Key takeaways
- MCP is the connection standard; it does not itself decide the legal answer.
- Tools let the assistant search, inspect and read external source material during a conversation.
- A useful GST workflow separates document discovery, page reading and answer drafting.
- OAuth and tool approvals help keep a remote connector permissioned and reviewable.
MCP in one sentence
The Model Context Protocol, usually shortened to MCP, is an open standard that lets an AI application connect to external tools and information through a consistent interface. The official MCP documentation describes a client–host–server architecture: Claude or ChatGPT acts as the host experience, a client manages the connection, and a remote server exposes focused capabilities.
For GST work, that means the assistant does not have to depend only on the text already present in the conversation or on what the underlying model happens to remember. A connected server can offer a controlled way to find a circular, identify a relevant notification, list the available documents or read the pages needed for the answer.
MCP is therefore the connection layer, not a separate legal brain. The quality of the result still depends on the document library, the retrieval method, the tool descriptions, the assistant’s reasoning and the user’s review.
The three MCP building blocks
The MCP specification identifies three main server primitives. They are related, but they do different jobs.
| Primitive | Who normally controls it | Practical GST example |
|---|---|---|
| Tools | The AI model, subject to host controls | Search the library, list documents, read pages or retrieve an image |
| Resources | The host application | Attach a maintained reference, file or structured item as context |
| Prompts | The user | Choose a reusable workflow such as researching an issue or drafting a reply |
TaxByKK primarily uses tools to give the assistant repeatable research operations. A prompt template can help the user begin with a well-structured instruction, but it should not be confused with retrieval: the useful evidence still comes from the source documents reached through the tools.
What happens after you ask a GST question
Consider the question: “What was the time limit for taking ITC on an FY 2019–20 invoice, and did section 16(5) change it?” A sensible connected workflow can be broken into four stages.
1. The assistant identifies the research task
Claude or ChatGPT reads the user’s question and sees the tools exposed by the connected server. Tool metadata tells the assistant what each operation does and which inputs it accepts. The host decides how tool approvals are displayed and whether the user must confirm a call.
2. The server discovers the relevant documents
The assistant asks the server to locate material matching the issue. In this example, likely sources include the current text of section 16, the Finance (No. 2) Act, 2024 and CBIC Circular No. 237/31/2024-GST. Good discovery should return document identity and relevance, not silently manufacture a legal proposition.
3. The assistant reads the relevant pages
Document discovery is not the same as evidence. The next step is to read the relevant statutory or circular pages, including qualifications that may change the answer. Page-level retrieval also makes it easier to show where a proposition came from.
4. The assistant drafts from the retrieved evidence
Only after the source is available should the assistant explain the rule, distinguish the financial years and state any limitation. The final answer should make the legal period visible and cite the source rather than presenting an uncited conclusion as certain.
How a remote connector is secured
A remote MCP server is reached over the internet. In supported connector flows, the server can use OAuth so the user signs in and grants access without sharing the account password with the assistant. Anthropic’s connector guidance explains that users typically complete an OAuth flow and can later revoke access by disconnecting the connector.
Security also depends on the nature of the tools. A research connector should expose narrowly described operations, validate their inputs and avoid unnecessary write access. Users should review tool requests, especially when a connector can change external data. TaxByKK’s research workflow is designed around retrieving and reading tax material; it does not need permission to modify a user’s unrelated services.
What TaxByKK adds to Claude and ChatGPT
TaxByKK connects the assistant to a curated tax document workflow. In practical terms, it helps the assistant:
- identify which source document is relevant;
- trace connected provisions, amendments or circulars;
- read the necessary pages rather than guessing from a title;
- carry the source into the reasoning and drafting process; and
- give the user evidence that can be independently checked.
The assistant still controls the conversation and prepares the prose. TaxByKK supplies the research operations and connected material. This separation is useful: the same source workflow can support an explanation, a notice reply, a checklist or an article without pretending that one fixed prompt is a substitute for legal analysis.
MCP versus a generic AI answer
| Generic answer without connected research | Answer using a source workflow |
|---|---|
| May rely on model memory | Can retrieve the current connected source |
| Often gives a provision without the underlying page | Can read and cite the relevant page |
| May collapse different financial years into one rule | Can compare period-wise amendments |
| Harder to audit | Gives the user a document trail to verify |
Connection does not guarantee correctness. A wrong document can be selected, a qualification can be missed, or the source library can be out of date. The advantage is that the research becomes observable and correctable.
A reliable way to ask questions
The strongest instructions provide the issue, the period and the required output. For example:
Research whether ITC on invoices from FY 2019–20 is protected by section 16(5). Use the statutory text and CBIC clarification, distinguish eligibility from the rectification procedure, and give page-level citations.
This works better than “Explain ITC” because it tells the assistant what must be separated. When facts matter, add the invoice year, return filing date, procedural stage and whether an order or appeal already exists.
The limit that matters most
MCP makes external research available; it does not remove the need for professional judgment. GST outcomes can depend on the applicable date, state enactment, documents on record, limitation, procedural history and later judicial developments. Use the cited material as the starting point for verification, not as a reason to stop checking.
Primary sources
Frequently asked questions
Does an MCP server train Claude or ChatGPT on my documents?
No. MCP is a protocol for connecting an AI application to external tools and context. A tool can retrieve source material for the current task without turning that material into the model's permanent training data. Platform data-handling terms still need to be reviewed separately.
Can the MCP server read my whole conversation?
The MCP architecture is designed so a server receives the inputs required for a particular tool call, not unrestricted visibility into every conversation or every other connected server. The host application controls what is sent.
Why use MCP instead of pasting a PDF into every chat?
A connector can search a maintained library, retrieve only the relevant material and expose repeatable research operations. Pasting a PDF can still be useful for a one-off matter, but it does not provide the same reusable discovery workflow.
Does TaxByKK replace review by a tax professional?
No. TaxByKK assists research and drafting. The user remains responsible for checking the cited source, the relevant period, the taxpayer's facts and any later amendment or judgment.
Is a retrieved document automatically the current law?
No. Retrieval proves what the document says, not that it remains operative. Check its date, amendments, effective date, later circulars and any binding judicial development before relying on it.
What should a GST MCP answer cite?
Prefer the statutory provision, notification, circular, official advisory or judgment that supports the proposition. A search result or AI summary is a discovery aid, not the final authority.
Explore the GST authority library
Research the underlying documents inside your AI assistant
Connect TaxByKK to Claude or ChatGPT to trace provisions, read source pages and build a cited working draft.
View TaxByKK plans