What to have ready.

• Record your plan, workspace role, approved training account, exact actions available and spending limit. Check the actual tool list before planning a write.
• For custom MCP in ChatGPT, Pro supports read/fetch. Full MCP on web requires an eligible Business or Enterprise/Edu workspace and authorization. In the current Plugins interface, use Add custom MCP server with an approved HTTPS endpoint or Secure MCP Tunnel.
• For an existing signed-in desktop tab, verify the desktop app, supported browser extension and site permission. Check the approved sandbox separately.
• If access is missing, use the local engineering starter or arrange supervised execution with a technical partner. Use fictional cases to practise design while waiting. Keep simulations separate from observed results at the practical review gate.

Verify current MCP access and setup

Verify desktop browser setup

Use the fictional accounts in the Academy delivery lab or the downloadable starter. This exercise needs no external course or evaluation service.

1. Open evaluation.csv in the starter archive. Before running anything, predict outcomes for an allowed read, a restricted read, a write without approval and a repeated approved write.
2. Run each case. Use the lab results or run node --test and the starter’s documented client. Record actual outcomes and retain sanitized traces. Check the stored outcome when testing a retry.
3. Keep unauthorized reads and writes, and duplicate changes, in a separate critical-failure column. Investigate any critical failure before recommending release.
4. Hold back one changed-input case. In Week 10, rerun after one controlled change and compare results against the same criteria.

Keep the completed CSV, expected and actual outcomes, relevant traces and your release decision with your project evidence.

Both tracks

Use your own ChatGPT account. Start with available basic chat and document features; check limits before upgrading.

1. Create a separate project conversation with synthetic records from your plan. Define the user, source boundary, output format and escalation rule.
2. Ask one supported question, one unknown question and one conflicting-source question. Save the actual responses and citations.
3. Compare an unstructured request with your revised instructions. Identify a remaining failure and explain how you would test it.

Required evidence: Submit your prompt, three observed outputs, source mapping, revision and a short explanation in your own words.

Both tracks

Use an app available to your ChatGPT account. Permissions and features vary by plan. Only connect an account you own or are authorized to use.

1. Choose a read-only connection to a synthetic document or training folder. Inspect the permissions before connecting. Do not authorize company-wide data access for a practice exercise.
2. Ask ChatGPT to retrieve a specific synthetic fact, cite its source and distinguish a missing result from an unavailable connection.
3. Disconnect or restrict the connection, retry the request and record what changes. Do not use a live write action in a real business system.

Required evidence: Submit the app used, permission scope, observed retrieval output, citation and a denied/disconnected result. If access is unavailable, record that blocker and complete a supervised equivalent; do not claim a successful connection.

Engineering required; practitioners use a supported hosted or supervised client

The downloadable Node.js 22+ lab uses MCP 2025-11-25 compatibility and works locally without an OpenAI API key. Connecting it to ChatGPT requires an eligible account and an authorized remote HTTPS endpoint or Secure MCP Tunnel.

1. Engineers: run node --test, then LAB_TOKEN=<local training value> node mcp-server.mjs and run the provided mcp-client.mjs in a second terminal with the same local token. The README explains supported protocol scope.
2. Inspect tools/list before calling read_service_status and create_support_ticket. Compare the declared input contracts with actual outcomes.
3. Test an allowed read, denied account, missing approval and a repeated idempotency key. Capture the actual tool trace and distinguish protocol errors from business denials.
4. Practitioners: complete the same discovery and outcome review in an available approved MCP client or a supervised clinic. Identify the actor, permissions, approval and handoff. Custom MCP access must be confirmed before promising the clinic.

Required evidence: Submit tool discovery, four observed outcomes, permissions, approval evidence and an explanation of a retry. A simulated plan alone does not satisfy the engineering MCP exercise.

Download MCP 2025-11-25 compatibility lab. The July 2026 specification uses a stateless protocol lifecycle. Run this archive with its bundled client.

Both tracks, deeper implementation for engineers

Your personal ChatGPT subscription and OpenAI API billing are separate. The in-app coach uses NeoAistriq-provided API access within its allowance.

1. Draw the boundary between the ChatGPT user interface, MCP client/server, model API, application backend and system of record.
2. Explain where credentials, authorization, approval, usage limits and audit records live.
3. For engineers, demonstrate one server-side model request in your own approved environment and document cost/response time. For practitioners, explain the configuration and engineering handoff using observed lab evidence.

Required evidence: Submit an environment diagram, tool ownership/cost sheet and a failure/rollback plan. Never submit API keys or real customer records.

Both tracks, with different implementation depth

Use an authorized training account or sandbox with a supported app/MCP connection. CRM licenses, connector permissions and API access are not automatically included. Confirm access before planning the exercise.

1. Choose one limited change: a lead-routing rule, required qualification fields, a follow-up workflow or a reporting view. Use synthetic contacts and define a baseline acceptance test.
2. Ask ChatGPT to inspect available tools and propose the configuration. Check the target environment, objects, permissions and exact intended change before approving a write. Start with read-only inspection.
3. Apply the approved change only in the sandbox using available MCP tools. Verify it in the app itself with an allowed case and a denied or invalid case. Record the actual before/after state.
4. Test a duplicate request, incorrect field mapping and rollback. Engineering learners show tool inputs and outputs, required data format and authentication code; industry learners show the configuration, acceptance results and handoff.

Required evidence: Submit the change brief, discovered tools, approved action, before/after configuration, actual acceptance results and rollback evidence. If a connector does not expose the needed operation, document that limit and use the authorized browser path if available, or arrange supervised execution. Record design-only simulation separately from actual results.

Both tracks

Use browser automation available in your own AI tool and an authorized sandbox. An Academy login does not itself provide a browser agent or a CRM license.

1. Write the intended change and success criteria before opening the app. Identify read, draft, approval, submit and verify steps.
2. Have the assistant inspect the actual page and prepare the change. Check the record, environment and values before any consequential submission. Credentials and verification codes stay in the approved secure sign-in flow.
3. Apply the approved change, then independently inspect the resulting app state. A clicked button is not proof of success. Capture a failed validation and recovery.
4. Repeat after a harmless navigation or layout variation. Explain when an API/MCP operation is more reliable than browser interaction, and document an escalation and rollback.

Required evidence: Submit a sanitized observed trace or screenshots, approval point, resulting configuration, failed case, recovery and choice of integration method.

Both tracks; revisit in weeks 4 and 12

Choose an authorized low-risk task and synthetic or non-confidential material. Employer policies and account permissions apply.

1. Choose one repeated task such as meeting preparation, a CRM account brief, weekly reporting, document comparison or an SOP draft. Record how you currently do it.
2. Measure a baseline across several representative examples: minutes spent, correction effort, missing facts and an output-quality check.
3. Run the same task with ChatGPT, an authorized connected app or a controlled browser workflow. Include prompting, verification, corrections and tool costs in the assisted time.
4. Compare the results using the same quality standard. Report failures and limitations, then create a reusable instruction and checklist. Do not extrapolate a percentage gain from a single successful demo.

Required evidence: Submit a baseline/assisted comparison table, sample outputs, quality criteria, correction effort, recurring instructions and an honest decision about whether the change is useful in your existing job.