The first official certification for the Model Context Protocol, from the Agentic AI Foundation and LF Education. Prepare against the 2026-07-28 MCP spec, which is stateless (no initialize handshake or sessions) and deprecates roots, sampling and logging. Many existing courses still teach the older protocol. Works across Claude, OpenAI, Google and open-source agents.
Who it's for: Developers building agent tools and integrations, whichever model vendor they use.
LF's recommended background: JSON-RPC, experience calling an LLM API, agentic AI concepts, basic security literacy, and the ability to read MCP server manifests
Node.js or Python installed locally to run MCP servers and the MCP Inspector
Week 1 · 8 h
MCP fundamentals and architecture
Explain the problem MCP solves (N×M integrations) and the host, client and server roles
Describe the primitives (server-side tools, resources and prompts; client-side elicitation, plus the now-deprecated sampling and roots)
Read JSON-RPC 2.0 requests, results and errors, and tool inputSchema/outputSchema and structuredContent (JSON Schema 2020-12)
Compare the STDIO and Streamable HTTP transports, and know that HTTP+SSE is deprecated
Interactions and execution in the 2026-07-28 spec (largest domain)
Explain the stateless model, with no initialize handshake; protocol version and client capabilities travel in each request's _meta, and servers implement server/discover
Walk through the tool-invocation lifecycle, resultType "complete" vs "input_required", and Multi Round-Trip Requests (inputRequests/inputResponses)
Use subscriptions/listen for list-changed and resource notifications, and per-request logLevel instead of logging/setLevel
Handle errors (JSON-RPC error codes vs tool isError results, UnsupportedProtocolVersion) and caching hints (ttlMs, cacheScope)
Hands-on: Build a small MCP server with the Python or TypeScript SDK that exposes two tools (one returning structuredContent), one resource and one prompt. Call each one from MCP Inspector and save the raw JSON-RPC messages.
Map trust boundaries between host, client, server and upstream APIs, and explain user consent for tool calls
Explain MCP authorization (OAuth 2.1, protected resource metadata, Client ID Metadata Documents vs deprecated Dynamic Client Registration, iss validation)
Recognise attacks such as tool poisoning, prompt injection through tool output, confused deputy and token passthrough, and their mitigations
Design auditability with OpenTelemetry trace propagation in _meta and least-privilege tool scopes
Hands-on: Run your server over Streamable HTTP. Add a tool whose description contains a hidden instruction, and see how Inspector and a real host display it. Then write a three-item mitigation checklist (allowlisting, description review, human confirmation).
Write out what changed between the 2025-11-25 and 2026-07-28 specs, because older courses teach the removed initialize, ping and SSE resumability features
Re-drill Interactions & Execution (26%) and Security & Governance (24%), half the exam combined
Do a timed 90-minute self-test using your own question bank from the spec
You can sketch the full message flow of a tool call that needs user input under the 2026-07-28 spec, including the MRTR retry
You can list what 2026-07-28 removed or deprecated (initialize, ping, sessions, roots, sampling, logging, Dynamic Client Registration) and what replaced each
You have built and run your own MCP server over both STDIO and Streamable HTTP
You can name three MCP-specific attacks and a mitigation for each
Exam tips
LF says to prepare against the 2026-07-28 spec. Many tutorials and courses, including older Hugging Face and Anthropic material, describe the 2025 stateful protocol, so check every fact against the current spec.
The $250 exam includes one retake, eligibility lasts 12 months, and the certification is valid for 2 years.
Start with the provider's official exam guide and free training, then use practice exams to find your weak domains. We never link to exam dumps, which break exam agreements and can get you decertified.