Learn
Learn MCP and AI agents, one track at a time
Short guides in reading order. Each one explains one idea in plain words, shows it in L2M, and most link to something you can run.
Start here
- 1
GuideWhat is MCP?MCP explained: hosts, clients and servers, the tools, resources and prompts a server offers, how a tool call travels, and a runnable agent to try.
- 2
GuideWhat is an AI agent?An AI agent is a model in a loop that calls tools until it can answer. See the loop step by step, where it fails, and run one locally with no API key.
- 3
- 4
GuidePrompt injectionDirect and indirect prompt injection explained with agent examples, why filters alone fail, and the layered defenses OWASP and recent research recommend.
- 5
GuideRegression testing agentsCatch a prompt, model or tool change that breaks your agent before users do: save known-good runs, replay changes against recordings, gate CI.
Tracks
01Beginner6 guides · about 52 min
What MCP is, its host, client and server roles, tools, resources and prompts, the transports, OAuth for remote servers, and tool annotations.
- What is MCP?MCP explained: hosts, clients and servers, the tools, resources and prompts a server offers, how a tool call travels, and a runnable agent to try.
- Hosts, clients and serversThe three MCP roles in plain terms: what the host app, each client connection and each server own, local vs remote servers, and where credentials live.
- Tools, resources and promptsWhen an MCP server should offer a tool, a resource or a prompt, who decides each is used, and what output schemas, tool errors and elicitation add.
- MCP transportsHow MCP messages travel: stdio for local servers, Streamable HTTP for remote ones, sessions, Origin checks, and what the stateless 2026-07-28 spec changes.
- MCP authorizationHow a remote MCP server asks for OAuth: the 401 challenge, resource metadata, PKCE, the resource parameter, and why tokens must never be passed through.
- MCP tool annotationsWhat readOnlyHint, destructiveHint, idempotentHint and openWorldHint mean, why the spec calls them untrusted, and how to use them in tool rules safely.
Start the track: MCP fundamentals →02Beginner6 guides · about 53 min
The tool-calling loop, when a fixed workflow beats an agent, supervisor and worker agents, memory, long tool lists and human approval.
- What is an AI agent?An AI agent is a model in a loop that calls tools until it can answer. See the loop step by step, where it fails, and run one locally with no API key.
- Workflows vs agentsFixed workflows and model-driven agents compared: cost, control and testability, the common patterns between them, and how to mix both in one graph.
- Supervisor and workersHow a supervisor agent delegates to worker agents exposed as tools: when it helps, what it costs, how to test it, and a sample that runs with no API key.
- Agent memoryShort-term history, session memory and cached tool results: what each is for, how much to keep, and how agent memory meets retention and erasure.
- Too many toolsWhy long MCP tool lists hurt agents (context, wrong picks, cost), and practical fixes: shortlisting, toolsets, per-agent servers and rules that hide tools.
- Human in the loopWhere a person belongs in an agent's path: before risky tool calls, on a draft, or as a question. Which actions need it, timeouts, and approval fatigue.
Start the track: How AI agents work →03Intermediate6 guides · about 56 min
Prompt injection, the OWASP Top 10 for LLM apps, the lethal trifecta, MCP risks, least privilege, and keeping personal data out of prompts.
- Prompt injectionDirect and indirect prompt injection explained with agent examples, why filters alone fail, and the layered defenses OWASP and recent research recommend.
- OWASP Top 10 for LLM appsThe ten OWASP LLM risks for 2025, one by one, with what each looks like in an agent that calls tools and a concrete control for each one.
- The lethal trifectaWhy an agent with private data, untrusted content and a way to send data out is one prompt injection from a leak, and how to break the combination.
- MCP security risksThe main MCP security risks from the spec and OWASP: tool poisoning, rug pulls, shadowing, confused deputy, token passthrough, SSRF and their controls.
- Least privilege for agentsWhy AI agents drift to more access than they need, the layers to narrow (servers, tools, arguments, data, people, keys) and how to keep them narrow.
- Sensitive data in promptsWhere personal data and secrets leak in an AI app (prompts, tool arguments, logs, answers) and how masking, tokens and a controlled reveal contain them.
Start the track: Agent security →04Intermediate4 guides · about 34 min
Evals, regression tests built from recorded runs, testing MCP tool calls, and release gates that ship only what passed.
- Agent evalsWhat an agent eval is: tasks, trials, graders and transcripts, capability vs regression evals, pass@k vs pass^k, and how to write your first 20 cases.
- Regression testing agentsCatch a prompt, model or tool change that breaks your agent before users do: save known-good runs, replay changes against recordings, gate CI.
- Testing MCP toolsWhy live MCP calls make agent tests flaky, and what recorded replays, schema-built mocks and breaking-change checks each catch and miss.
- Release gatesA release gate lets a workflow version reach production only after its tests pass on that exact version. How gates, drafts, overrides and rollback work.
Start the track: Testing and shipping agents →05Intermediate6 guides · about 45 min
LLM and MCP gateways, cost caps, self-hosting, retention and erasure, audit trails, and identity: SSO, MFA, roles and keys.
- LLM gateway vs MCP gatewayAn LLM gateway sits between apps and models; an MCP gateway sits between agents and tools. What each controls, where they overlap, and when you need both.
- LLM and agent cost controlWhy agent costs run away (loops, long tool lists), where to put spend and token budgets (user, project, client), warn vs block, and estimating usage.
- Self-hosting AI agentsWhat a self-hosted agent platform runs (API, workers, database, models), the trade-offs against hosted tools, and a checklist for a first production setup.
- Data retention and erasureWhere an AI app keeps personal data, how retention purges it, how to answer access and erasure requests, and how legal holds stop deletion.
- Audit trails for agentsWhat an agent audit trail should record (inputs, tool calls, approvals, versions), how a signed run manifest proves it is intact, and offline checks.
- Identity for AI platformsWho may build, run, approve and call agents: single sign-on with OIDC or SAML, MFA per role, project roles, service accounts and scoped API keys.
Start the track: Running AI in production →
Try it and look things up
The real studio with recorded runs, in your browser. No install, no sign-up.
Working workflows to download and import, with a diagram of every node.
Short, plain definitions of the terms the guides use.