Every public page is now readable without JavaScript — plus llms.txt
All public pages (English and Hungarian) are now also served as full, pre-rendered static HTML. Crawlers, AI assistants and anyone fetching a page without executing JavaScript get the complete content — including per-page titles, descriptions and language links — instead of an empty application shell. We also published /llms.txt, a machine-readable summary of what GrantCortex is and where to find the docs, following the emerging convention for AI-assistant discovery. Nothing changes visually for browser visitors.
WebSEO
Roadmap page refreshed; new roadmap overview on the homepage
The public Roadmap page got a refresh: the "live now" lane no longer carries quarter labels (it reflects what is actually shipped today), and upcoming milestones were updated — lump-sum budgeting and deadline alerts, then writing-support tooling and an early look at the next EU framework programme. The homepage now includes a compact roadmap overview section after the feature grid, linking through to the full page. The interactive tool-call playground moved from the homepage to the bottom of the Docs page, next to the tool reference it demonstrates.
WebUX
Privacy-first visitor analytics — cookieless and self-hosted
We now measure site traffic with a self-hosted Umami instance running on our own EU server. No cookies, no cross-site tracking, no data shared with third parties (we deliberately chose this over Google Analytics), and the script does not even load if your browser sends Do Not Track. We only see aggregate, anonymous statistics — page views, country, referrer, device type. The Cookies and Privacy pages document the details, and no consent banner is needed because nothing is stored on your device.
WebPrivacy
Public beta — the site is now open
The access-code gate is gone: grantcortex.eu is now publicly browsable. You can read the docs, check live source coverage, create an account and generate your tester API key without an invite code — free during the beta. The MCP endpoint itself (mcp.grantcortex.eu) has been publicly reachable for authenticated testers all along; what opens today is the site around it.
WebLaunch
ChatGPT connection guide updated for the new Plugins flow
ChatGPT recently reorganised how custom MCP connectors are added: Developer mode now lives under Settings → Security and login, and connectors are created as Plugins (Settings → Plugins → Browse plugins → ＋ → New Plugin). The animated walkthrough on the Docs quickstart and in the account dashboard now mirrors the new UI click by click — including the New Plugin dialog and picking GrantCortex from the composer’s ＋ menu. If you connected under the old Apps flow, your connection keeps working; the guide simply matches what new testers see today.
DocsOnboarding
Self-service tester keys — register and generate your key in minutes
Tester access no longer waits on a manual onboarding step: create an account, confirm your email, and generate your tester API key right in your dashboard — free during the beta. The waitlist stays as an update list for launch news. The Docs authentication section and every call-to-action on the site now describe this flow.
ProductOnboarding
Truth-pass across the site: live-verified examples, Docs matched to the live API
Every example on the site now comes from live data: the partner-search sample on the Coverage page shows real PIC numbers and project counts from a dated corpus snapshot, and the homepage demo conversation uses real, currently open call identifiers. Source counts read the same everywhere — 3 searchable live sources plus the CORDIS partner archive — and the Coverage page now distinguishes sources being prepared for licensing from those currently excluded. The Docs tool reference was regenerated against the live API: search is hybrid (keyword + semantic) by default with a lexical_only opt-out, aggregated queries default to EU-level sources, and the quickstart examples resolve to real cached records. The landing demo answers in clean plain text and explains transparently which national catalogues are not served yet and why.
WebDocsData quality
Legal pages are final, plus a new Data sources & attribution page
The legal surface has left draft mode: the Terms, Privacy, Imprint, Cookies, Acceptable Use and Takedown pages no longer carry a draft banner or internal review notes, and the Imprint now states the operator’s real legal identity. A new page, "Data sources and attribution" (/legal/sources, also reachable at /adatforrasok), explains exactly which sources GrantCortex serves today, under which licences (CC BY 4.0, CC0, Dutch open-data rules), how attribution is carried, and how to reach us about any rights concern. All contact routes are now a single address: info@grantcortex.eu.
