Connecting an AI agent to public tender data
A chat window over a dashboard is not the interesting part. Giving an agent the underlying data, with provenance, is.
Almost every procurement platform has added an AI feature in the last two years. Most of them are a chat box over a dashboard you already had. The genuinely useful version is the other way round: your assistant — the one you already work in — gets the ability to query the data itself.
What MCP is, briefly
The Model Context Protocol is an open standard for connecting an AI client to a data source or tool. The client discovers what tools exist, calls them with structured arguments, and gets structured results back. It is a plug format, not a product, and it removes the need for a bespoke integration per assistant.
It is being adopted in procurement specifically: GovSpend shipped an MCP server for US public-sector procurement data in 2026, and there are MCP servers for Indian and Moldovan procurement portals. This is not a novelty any more, which is good — a standard is only useful once several people have implemented it.
What it changes in practice
The difference is not that you can ask questions in English. It is that the question and the answer stay in the place where you are already working, and the agent can chain steps you would otherwise do by hand.
- "Find open tenders matching our profile closing in the next three weeks, qualify each one against our capability, and tell me which three are worth a bid meeting."
- "Who has won facilities management work from Liverpool City Region since 2023, and what were the contract terms?"
- "Set up a monitor for structural engineering work in the North West above £250,000, and tell me what is new each morning."
- "This tender mentions UKAS 17020 — do we hold that, and which of our past contracts used it?"
Each of those is several searches, a spreadsheet and a judgement call when done manually. The value is not the chat interface; it is that the agent can hold the whole sequence.
The questions to ask before you trust it
An agent will state whatever the data source returns, fluently and without hedging. That makes the quality of the underlying data more important than it was when a human was reading a dashboard and applying scepticism. Ask any vendor these five.
| Question | Why it matters | Bad answer |
|---|---|---|
| How do you treat framework ceilings? | A framework's published value is capacity, not revenue. Summed as revenue it can overstate a supplier tenfold. | "We report the published value." |
| What do you do with a multi-supplier award that has no split? | Dividing evenly invents a number; giving each supplier the full value inflates the market. | "We divide it between them." |
| Do you sum sterling and euro? | Any single UK-and-Ireland total has either converted at an unstated rate or simply added them. | A single combined figure with no rate or date. |
| Do you deduplicate by notice or by procurement? | One procurement generates several notices across several services. | "We hold 400,000 notices." (a volume claim, not a coverage one) |
| Does every record carry a link to the source notice? | Without it, nothing the agent says can be checked. | "Our data is verified." |
How to connect one
- Decide where the agent lives — a desktop assistant, a browser assistant, or something you run yourself against an API.
- For desktop and browser assistants, add the MCP endpoint in the connector settings and authorise it. Claude and ChatGPT both support this today; Perplexity and Grok take an API key.
- For your own agent, an MCP client library or a plain REST call both work. If you are building against the OpenAI or Anthropic APIs, the server can be passed as a remote tool.
- Start with read-only questions until you trust the answers. Ask it something you already know the answer to.
- Check one number by hand, against the source notice, before quoting any of it to anyone else.
TenderGraph exposes fifteen tools over MCP and the same data over REST, with the source notice on every record and the attribution rules published rather than implied. The connection guides cover Claude, ChatGPT, Perplexity, Grok and anything that speaks the protocol; the integrations page explains what each tool does.
Common questions
- What is the Model Context Protocol?
- An open standard for connecting AI clients to external tools and data. The client discovers available tools, calls them with structured arguments and receives structured results — which means one server works with any compatible assistant rather than needing a bespoke integration each time.
- Is an MCP connection secure?
- It carries the same considerations as any API access: it authenticates as you, it should be scoped, and the key or token should be revocable. Prefer OAuth or a header-based key over one embedded in a URL, because URLs reach access logs in a way headers do not.
- Will an AI agent hallucinate tender data?
- A model asked to recall procurement facts from memory certainly will. A model calling a tool returns what the tool returned — which moves the risk from invention to data quality, and makes the vendor's methodology the thing to interrogate.
- Do I need to be technical to use this?
- Not for the assistant route. Adding a connector in Claude or ChatGPT is pasting an endpoint and approving it in the browser. Building your own agent against the API is a developer job.
Read next
- How to find out who won a public contract — and what they were really paidThe winner is easy. The amount is where award data goes wrong, and a framework ceiling read as revenue can overstate a company by a factor of ten.
- A bid/no-bid decision you can defend on Monday morningWin rate is not the metric that matters. Cost per win is — and the fastest way to improve it is to stop entering competitions you were never going to win.
- Your tender alerts are either noise or a blind spotThere is no volume setting that fixes a badly built feed. The failure is usually in the targeting, and it is diagnosable in an afternoon.