Frequently asked questions

Pricing is still being finalised. Tester access is currently free and available right away with signup. Join the updates list and we will notify you when paid plans launch. See pricing →

Tester access is free: sign up, verify your email, and generate a tester key in your account right away. This provides full MCP server access (read scope) during the testing period. Terms may change at commercial launch.

We plan team and agency pricing, but it is not finalised yet. Contact us via the contact page and we can discuss your needs.

The content of grant searches and eligibility checks through the MCP server is never logged. The server writes one metadata row per request (timestamp, token/account ID, tool name, IP address, response status) but does not store the query content or response payload.

Account registration requires an email address. Full details: Privacy policy →

GrantCortex is an independent project operated from Hungary (EU) and treats GDPR as a baseline design principle. Data Processing Agreements (DPAs) required for user data processing will be in place before commercial launch.

The grant data itself (calls, organisations, work programmes) is indexed from source portals — this is not personal data.

No. We do not use the content of grant searches or user inputs for model training.

The MCP server and database currently run on Hetzner Cloud (Helsinki, Finland, EU). Transfers outside the EU happen only in the narrow cases listed in the Privacy policy (e.g. transactional email).

The grantcortex_check_eligibility tool uses a deterministic 4-bucket rule engine, not a probabilistic AI. This means results are reproducible and traceable, but they are not a binding decision.

• The country and organisation-type buckets are reliable when the work programme carries a fixed eligibility list
• The special-conditions bucket (free-text rules) always returns "unknown" — the engine does not interpret natural language
• The tool response always includes a note field which the agent must surface verbatim to the user

The data pipeline refreshes daily at ~03:30 UTC. Every response includes a provenance.oldest_fetched_at freshness floor that can be inspected.

For final decisions, always verify at the source portal URL — found in every response's provenance.source_urls field.

GrantCortex MCP tools read only from the daily-refreshed cache. They do not generate text beyond the raw data, and return a ParseError on empty or garbage input rather than a fabricated result. Narrative summaries and analysis are produced by the client agent (Claude, ChatGPT, etc.) — quality guarantees for those depend on the client.

GrantCortex runs on a hosted streamable-HTTP MCP server. Any MCP-capable client can connect:

• Claude.ai — Settings → Connectors → Add custom connector
• Claude Desktop — Customize → Connectors
• ChatGPT — Developer mode → Connectors
• Claude Code / Cursor / Cline — claude mcp add command

Full guide: Documentation →

Self-service: sign up, verify your email, then generate a tester API key in your account. The key is shown only once at creation — store it in a password manager. Free during the beta.

The public launch serves 10 core tools, including:

• grantcortex_search_grants — Search across 3 live sources + the CORDIS partner archive
• grantcortex_check_eligibility — Deterministic eligibility verdict
• grantcortex_recommend_grants — Rank open grants that fit a business
• grantcortex_build_prep_timeline — Prep schedule reverse-engineered from deadline
• grantcortex_get_evaluation_criteria — Award criteria from the work programme

Full list: Tool reference →

Yes. The edge applies a per-IP and per-token ceiling (approximately 120 / 60 requests per minute), with the /health endpoint exempt. If you receive HTTP 429 due to a burst, the response includes a Retry-After header — back off for that interval rather than retrying immediately.

EU funding data is huge — the full catalogue we cache is millions of words. If you connected an AI straight to the raw portals, a single question could pull back more text than the AI can hold in memory at once — and it still would not be sorted by relevance.

GrantCortex does the searching, filtering and ranking on the server. The AI receives a short, ranked answer — in our measurement about 30× less data per result than the raw records — in a single round-trip instead of dozens of separate fetches. One measured call covered 36 upstream fetches behind a single tool call.

It also searches across languages: you can ask in Hungarian and it still finds the relevant English-language EU call — something a raw API's keyword matching cannot do.

These figures come from one 2026 measurement and change as the catalogue grows; they describe efficiency, not your chance of an award. The corpus refreshes daily (not a live fetch), and every answer is a research aid, not advice.

Most sources refresh daily at ~03:30 UTC (±15 min jitter). The CORDIS archive is scheduled on a weekly cadence; the timestamp of its last refresh is shown on the Coverage page.

The current refresh status is shown in real time on the Status page.

The MCP server's delta-monitoring tools:

• grantcortex_list_new_grants — newly indexed calls since your last check
• calls_changed_since (coming soon) — updated rows (deadline changes, status updates) since a given time

These can be called on a schedule in your agent pipeline — no need to re-scan the full catalogue each run.