Using the GST MCP Server

01 Your first query

A connected server proves nothing until it answers something you would otherwise have paid an analyst a morning for. This one is a good first test: the corpus spans 123 regulatory records, the answer is checkable against public law, and it needs no history with GST. Start a fresh conversation and paste it.

• An MCP tool call runs and returns

  Claude names the MCP tool before reaching for it, runs two GST MCP calls (List Regulatory Map Facet Values, then Search Regulatory Map) and answers from what they return. The capture below is that exact exchange.

Then try one of these

02 Run a prompt with arguments

Everything above is free-text prompting: you describe the job and hope the agent routes it well. The twelve GST prompts are the other path. Each one is a workflow a senior consultant already wrote (which MCP tools to call, in what order, what to do with the results), and Claude Desktop collects its inputs from you in a form before the model sees anything. Same data, no improvisation.

1. 01

   Open the prompt picker

   In the composer, click + → Add from GST MCP, the same menu you checked when you verified the connection. The twelve entries prefixed gst_ are prompts, not MCP tools: you invoke them, the agent doesn't.

2. 02

   Fill the argument form

   Picking a prompt opens a field per argument, in the order the server declares them; the ones you always supply come first. Required fields are marked; leave an optional field blank and the workflow adapts rather than guessing. Values are validated server-side before the prompt renders, so a malformed number or an unrecognised enum comes back as an error, not a confident wrong answer.

3. 03

   Submit and read the tool calls

   The rendered workflow lands in the conversation and the agent works through it. You should see the MCP calls the prompt names: several in sequence, not one. That sequence is the value: it is what you would have had to orchestrate by hand.

Worked example: first look at an unfamiliar target

gst_target_quick_look takes five required arguments and drives four analyses (cost-governance maturity, unit-economics benchmark, tech-debt range, regulatory exposure) collapsed into one page. Fill it like this:

Read the assumptions section firstThis prompt fills the twenty-question cost-governance assessment from what you gave it, and marks every answer it could not derive as "Not sure" rather than guessing. Those show up in the brief under Assumptions; if most of them are unknowns, treat the cost-governance read as a placeholder and run the full assessment in the Hub.

The five bare arguments are the floor, not the intended run. The prompt is evidence-conditional: with an IRL extract record already in the conversation it resolves its inputs from the record instead of stage norms, cites the workbook reference behind each number, and states which branch produced every figure. Run gst_irl_extract first (zero tool calls) and the same form fields become labels on evidence rather than the evidence itself. How the record works, and which other prompts read it, is covered in Advanced operations.

Two more worth knowing

03 Attach a resource

The server exposes three kinds of content, and they are not interchangeable. MCP tools are called by the agent. Prompts are invoked by you. Resources are attached by you: inert documents that do nothing until you pin one to a conversation, at which point the model is reading GST's canonical text instead of recalling something adjacent to it.

Pin one the same way you attached the prompt: + → Add from GST MCP, then choose from the resource list rather than the gst_ entries. It stays attached for the rest of the thread.

A prompt cannot fetch a resource for youResources are user-pinned by design: the model cannot read one from inside a prompt expansion. This trips people up because it looks like a gap: you invoke a workflow that clearly needs a canonical document, and nothing is fetched. It isn't a gap. Where a workflow depends on GST text, the server embeds that text in the prompt body before you ever see it, so the document arrives whether or not you pinned it. Pinning is for your follow-up questions afterwards, the conversation where you interrogate the framework rather than run the workflow.

04 Troubleshooting

The five failures a first connection runs into, what each one means, and the one fix that lives in your Claude settings rather than on the server.

For any question about GST's portfolio, engagements, radar, regulatory
analysis, TechPar, infrastructure cost governance, diligence, technical
debt, or information request lists:

1. Your FIRST sentence must name the GST MCP tool you intend to call.
   Do not open with "I'll search past conversations" or "Based on what
   I know"; the MCP tools are the authoritative source, not memory.
2. Call the GST MCP tool first, even when the prompt does not mention GST.
3. For "what does this framework cover?" questions, call the MCP tool with
   empty arguments rather than describing it from memory.
4. When a task matches a published GST prompt, invoke the prompt rather
   than orchestrating its tool calls yourself.

05 Next steps

Three routes on from everyday use: the IRL ingestion workflow, the wire detail, and a wider grant.