WebLegal
Search defaults to EU-level sources — national catalogues on request
Aggregated grant searches now cover the EU-level sources (Funding & Tenders Portal and CORDIS) by default, so a query no longer surfaces Swedish or Dutch national calls unasked. The national catalogues (Vinnova SE, RVO NL) remain fully searchable — your agent simply names the source when you want them. Recommendations follow the same scoping, and the in-chat result cards now label EU-level rows with an "EU" chip.
MCPSearch
Animated connection walkthroughs in your dashboard, ChatGPT first
The "Connect your agent" card on your account dashboard now shows the same step-by-step animated walkthrough as the Docs Quickstart — pick ChatGPT, Claude Desktop or Perplexity and watch exactly where to paste the endpoint and your key, with a compact copy block underneath and a link through to the full setup. ChatGPT now leads the walkthrough on both the dashboard and the Docs Quickstart.
WebUX
Client list refreshed and a leaner Docs tool reference
The "works with any MCP client" line and the scrolling client strip now highlight the AI assistants people actually use for grant work — Claude, ChatGPT, Gemini, Perplexity, Microsoft Copilot, Raycast, n8n, Goose and LibreChat — instead of developer-focused code editors. On the Docs page the tool reference now lists only the tools available to you: the internal admin-only refresh trigger (which runs automatically every day) is no longer shown, so the tool list and the sidebar match what a tester key can actually call.
WebDocs
Site polish: steadier hero, cleaner coverage, docs and stats layout
A round of interface refinements across the public pages. On the home page the animated chat demo now reserves a fixed height, so the layout no longer jumps as the conversation streams in. On the Coverage page the country tile-map is centered and the status legend (Active / Forthcoming / Expired / No structured deadline) moved directly beneath the sources table as plain, readable rows instead of a boxed grid. On the Docs page the left-hand tool list no longer overflows — long tool names wrap into the column — and the coming-soon tools are grouped into their own "Coming soon" section below the live ones. On the Stats page the four F&T Portal programmes (Horizon Europe, Digital Europe, EIC, Horizon Pillar 1) are now grouped under a single "F&T Portal / SEDIA" heading, so the source breakdown matches the "3 integrated sources" figure.
WebUX
Grant search now resolves exact topic/call IDs directly
Searching with an exact topic or call identifier (a SEDIA topic code, a Vinnova or RVO reference) now returns that record directly instead of an empty result — identifiers were never part of the keyword index, so an exact-ID search used to come back empty. Recommendation requests are also more reliable: the business-profile input now states explicitly that the requested result count is a separate argument, not part of the profile, which occasionally got confused in longer conversations.
MCPSearch
All tools renamed with a grantcortex_ prefix + richer schemas and leaner responses
Breaking change: every tool now carries a grantcortex_ prefix (e.g. grant_search → grantcortex_search_grants), so GrantCortex tools stay unambiguous next to other MCP servers in your client. Tool descriptions were rewritten with explicit "when to call" guidance, every parameter now carries its own documentation, and several inputs became stricter and clearer: enterprise inputs of the SME calculator are fully typed, the submission-guide programme key is a real enum, dates are native date fields, and partner-search pagination uses separate controls for organisations and projects. Responses got leaner and more machine-friendly too: every tool returns the same provenance envelope, list results report total_count / has_more / next_offset so truncation is never silent, grant search defaults to compact rows (pass verbosity="detailed" for the full text), and the eligibility check returns its verdict without the raw criteria dump unless you ask (include_criteria=true). If you granted per-tool permissions in your client, you may need to re-approve the renamed tools; everything else about your connection is unchanged.
The in-chat result cards were redesigned to blend into your chat client instead of carrying our own visual branding: they now follow the host’s typography and theme, with a frameless, quieter layout. The grant-search card also became interactive — you can solo a data source with one tap and toggle “open calls only”, without typing a new query. Clients without MCP Apps support keep the exact same text answers.
