Augmenting Agents with Tools and Data

Resources

• Writing tools for agents, not wrapping endpoints — Anthropic explains tool selection, namespacing, meaningful responses, context efficiency, actionable errors, and evaluation-driven iteration. This is the central reference for designing useful capabilities rather than exposing every API method unchanged. https://www.anthropic.com/engineering/writing-tools-for-agents

• MCP’s host–client–server architecture — The MCP architecture specification explains component responsibilities and the separation of server connections. Use it to locate configuration, context assembly, and trust boundaries before adding an external integration. https://modelcontextprotocol.io/specification/2025-11-25/architecture

• Connection setup and capability negotiation — MCP’s lifecycle specification describes initialization, version negotiation, capability exchange, and shutdown. It gives teachers a concrete debugging sequence for failures that occur before a tool can even be discovered or called. https://modelcontextprotocol.io/specification/2025-11-25/basic/lifecycle

• The MCP tool contract — The tools specification covers discovery, input and output schemas, structured results, annotations, and tool-level errors. It is the primary reference for reliable return contracts and for distinguishing recoverable execution errors from protocol failures. https://modelcontextprotocol.io/specification/2025-11-25/server/tools

• Resources are not tool calls — MCP’s resources specification defines addressable context, resource templates, listing, reading, and update notifications. Use it to decide when internal data should be exposed as a resource rather than hidden behind an action-shaped tool. https://modelcontextprotocol.io/specification/2025-11-25/server/resources

• Authorization for remote MCP — MCP’s authorization specification describes OAuth-based access to protected HTTP servers. It anchors the external-integration discussion in token audiences, authorization discovery, and scopes rather than embedding credentials in prompts or tool arguments. https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization

• MCP security failure modes — The protocol’s security guidance examines confused-deputy attacks, token passthrough, SSRF, and session-related risks. Use it to threat-model integrations and distinguish successful authentication from permission to perform a particular action. https://modelcontextprotocol.io/docs/2025-11-25/tutorials/security/security_best_practices

• Inspect the protocol before blaming the model — MCP Inspector provides interactive and command-line ways to exercise servers, inspect capabilities, and test requests. It is a practical reference for reproducing configuration, schema, transport, and result-format problems outside an agent conversation. https://modelcontextprotocol.io/docs/2026-07-28/tools/inspector

• The official TypeScript MCP SDK — The SDK repository supplies server and client implementations, examples, and transport documentation. Use it as the implementation reference for building and hosting internal tools, with attention to SDK versions and testing the deployed interface. https://github.com/modelcontextprotocol/typescript-sdk

• Command-line interfaces that compose — The Command Line Interface Guidelines cover help, structured output, stdout versus stderr, exit codes, configuration, and predictable behavior. These conventions explain why a well-designed CLI can be an excellent agent interface without making CLI and MCP interchangeable in every environment. https://clig.dev/

• Connecting a real harness to MCP — Claude Code documents local and remote server setup, configuration scopes, authentication, tool discovery, and troubleshooting. It grounds the lesson in practical integration work, including which settings belong to an individual and which can be shared with a project. https://code.claude.com/docs/en/mcp

• Code execution with MCP — Anthropic demonstrates discovering tool interfaces as files, composing calls in code, and filtering results before they enter model context. Use it to discuss large tool catalogs and data-heavy workflows while retaining the sandboxing and monitoring responsibilities that code execution introduces. https://www.anthropic.com/engineering/code-execution-with-mcp

• GitHub’s production MCP interface — GitHub’s official server documents local and hosted access, authentication, selectable toolsets, and read-only configuration. It is a concrete case for narrowing an integration to the task, reviewing permissions, and managing the difference between available tools and authorized operations. https://github.com/github/github-mcp-server