AGENTS, TOOLS, AND DATA ACCESS

Chatbot: Responds to prompts using only its training data. Can't take actions, access external information, or modify files. Like talking to someone who can only answer from memory.

Agent: Can use tools, access data sources, and take actions. Breaks complex tasks into steps, decides what tools to use, and executes multi-step plans. Claude Code is an agent -- it can read files, run commands, and make changes to your projects.

Running claude -p "your prompt here" sends a single prompt without entering an interactive session. Claude processes the prompt, prints the result to stdout, and exits. This is what makes scripted pipelines and autonomous workflows possible.

Practical uses: a shell script that runs claude -p on every file in a folder, a cron job that checks data daily, or piping output from one command into another. Once you can call Claude from a script, you can build workflows that run without you.

Agents don't just answer questions -- they break work into steps and execute each one. Claude Code can read a file, decide what needs to change, make the edit, run a test, and fix problems it finds along the way. Each step informs the next.

This also means agents can delegate. Claude Code can spawn subagents to handle parts of a larger task in parallel. The risk: errors in early steps become assumed facts in later ones. The skill is knowing how to verify intermediate results, not just the final output.

Language models only know what's in their training data. To work with your files, your databases, or live information from the web, the model needs a connection layer. Model Context Protocol (MCP) is an open standard for this -- it lets you plug data sources into Claude the same way you'd add plugins to a browser.

This is sometimes called retrieval-augmented generation (RAG): the model retrieves relevant information from an external source and uses it to generate a grounded response instead of relying on memory alone. MCP is how you set up that retrieval in practice.

Every tool you connect costs context. Each MCP server you configure consumes tokens in your session before you've asked anything -- the server's schema, its available tools, its connection overhead. Before connecting a data source, ask whether the capability it adds justifies what it takes from your context budget. A database connection you rarely need still consumes context on every session where you don't use it. A lean, purposeful set of integrations outperforms a crowded one.

Setting up an MCP server is straightforward. Debugging it when things go wrong takes practice. Common failures:

• > Permissions: The server can't read the files you pointed it at, or the database user lacks access to the table you need
• > Auth failures: API keys expired, tokens revoked, OAuth scopes too narrow -- the connection looks configured but returns nothing useful
• > Schema mismatches: The data source changed its column names, field types, or API response format since you set things up
• > Silent failures: The server starts, Claude tries to use it, but gets empty results or malformed data -- and keeps going without flagging the problem

The pattern for journalists: when Claude gives you an answer based on external data, check the source. Ask it to show you the raw data it retrieved. If it can't, the connection may be broken and the answer may be fabricated from training data instead.

Reroute NJ is a public transit guide built during a real infrastructure disruption in New Jersey -- 11 languages, zero build step, serving a real community. The project applies concepts from all four modules: a CLAUDE.md for project context (Module 1), reusable scripts for page generation and translation (Module 2), a cron-scheduled scraper that commits and pushes automatically (Module 3), and autonomous data connections to external sources like news sites and NJ Transit alerts (Module 4).

The Jay Rosen digital archive is a collection of 765+ records from journalist Jay Rosen's career. The project applies entity extraction and grounded knowledge retrieval to primary source materials, turning raw documents into a searchable knowledge base.