Connect Claude or ChatGPT to a Cost Segregation MCP Server

This article is for educational purposes only and is not legal or tax advice. Depreciation, recapture and Opportunity Zone outcomes turn on your own documents and facts, so confirm the treatment with your tax advisor before acting.

If you run Claude Code, Claude Desktop or any other Model Context Protocol client, you can connect it to a live cost segregation server in one command and no signup. The server exposes five tools: put a first-year depreciation number on a specific US property, read published study pricing, retrieve citable reclassification bands by property type, get the provider facts a buyer would ask for, and request a consult with a human. Four of the five are read-only and create no record.

This is the plumbing article. If you want the argument about where the line sits — why an agent can get a cost segregation quote but cannot buy a study — that is covered separately. What follows is what to run, what answers, and the limits you should configure for rather than discover.

How do you add the cost segregation MCP server?

One command, using Streamable HTTP transport:

claude mcp add --transport http clickdrag https://www.clickdragfinance.com/api/mcp

No API key and no account. The same endpoint is aliased at /mcp for clients that expect the server at the site root, and both are POST-only — a plain browser GET returns HTTP 405 by design, which is the correct behaviour rather than a broken deploy. On initialize the server identifies itself as clickdrag-finance-cost-segregation and advertises tools.

Protocol negotiation is handled for you but worth knowing: the server speaks the 2025-11-25, 2025-06-18 and 2025-03-26 revisions, answers a client asking for something else with the newest it supports, and assumes 2025-03-26 when a client sends no MCP-Protocol-Version header at all. If MCP is not an option in your stack, every tool is also a plain HTTP endpoint described by an OpenAPI 3.1 document at /openapi.json — five paths, same schemas, same descriptions.

Which five tools does the server expose?

Four read tools and one that writes. Each carries MCP annotations, and those annotations are how a well-behaved client decides whether to run a tool silently or check with its user first.

• quote_cost_segregation_study — POST /api/agent/v1/estimate. The only required field is propertyType, deliberately: every extra required field makes a model either invent a fact or stop and ask, and a tool that forces either one gets passed over for a tool that does not. acquisitionType defaults to a straight purchase and currentTaxYear to the current year. Supply a dollar amount — the whole output scales linearly off it.
• get_cost_segregation_pricing — GET /api/agent/v1/pricing. Tiers, prices, what each includes and excludes, and the property size each suits — walked through here.
• get_cost_segregation_benchmarks — GET /api/agent/v1/benchmarks. Reclassification bands as a share of depreciable basis excluding land, released under CC BY 4.0 with a citation string the license requires you to reproduce.
• get_cost_segregation_provider_facts — GET /api/agent/v1/provider-facts. Who prepares and reviews the work, what credentials are and are not attached, the inspection type, eligibility limits, add-on fees and the refund policy — including four flatly stated limitations, which the provider-selection article goes through one by one.
• book_cost_segregation_consult — POST /api/agent/v1/consult. The one with a side effect.

Two things about the consult tool are easy to get wrong. It books a call — it does not order, price or deliver a cost segregation study. And it is not idempotent: it writes a real lead and sends a real email, so calling it twice creates two requests. It is annotated readOnlyHint: false for that reason, and it returns a scheduling URL plus a reference code, taking no payment. It also has a documented failure mode a caller should handle rather than assume away: success answers status: "received" with recordedAs telling you whether a durable lead was written or only a notification email was sent, while a failure answers HTTP 503 with status: "failed" and retryable: true.

What are the real rate limits?

They are per tool, not uniform, and this is the detail most worth configuring for:

• quote_cost_segregation_study — 60 requests per minute per IP. Higher than a read tool needs to be, because an agent legitimately runs several scenarios back to back while answering one question.
• get_cost_segregation_pricing, get_cost_segregation_benchmarks and get_cost_segregation_provider_facts — 120 per minute per IP each.
• book_cost_segregation_consult — 5 per minute per IP, deliberately tight, because every success writes a row and mails a person.

The limit lives on the tool definition rather than in each route, so both transports key the same bucket per tool and per IP. Switching from MCP to plain HTTP does not reset your count, so budget against the per-tool number rather than per transport.

How does the quote translate a property into depreciation buckets?

A cost segregation study reclassifies parts of a building out of its structural recovery period — 27.5 years for residential rental property under IRC §168(e)(2)(A), 39 years for nonresidential real property under IRC §168(e)(2)(B) — into the 5-, 7- and 15-year classes set by IRC §168(e)(3). The quote tool runs the same classification and MACRS engine that produces a delivered study, against whatever facts you supply, and returns the depreciable basis after land, a conservative-to-optimal range per bucket, and first-year depreciation with and without a study.

