Awesome Testing

MCP lesson 08 · Ship a modern MCP system

Can the host and server survive scaling, migration, hostile content, and long-running work?

2026-07-28 core · deterministic JSON-RPC fixtures

Atlas Release Evidence Server. Read-only evidence, deterministic scoring, and an explicit evaluation Task across replicas and protocol eras.

01 · Smallest useful mechanism

Integrate modern stateless MCP without hiding legacy or product risk.

The capstone combines semantic primitives, stdio and HTTP, per-request metadata, discovery, explicit handles, MRTR, Tasks, authorization, hostile-content controls, telemetry, and a compatibility suite.

Ship one coherent modern path, one explicit legacy path if required, and one product acceptance boundary around both.

02 · Experiment

Test the prediction

Deterministic protocol fixture

This workbench uses inspectable JSON-RPC fixtures and injected failures. It does not make live calls or execute external effects.

Protocol workbench · atlas-capstone-2026-07-28

Choose the system ready for a bounded internal pilot.

Injected failure

All candidates list tools; only one survives replica changes, hostile content, and cancellation races.

Choose the implementation

The trace has not run.

Commit to an implementation, then inject the failure and inspect the actual state transitions.

03 · Implementation brief

Build it, break it, prove the outcome.

Implement
Deliver architecture, primitive catalog, version strategy, transport profile, handle and task lifecycle, auth model, threat model, trace schema, test matrix, and rollout plan.
Break it on purpose
Combine rolling-version inconsistency, replica loss, auth revocation, handle replay, malicious content, task cancellation race, and fallback pressure.
Completion evidence
The system remains coherent, denies unauthorized work, resumes truthfully, prevents unsafe fallback, and produces a versioned release dossier.

Your learning artifact

Write the argument you would defend

Stored only in this browser. No account required; course reset does not delete it.

Capstone decision rubric

Would this MCP system be safe to pilot?

Ready

The semantic contract, state ownership, security boundary, failure recovery, budgets, traces, and acceptance suite remain coherent under injected failures.

Incomplete

The happy path works, but a capability, terminal state, ownership rule, compatibility case, or recovery invariant remains implicit.

Unsafe

The design executes untrusted proposals, relies on stale or hidden authority, retries unknown effects blindly, leaks sensitive data, or hides failure behind a polished final answer.

04 · Check your understanding

What is the correct default for a new MCP implementation?

Primary sources · verified 2026-08-02

What the fixture does not prove

Each feature can pass alone while mixed-era fallback, load balancing, cancellation, token handling, or changed server metadata breaks the complete product.

Common mistake: A stateless server plus OAuth is enough for production. Production also needs semantic tool policy, tenant binding, safe handles, input and output validation, approvals, limits, drift detection, fault recovery, and evaluation.

Transfer exercise: Run a release-evidence workflow across random replicas while rolling one backend to a legacy version and injecting a malicious tool result.

Optional referenceDeep dive / reference chapterOpen the complete essay, diagrams, mathematics, exercises, glossary, and sources when you want more depth.

Next: Return to the evaluation course and turn the MCP conformance, security, and product suites into a monitored release argument.