MCP integration

One hosted MCP server: tender and grant search, buyer intelligence, and the document, text and evidence checks a bid needs. No install required.

Start with your account

  1. Sign in — every checker is free under a weekly limit.
  2. Open Your company → Account → Developer access and create a connector URL. Paste it into your agent's MCP connection settings. Clients that support headers can use an account API key instead.
  3. List the tools, then try tender_sources to check the connection — it is unmetered. A free tool succeeding does not confirm access to the rate-limited tools.
  4. Run the tool you need. If you hit the weekly limit, the reply says when it resets.

A sign-in error means the connection needs your account credentials. A rate-limit error names when it resets. Repeated retries will not resolve either.

The server

Everything lives on https://www.stipple.sh/mcp (Streamable HTTP), listed in the official MCP registry as sh.stipple/openwarrant:

ToolWhat it does
find_tendersFree, no key. Open tenders across Australia; New Zealand rows only when you ask for them (jurisdiction=NZ). Filter by kind (grant or tender), jurisdiction, tier, source, category, closing date or keyword. Every reply names the sources searched and the ones that returned nothing.
match_tendersRank open tenders against what a company does, read from its own website: why[] and gaps[] on each. The score is relative fit, never a chance of winning.
buyer_awardsFree. What a buyer has awarded, what is ending within 12 months with the incumbent, what they plan, the suppliers who win from them, and their open tenders. Anonymous callers see counts, values, dates and buyers; supplier and incumbent names after a free sign-in, and the reply says so.
find_signalsFree. Contract expiries, planned procurements and recurring tenders, each with its window (never invented), its evidence and a confidence word: published or derived.
tender_sourcesFree. Where the tender data comes from, and why a search came back empty: sources behind login walls, sources robots.txt refuses, sources that returned nothing.
detect_ai_textProbability a document's prose was AI-written, with the tells.
verify_referencesCitations resolve & match, internal math recomputed, claims judged; returns a public permalink.
verify_documentForensic authenticity inspection — risk band, findings, shareable warrant.
check_source_overlapDoes this text overlap text found on the public web (plagiarism-style)? Exact matches are confirmed quotes with source URLs; paraphrase flags are labelled model judgements. Overlap, never accusations.
extract_fieldsPull structured fields from any document — ad-hoc list, template, or auto-detected — plus explicit options (all off by default): whole tables as rows (deterministic and free on born-digital PDFs), per-value bounding boxes (native-PDF text layer only), classification, PII redaction, layout, links, figures, retrieval chunks with provenance, page-splitting, and confidence triage (flag_below). GET /v1/extract/capabilities is the machine-readable contract.
screen_adverse_mediaScreen a person/organisation for adverse media + sanctions exposure (PEP is not screened; the reply says so). Corroboration-gated sanctions matches (name-only = possible, never confirmed) plus entity-disambiguated news, same-name articles excluded. Screening coverage, not a determination.
verify_identityAustralian identity check over a set of documents (AFP 100-point or AUSTRAC safe-harbour) — established + the gap.
check_packDoes a set of documents satisfy a checklist? Classifies each (one cheap read) and matches the required slots — a named scheme or an ad-hoc requirements list; reports complete + what's missing.
check_document · get_warrant · submit_feedbackCache check, warrant retrieval, reviewer feedback.

The scoped servers, and the bid workspace

https://www.stipple.sh/mcp-tenders serves the five tender and intelligence tools alone (registry id sh.stipple/stipple-tenders); /mcp-aitext, /mcp-verify and /mcp-plagiarism serve one checker each. Same implementations, same limits; a new install wanting everything uses /mcp.

The bid workspace's own agent talks to a separate gateway, /mcp-bidline, with a key scoped to one bid run: it reads the bid's documents and requirements and records findings. It is not a public install; the workspace connects it for you when you prepare a response.

