Practice · MCP integration

Inspect MCP tools before connecting a workflow

Run a pinned MCP client, inspect tool discovery and retrieve a named workflow with a dated result you can reproduce.

Jason Colapietro · Suede AI SEO ·

Start with the workflow request

An assistant connected to a workflow library needs an inspectable route from a request to the instructions it receives. The Suede Creator Skills demonstration starts with the text “technical SEO audit,” discovers the available tools, then retrieves the named suede-seo-audit workflow. You can inspect that sequence before adapting it to an internal knowledge or operations library.

The saved run on 21 October 2026 records five protocol requests and nine discoverable tools. It retains the server version, negotiated protocol, search query, returned workflow and instruction hash. Those fields make a later comparison useful: a reviewer can identify a changed catalog or changed instructions without guessing from a screenshot.

Pin the source and run the client

Use a checkout of suede-seo and Node 22 or newer. Clone the public skill pack beside that checkout, then select the exact commit used by the saved demonstration. The runner verifies the server, catalog and selected workflow against that commit before starting the process.

The pin is part of the experiment. Running a new catalog against an old expected response would mix two versions and make a changed output harder to diagnose.

git clone https://github.com/JasonColapietro/suede-creator-skills.git ../suede-creator-skills-demo
git -C ../suede-creator-skills-demo checkout --detach 46445415a767df5827f5e124f41749ef369cca83
node examples/training-proof/run.mjs --skills-repo ../suede-creator-skills-demo --out /tmp/suede-training-proof

Inspect the discovery contract

Read run.mjs in order: initialize, send the initialized notification, request tools/list, search, and retrieve a skill. The MCP tools specification defines the discovery and invocation message shapes. The example keeps those requests visible instead of hiding them behind a model session.

Inspect each tool’s name, input schema and annotations in the pinned server source. The saved result reports readOnlyHint on all discovered tools. An annotation describes the server’s declared behavior; inspect implementation and access controls before granting a production client access to business systems.

Review the returned instructions

Open the generated results.json and compare mcp.search_query, mcp.matches and mcp.returned_skill. The search returns three named matches in the recorded run; the client explicitly selects the named workflow. It then checks that the returned body has content and saves its SHA-256.

That separation is useful when retrieval becomes ambiguous. A search result is a candidate, while the selected workflow is the exact object the client requested. Keep both in an integration report so a reviewer can reproduce the selection and inspect the instructions that followed.

Turn the example into an acceptance checklist

For your own server, write down the intended reader request, required tool, allowed arguments and expected response before connecting a model. Add a negative case for an unknown identifier and retain the original response for review. The companion error-contract article develops that step.

Jason Colapietro’s Introduction to Model Context Protocol course completion badge and Integrate MCP Tools with Azure AI Agents training module achievement connect to this implementation exercise. Their issued records appear below. The working proof here is the pinned client/server exchange, which used zero paid model calls.

  • Record the server commit and protocol version.
  • Confirm discovery returns the intended tool and schema.
  • Retain the selected workflow and instruction hash.
  • Inspect a failed lookup as well as a successful lookup.

Issued training connected to this exercise

Inspect the full training inventory · Open the qualifications sheet

Source and recorded result

The demonstration report is dated 2026-10-21. Repeat the command against the revision you intend to inspect and retain the resulting artifact.

Put the workflow to work

Define a workflow pilot. Bring the workflow, source page or answer your team needs to check so the first deliverable has clear acceptance criteria.

Test the failed lookup in an MCP workflow · All three working demonstrations