...

  if (!context.activeFilePath) {
    return { success: false, error: 'No active file is open.' };
  }

  await context.extensionContext.services.filesystem.writeFile(
    context.activeFilePath,
    serializedData
  );

  return {
    success: true,
    message: 'Row added successfully',
  };
}

Nimbalyst will:

1. Persist the updated file content

2. Notify the active editor through file watching

3. Let the editor reload or reconcile its in-memory state

Tool Naming Conventions

Use a prefix for your tools to avoid conflicts:

extensionname.action_name

Examples:

• csv.get_schema

• csv.add_row

• diagram.add_node

• datamodel.get_entities

Writing Good Tool Descriptions

Claude uses the description to decide when to use your tool. Be specific:

Good:

description: 'Get the column names and data types from the current CSV file. Returns an array of column definitions.'

Bad:

description: 'Get schema'  // Too vague

Error Handling

Return errors as objects, not thrown exceptions:

handler: async (args, context) => {
  if (!context.activeFilePath) {
    return { success: false, error: 'No active file is open' };
  }

  if (!args.columnName) {
    return { success: false, error: 'columnName parameter is required' };
  }

  try {
    // ... do work ...
    return { success: true, data: result };
  } catch (e) {
    return { error: `Failed to process: ${e.message}` };
  }
}

Input Schema

The inputSchema follows JSON Schema format:

inputSchema: {
  type: 'object',
  properties: {
    // String parameter
    name: {
      type: 'string',
      description: 'The name to use',
    },

    // Number parameter
    count: {
      type: 'number',
      description: 'How many items',
    },

    // Boolean parameter
    includeHeaders: {
      type: 'boolean',
      description: 'Whether to include header row',
    },

    // Enum parameter
    format: {
      type: 'string',
      enum: ['json', 'csv', 'xml'],
      description: 'Output format',
    },

    // Array parameter
    columns: {
      type: 'array',
      items: { type: 'string' },
      description: 'List of column names',
    },

    // Object parameter
    options: {
      type: 'object',
      properties: {
        sortBy: { type: 'string' },
        ascending: { type: 'boolean' },
      },
    },
  },
  required: ['name'], // Required parameters
}

Best Practices

1. Keep tools focused - One tool, one job

2. Return structured data - Objects are easier for Claude to work with

3. Include context in responses - Return relevant metadata

4. Handle missing files gracefully - Check if activeFilePath exists and read through the filesystem service

5. Validate inputs - Check required parameters

6. Use descriptive names - get_column_stats not stats

Testing Tools

Test your tools by asking Claude to use them:

"What columns are in this CSV file?"

Claude should invoke your csv.get_schema tool and report the results.

Example: Data Model Tools

For a more complex example, here are tools for a data modeling extension:

export const aiTools: ExtensionAITool[] = [
  {
    name: 'datamodel.get_entities',
    description: 'List all entities (tables/models) defined in the data model',
    inputSchema: { type: 'object', properties: {} },
    handler: async (_args, context) => {
      if (!context.activeFilePath) {
        return { success: false, error: 'No active file is open' };
      }

      const content = await context.extensionContext.services.filesystem.readFile(
        context.activeFilePath
      );
      const model = parseDataModel(content);

      return {
        success: true,
        data: {
          entities: model.entities.map(e => ({
            name: e.name,
            fieldCount: e.fields.length,
          })),
        },
      };
    },
  },

  {
    name: 'datamodel.get_entity',
    description: 'Get detailed information about a specific entity',
    inputSchema: {
      type: 'object',
      properties: {
        name: { type: 'string', description: 'Entity name' },
      },
      required: ['name'],
    },
    handler: async (args, context) => {
      if (!context.activeFilePath) {
        return { success: false, error: 'No active file is open' };
      }

      const content = await context.extensionContext.services.filesystem.readFile(
        context.activeFilePath
      );
      const model = parseDataModel(content);
      const entity = model.entities.find(e => e.name === args.name);

      if (!entity) {
        return { success: false, error: `Entity '${args.name}' not found` };
      }

      return {
        success: true,
        data: {
          name: entity.name,
          fields: entity.fields.map(f => ({
            name: f.name,
            type: f.type,
            required: f.required,
          })),
          relations: entity.relations,
        },
      };
    },
  },

