Government tender search from an AI agent: the free tools, the one that costs, and coverage first

The three tools

`find_tenders(q, jurisdiction, tier, source, category, closing_before, first_seen_after, limit)` searches the open corpus and returns `{total, results[], coverage}`. Each result carries the title, buyer, jurisdiction, closing date, categories, a summary, a link to the notice, and which source it came from. It is free and needs no account, from any client — which makes it the natural first tool to try after connecting the server.

`tender_sources()` returns the registry: every source searched, what it is, how it is accessed, what its robots file permits, how many tenders are held from it, and its status on the most recent run — including the sources behind login walls and the ones that returned nothing. Also free.

`match_tenders(company_url | example, jurisdiction, closing_before)` reads what a company says it does on its own website, ranks open tenders against it, and returns for each match a `score`, the `why[]` — the company’s stated capabilities the tender needs — and the `gaps[]` — what the tender asks for that the site does not mention. This is the metered one: it reads a website and runs a scoring pass, so it needs a signed-in account’s key. Over REST the same three are `GET /v1/tenders`, `GET /v1/tenders/sources` and `POST /v1/tenders/match`.

Searching: filters that exist, and one that deliberately does not

Jurisdiction takes AU, NZ, or an Australian state or territory — AU-NSW, AU-VIC, AU-QLD, AU-WA, AU-SA, AU-TAS, AU-ACT, AU-NT. Tier is federal, national (New Zealand), state, council, university or health. Category and a free-text query narrow further; `closing_before` bounds the deadline; `first_seen_after` answers “what is new since my last look” — first-seen being the day the corpus first saw the tender, which is the honest clock for newness.

There is deliberately no location filter. Location is populated on a small minority of rows while jurisdiction is populated on all of them, so a location filter would look useful and silently hide most of the corpus from anyone who used it. An assistant asked for “tenders near Geelong” should search by jurisdiction and keyword, and say why.

A good agent turn: “Search open tenders in AU-NSW, tier council, query ‘road resurfacing’, closing before the end of next month” — one call, and the reply carries the notices and the coverage together.

Coverage first, especially when the result is empty

Every search reply includes `coverage`: how many sources were searched on the last nightly run and which returned nothing, with the reason recorded per source. An empty result therefore means “no match in what we searched” and never “no such tender exists”. For a supplier the difference is a missed bid; for an agent it is worse, because it would otherwise report “no matching tenders” with the same confidence in both cases.

The corpus comes from about twenty registered sources refreshed nightly: the federal procurement portal and the federal grants feed, every state and territory portal, council platforms and individual councils, universities and a health service, and New Zealand’s national tender service. Some are read from public feeds, some are scraped, some are read through a rendering service because they are single-page applications.

And some cannot be reached, which the registry says out loud. Portals that require a supplier login are named as login-walled. Sources whose robots file refuses automated access are named as refused rather than fetched anyway. Sources that returned nothing on the last run are listed as empty with the diagnosis. An assistant that reads `tender_sources` before answering “is my state covered?” answers correctly.

Matching a company: why and gaps

Give `match_tenders` a company website. It reads what the site says the company does — its stated capabilities — and ranks open tenders against them. Each match comes back with `why[]`, the capabilities the tender needs that the site states, and `gaps[]`, the things the tender asks for that the site does not mention. The gaps are the useful half: they are the questions to answer before writing the bid.

The score is a number from 0 to 100 that means relative fit within these results, and the reply says so in a field of its own. Nothing in a tender notice can tell you about incumbent relationships, pricing or the twenty other bidders, so a score that implied odds of winning would be selling certainty about a process it cannot see. An assistant should relay it as “this fits your stated capabilities well” and not as “you are likely to win”.

On the website and over REST, an anonymous match shows three results and counts the rest as withheld, so the shape can be tried before the meter runs. Over MCP the tool needs the account’s key, and returns the full ranked list.

From opportunity to bid/no-bid: the evidence an agent can gather

A bid/no-bid decision is a judgement, and no tool here makes it. What an agent can do is put the evidence in front of the people who do. The match reply already carries the two halves that matter most: `why[]`, the stated capabilities the tender needs, and `gaps[]`, what it asks for that the company has not said it has — a certification, a location, a licence. The gaps are the questions a bid team answers before it commits a week to writing.

The rest of the evidence comes from the same API — one piece of it over REST rather than MCP today. `POST /v1/companies/resolve` confirms a partner or subcontractor is the company it claims and reports its registration standing. `screen_adverse_media` screens a proposed joint-venture partner or the buyer’s named contact, with candidates and evidence rather than a verdict. `verify_document` checks the certificates and declarations the bid will attach, and `check_pack` — given the tender’s own document list as an ad-hoc checklist — names what is still missing from the submission.

Once a bidder is in front of a buyer, the same tools run the other way: the supplier onboarding and award gate takes a bidder’s pack, company and people through five questions before award. That is its own guide. This one ends where the opportunity is found, matched, and the gaps are named.

Alerts, and going to the original

Alerts live on the site rather than in the tools: sign in, run a search on the tenders page, and save it — new matching tenders arrive as a free weekly email, and a company profile with credits gets a daily ranked digest with the reasons and gaps. Each email carries a one-click stop link.

The corpus carries the notice and a link to it, plus a summary where the tender’s own page or documents were readable. It is not the tender documents, and a summary is not a specification. The aggregation exists to find the opportunity; the bid is written from the original, and an agent should hand its user the link rather than paraphrase the requirements.