Claude (claude.ai and Claude Desktop)

Customize → Connectors → + → Add custom connector. Name it, paste your connector URL from the account page (Account → MCP → Create connector URL; it looks like https://www.stipple.sh/mcp/c/…) as the remote MCP server URL, leave the OAuth fields empty, and add. The plain https://www.stipple.sh/mcp also connects, for the free tools only. Free accounts get one custom connector; on Team and Enterprise plans an owner adds it under Organization settings → Connectors first. The tools then appear in a conversation's tools menu.

ChatGPT

Turn on Developer mode in ChatGPT settings (the menu it sits under has moved between releases — search settings for the phrase; availability can depend on workspace policy). Then create a connector: a name and description, Connection → Public endpoint, and your connector URL from the account page (https://www.stipple.sh/mcp/c/…) — or the plain https://www.stipple.sh/mcp for the free tools only. ChatGPT lists the tools it discovered.

Use the developer-mode connector, not the deep-research kind. Deep research requires two tools named search and fetch, which this server does not expose — its tools do specific jobs with typed arguments.

Claude Code

bash
claude mcp add --transport http stipple https://www.stipple.sh/mcp

To use the metered tools, attach your account's key on the same command: --header "Authorization: Bearer stp_…". The key is on your account page after a sign-in.

Accounts and keys — what works where

Seven tools need no account from any client: find_tenders, buyer_awards, find_signals, tender_sources, check_document, get_warrant and submit_feedback. Anonymous calls to the two intelligence tools get counts, values, dates and buyers; the supplier and incumbent names arrive with a key or a session. The metered tools — detection, references, source overlap, verification, extraction, screening, tender matching, identity and pack checks — need a signed-in account's key (free, no card), sent as an Authorization: Bearer header.

The claude.ai and ChatGPT connector dialogs take a URL and optional OAuth, not a header. For those two clients use your connector URL: Account → MCP → Create connector URL. It is an API key scoped to MCP, carried in the URL as /mcp/c/<token>. Every tool then works and spends this account's weekly limit. Treat the URL as a secret — it is shown once, it does nothing on the REST API, and you can revoke it from the account page. Claude Code, Cursor and config-file clients can keep passing the header instead.

The step-by-step version per client, with what to ask once connected, is in Connect an MCP server to ChatGPT and Claude.

Claude Desktop / Cursor / generic clients

Add to your client's MCP configuration:

json
{
  "mcpServers": {
    "stipple": { "type": "http", "url": "https://www.stipple.sh/mcp" }
  }
}

For tenders and intelligence alone, https://www.stipple.sh/mcp-tenders; the other scoped servers are listed above. New installs wanting everything should just use /mcp.

Calling the tools

The document tools accept a public url — the cheapest call: one short string, no download or encoding — or inline base64 (bytes_b64); the identity and pack checks take a documents list; the text tools take raw text or a url.

text
# In any connected client, just ask:
"What council road tenders are open in Queensland?"
"Which of these could we bid for? Our site is example.com.au"
"Who holds Transport for NSW's contracts, and what is ending soon?"
"What is planned for ICT in Queensland next quarter?"
"Fact-check this report: https://example.com/report.pdf"
"Is this payslip genuine? <attached file>"

verify_references returns a shareable permalink with every result, so an agent can hand its user a public fact-check record. Each server also publishes MCP prompts (ready-made starters like "Fact-check a report URL") that appear natively in clients that support them.

Limits & behavior

  • Same per-IP rate limits as the REST API (the tools share one quota).
  • Outbound URL fetches are SSRF-guarded: public http(s) hosts only, every redirect re-validated, size-capped.
  • Transport is stateless Streamable HTTP with JSON responses — no session affinity required.
For local or private files too large to inline, a local stdio server mirrors /mcp: it reads files on your machine and calls this same API for every tool. Setup instructions are in agents.md.

Not using MCP?

Agents can also discover the service through agents.md and llms.txt, or call the REST API directly.