  {
    name: 'datamodel.add_field',
    description: 'Add a new field to an entity',
    inputSchema: {
      type: 'object',
      properties: {
        entityName: { type: 'string' },
        fieldName: { type: 'string' },
        fieldType: { type: 'string' },
        required: { type: 'boolean' },
      },
      required: ['entityName', 'fieldName', 'fieldType'],
    },
    handler: async (args, context) => {
      if (!context.activeFilePath) {
        return { success: false, error: 'No active file is open' };
      }

      const content = await context.extensionContext.services.filesystem.readFile(
        context.activeFilePath
      );
      const model = parseDataModel(content);
      const entity = model.entities.find(e => e.name === args.entityName);

      if (!entity) {
        return { success: false, error: `Entity '${args.entityName}' not found` };
      }

      entity.fields.push({
        name: args.fieldName,
        type: args.fieldType,
        required: args.required ?? false,
      });

      await context.extensionContext.services.filesystem.writeFile(
        context.activeFilePath,
        serializeDataModel(model)
      );

      return {
        success: true,
        message: `Added ${args.fieldName} to ${args.entityName}.`,
      };
    },
  },
];

Calling AI Models Directly

Extensions can also call AI chat/completion models directly, without going through Claude. This is useful for summarization, classification, code generation, or any task where the extension itself needs an AI response.

Prerequisites

Your manifest must declare permissions.ai: true.

Listing Available Models

export async function activate(context: ExtensionContext) {
  const models = await context.services.ai!.listModels();
  // => [
  //   { id: "claude:claude-sonnet-4-6-20250514", name: "Claude Sonnet 4.6", provider: "claude" },
  //   { id: "openai:gpt-4o", name: "GPT-4o", provider: "openai" },
  //   ...
  // ]
}

Only models from chat providers the user has enabled and configured are returned (Claude, OpenAI, LM Studio). Agent providers like Claude Code are not included.

Non-Streaming Completion

const result = await context.services.ai!.chatCompletion({
  messages: [
    { role: 'user', content: 'Classify this text as positive or negative: "Great product!"' },
  ],
  model: 'claude:claude-sonnet-4-6-20250514', // optional, uses default if omitted
  systemPrompt: 'Respond with a single word: positive or negative.',
  temperature: 0,
  maxTokens: 10,
});

console.log(result.content);  // "positive"
console.log(result.model);    // "claude-sonnet-4-6-20250514"
console.log(result.usage);    // { inputTokens: 42, outputTokens: 1 }

Streaming Completion

For longer responses where you want to show results incrementally:

const handle = await context.services.ai!.chatCompletionStream({
  messages: [
    { role: 'user', content: 'Write a haiku about programming' },
  ],
  onChunk: (chunk) => {
    if (chunk.type === 'text') {
      // Append text to your UI
      appendToOutput(chunk.content!);
    } else if (chunk.type === 'error') {
      showError(chunk.error!);
    }
    // chunk.type === 'done' means the stream is complete
  },
});

// Optionally abort:
// handle.abort();

// Wait for the full result:
const result = await handle.result;
console.log(result.content); // Full response text

Multi-Turn Conversations

Pass multiple messages for conversation context:

const result = await context.services.ai!.chatCompletion({
  messages: [
    { role: 'user', content: 'What is the capital of France?' },
    { role: 'assistant', content: 'The capital of France is Paris.' },
    { role: 'user', content: 'What is its population?' },
  ],
});

Structured Output (JSON Mode)

Use responseFormat to constrain the model's output to valid JSON:

// Simple JSON mode - model returns valid JSON
const result = await context.services.ai!.chatCompletion({
  messages: [{ role: 'user', content: 'List 3 colors with hex codes' }],
  systemPrompt: 'Respond in JSON format.',
  responseFormat: { type: 'json_object' },
});
const data = JSON.parse(result.content);

For stricter control, use json_schema to enforce a specific shape:

const result = await context.services.ai!.chatCompletion({
  messages: [{ role: 'user', content: 'Classify this issue: login page crashes on Safari' }],
  responseFormat: {
    type: 'json_schema',
    schema: {
      type: 'object',
      properties: {
        category: { type: 'string', enum: ['bug', 'feature', 'question'] },
        severity: { type: 'string', enum: ['low', 'medium', 'high', 'critical'] },
        component: { type: 'string' },
      },
      required: ['category', 'severity', 'component'],
    },
  },
});
const classification = JSON.parse(result.content);
// => { category: "bug", severity: "high", component: "auth" }

Key Points

• Stateless: These calls do not create sessions in the session history. Each call is independent.

• Model selection: Use listModels() to discover available models, then pass an id to chatCompletion() or chatCompletionStream(). If you omit the model, the first available provider's default is used.

• Chat providers only: Claude, OpenAI, and LM Studio. Agent providers (Claude Code, Codex) are not available through this API.

• User configuration: The API respects the user's provider settings and API keys. If a provider is disabled or unconfigured, its models won't appear in listModels().

Next Steps

• See custom-editors.md to build the visual component

• Check manifest-reference.md for all configuration options

• Check the built-in extensions in packages/extensions/ in the Nimbalyst repository for production examples

• See api-reference.md for full type definitions