MCPUX
Tool results now render as in-chat cards (MCP Apps)
On MCP hosts that support the new MCP Apps standard (Claude Desktop, claude.ai, and others), five tools now render their results as brand-styled, interactive cards right inside the conversation: eligibility and SME-status verdicts, the grant-search results list, recommendations, and a preparation timeline. The cards follow your client’s light/dark theme and always cite the source. Clients that don’t support MCP Apps keep the exact same text answers — nothing is lost.
MCPUX
Statistics and Coverage pages redesigned
The public Statistics and Coverage pages were given a more visual design. Coverage now shows a tile map of the countries and sources in scope; Statistics adds a deadline histogram, source-composition bars, a 30-day growth series and a clearer ROI view. All figures are real aggregates from the live catalogue.
WebUX
Focus on openly licensed sources; 10 core tools
The grant catalogue now focuses on openly licensed sources: the EU Funding & Tenders Portal (Horizon Europe, Digital Europe, EDF), CORDIS, Vinnova (Sweden) and RVO (Netherlands). Licensing discussions with further national funding sources are under way, so coverage can expand on a clear legal footing. The public tool set was consolidated to 10 core tools.
SourcesTools
Cross-language search upgraded again (bge-m3)
The semantic search layer was upgraded from multilingual-e5-base to bge-m3. Hungarian → English grant matching improved substantially (MRR 0.47 → 0.84; recall@20 0.71 → 0.97), with a significant gain in pooled MRR across all languages. The 1024-dimension int8 model runs fully on our own infrastructure.
The semantic search layer was upgraded from e5-small to multilingual-e5-base. Hungarian → English grant matching improved by +28.6 pp MRR; pooled MRR across all languages improved +6.4 pp. The 768-dimension model runs locally — no third-party embedding API is involved.
API tokens now carry an expiry time and a rotating refresh token. The /oauth/revoke endpoint (RFC 7009) allows clients to explicitly revoke a token family. Reuse detection burns the whole token family on replay to protect against token theft. Admin UI hides stale tokens by default.
The new_grants_since tool (16th tool) is now live — it returns grants newly indexed since a given timestamp, enabling lightweight delta-monitoring pipelines without re-scanning the full catalogue. The /version endpoint now requires a bearer token.
ToolsMonitoring
Hungarian-language queries improved
FTS5 prefix matching was tuned for Hungarian agglutinative morphology, so queries using declined or conjugated Hungarian keyword forms now match correctly instead of returning zero results. Combined with the cross-language semantic layer, Hungarian queries reliably find relevant calls across the catalogue.
SearchBug fix
Proposal support: 3 new tools (15 total)
Three new tools are live: analyze_reviewer_feedback (structure an ESR for resubmission), check_proposal_compliance (page-limit and chapter QA), and find_consortium_partners (discover experienced partners from CORDIS history). Tool count reached 15.
ToolsWriting support
Semantic search live (local e5-small ONNX, RRF hybrid)
grant_search now supports semantic: true — a cross-language vector layer fused with FTS5 via Reciprocal Rank Fusion. A Hungarian keyword can now match an English-language call with no shared words. The recommend_grants tool was also added, ranking open grants by business-profile fit.
SearchNew feature
New grant sources: Vinnova (Sweden) and RVO (Netherlands)
The grant catalogue now covers two additional national sources: Vinnova (Sweden) and RVO (Netherlands). grant_search spans all sources by default; use source_system to scope to one.
SourcesCoverage
Public HTTPS MCP endpoint live (mcp.grantcortex.eu)
The GrantCortex MCP server is now publicly accessible over HTTPS at mcp.grantcortex.eu. Caddy reverse proxy with rate limiting (per-IP and per-token), Let's Encrypt TLS, and a daily refresh timer are all active. Any MCP-capable client can now connect without a local installation.