mcp-use
Version:
Opinionated MCP Framework for TypeScript (@modelcontextprotocol/sdk compatible) - Build MCP Agents, Clients and Servers with support for ChatGPT Apps, Code Mode, OAuth, Notifications, Sampling, Observability and more.
27 lines • 4.59 kB
TypeScript
import type { CallToolResult, Tool } from "@mcp-use/modelcontextprotocol-sdk/types.js";
import type { MCPClient } from "../../client.js";
import { BaseConnector } from "../../connectors/base.js";
export declare const CODE_MODE_AGENT_PROMPT = "\n## MCP Code Mode Tool Usage Guide\n\nYou have access to an MCP Code Mode Client that allows you to execute JavaScript/TypeScript code with access to registered tools. Follow this workflow:\n\n### 1. Tool Discovery Phase\n**Always start by discovering available tools:**\n- Tools are organized by server namespace (e.g., `server_name.tool_name`)\n- Use the `search_tools(query, detail_level)` function to find available tools\n- You can access `__tool_namespaces` to see all available server namespaces\n\n```javascript\n// Find all GitHub-related tools\nconst tools = await search_tools(\"github\");\nfor (const tool of tools) {\n console.log(`${tool.server}.${tool.name}: ${tool.description}`);\n}\n\n// Get only tool names for quick overview\nconst tools = await search_tools(\"\", \"names\");\n```\n\n### 2. Interface Introspection\n**Understand tool contracts before using them:**\n- Use `search_tools` to get tool descriptions and input schemas\n- Look for \"Access as: server.tool(args)\" patterns in descriptions\n\n### 3. Code Execution Guidelines\n**When writing code:**\n- Use `await server.tool({ param: value })` syntax for all tool calls\n- Tools are async functions that return promises\n- You have access to standard JavaScript globals: `console`, `JSON`, `Math`, `Date`, etc.\n- All console output (`console.log`, `console.error`, etc.) is automatically captured and returned\n- Build properly structured input objects based on interface definitions\n- Handle errors appropriately with try/catch blocks\n- Chain tool calls by using results from previous calls\n\n### 4. Best Practices\n- **Discover first, code second**: Always explore available tools before writing execution code\n- **Respect namespaces**: Use full `server.tool` names to avoid conflicts\n- **Minimize Context**: Process large data in code, return only essential results\n- **Error handling**: Wrap tool calls in try/catch for robustness\n- **Data flow**: Chain tools by passing outputs as inputs to subsequent tools\n\n### 5. Available Runtime Context\n- `search_tools(query, detail_level)`: Function to discover tools\n- `__tool_namespaces`: Array of available server namespaces\n- All registered tools as `server.tool` functions\n- Standard JavaScript built-ins for data processing\n\n### Example Workflow\n\n```javascript\n// 1. Discover available tools\nconst github_tools = await search_tools(\"github pull request\");\nconsole.log(`Available GitHub PR tools: ${github_tools.map(t => t.name)}`);\n\n// 2. Call tools with proper parameters\nconst pr = await github.get_pull_request({\n owner: \"facebook\",\n repo: \"react\",\n number: 12345\n});\n\n// 3. Process results\nlet result;\nif (pr.state === 'open' && pr.labels.some(l => l.name === 'bug')) {\n // 4. Chain with other tools\n await slack.post_message({\n channel: \"#bugs\",\n text: `\uD83D\uDC1B Bug PR needs review: ${pr.title}`\n });\n result = \"Notification sent\";\n} else {\n result = \"No action needed\";\n}\n\n// 5. Return structured results\nreturn {\n pr_number: pr.number,\n pr_title: pr.title,\n action_taken: result\n};\n```\n\nRemember: Always discover and understand available tools before attempting to use them in code execution.\n";
/**
* CodeModeConnector provides a special "code mode" interface for executing JavaScript/TypeScript
* code with access to MCP tools. Unlike other connectors, it doesn't establish its own external
* connection - instead, it wraps an already-connected BaseMCPClient and exposes special tools
* (execute_code, search_tools) on top of it.
*
* Since there's no connection phase to perform, the connector is immediately ready to use
* (connected=true) upon construction. The connect() and disconnect() methods exist for
* lifecycle compatibility with the BaseConnector interface but don't manage actual connections.
*/
export declare class CodeModeConnector extends BaseConnector {
private mcpClient;
private _tools;
constructor(client: MCPClient);
connect(): Promise<void>;
disconnect(): Promise<void>;
get publicIdentifier(): Record<string, string>;
private _createToolsList;
get tools(): Tool[];
initialize(): Promise<any>;
callTool(name: string, args: Record<string, any>): Promise<CallToolResult>;
}
//# sourceMappingURL=codeMode.d.ts.map