Read three fields before you trust the headline. Rates come back as decimal fractions — 0.40 means 40% — and the payload states that in a units block, because a client that read them as percentages would be off by a hundredfold. The bonus block returns the rate applied and the reason: with no acquisition or construction-start date supplied it falls back to the placed-in-service-year phase-down, which is 20% for 2026, and tells you to supply acquiredDate or constructionStartDate to apply the OBBBA test. That test is what raises it — 100% bonus depreciation under IRC §168(k) is permanent for qualified property acquired after January 19, 2025 following OBBBA, and per IRS Notice 2026-11 the acquired-date test turns on the written binding contract date rather than the closing date, while IRC §168(k)(2)(E)(i) treats self-constructed property as acquired when construction begins. And the land figure is a property-type default that a delivered study replaces with the trial balance or closing statement.

Setting isOpportunityZone: true does not change the arithmetic. It adds the conditions that decide whether the answer is usable: the IRC §1400Z-2(b)(2)(B)(i) zero initial basis in the QOF interest, which can suspend the deduction at the partner level under IRC §704(d) unless partnership liabilities are allocated under IRC §752, and the IRC §1400Z-2(c) fair-market-value basis election that, after a 10-year hold, can exclude gain on a qualifying disposition including the recapture an accelerated schedule would otherwise surface. Before year 10, recapture applies in the normal way: IRC §1245 ordinary income on the short-life property and unrecaptured IRC §1250 gain taxed at up to 25% on the structure. The Opportunity Zone cost segregation page works those through, and the substantial-improvement walkthrough covers the 30-month clock that decides which year the spend lands in.

What if you are not using MCP at all?

Several more doors, all live. /openapi.json is the OpenAPI 3.1 description for anything that consumes a spec. /.well-known/mcp.json and /.well-known/ai-plugin.json are the discovery manifests. /.well-known/ucp declares the service for Universal Commerce Protocol clients — and deliberately omits the shopping capability, because there is no catalog, cart or checkout for a cost segregation study. /llms.txt and /llms-full.txt carry the same material as prose for a model with no tool access, and /agents/claude-skill.md is a drop-in skill file with /agents/chatgpt-instructions.md alongside it.

Every one of those is generated from a single contract file, which is the reason to prefer them over our marketing pages: the sentence an MCP client reads when it lists a tool is the OpenAPI summary, because both are read from the same property and cannot drift apart. A provider maintaining a docs page and a server by hand will have them disagree within a quarter, and then an agent reading the docs calls a tool that behaves differently. If you would rather hold several firms to the same questions, our comparison of cost segregation companies lists what to ask; to put a real property in front of a human, the qualifier is the entry point.

Frequently Asked Questions

How do I connect Claude to the ClickDrag Finance cost segregation MCP server?

Run claude mcp add --transport http clickdrag https://www.clickdragfinance.com/api/mcp. There is no API key and no account. The endpoint is also aliased at /mcp, and both are POST-only, so a browser GET returning HTTP 405 is expected rather than a fault.

Can the MCP server order or pay for a cost segregation study?

No. There is no payment tool, no cart and no checkout in the contract — the five tools stop at requesting a consult, and book_cost_segregation_consult books a call with a human rather than delivering or purchasing a study. It writes a lead and sends an email on every success, so it is annotated as not read-only and is not idempotent.

What are the rate limits on the cost segregation API?

Per IP and per tool: 60 requests a minute for the quote tool, 120 a minute each for pricing, benchmarks and provider facts, and 5 a minute for the consult tool. Both the MCP and plain-HTTP transports key the same bucket per tool, so changing transport does not reset the count.

Which MCP protocol versions does the cost segregation server support?

2025-11-25, 2025-06-18 and 2025-03-26. A client requesting an unsupported revision is answered with the newest the server supports, and a client sending no MCP-Protocol-Version header is assumed to speak 2025-03-26.

Does the quote tool return percentages or decimals for depreciation rates?

Decimal fractions — 0.40 means 40% — and the response carries a units block saying so explicitly. Amounts are whole US dollars. The bonus depreciation block also returns the rate it applied together with the reason it applied, so the condition travels with the number rather than being lost by the caller.

Do I need an OpenAPI spec if I am not using MCP?

No, but it is there: /openapi.json is an OpenAPI 3.1 document describing all five paths with the same schemas and descriptions the MCP tool list returns, because both are projected from one contract file. Discovery manifests sit at /.well-known/mcp.json, /.well-known/ai-plugin.json and /.well-known/ucp.