Ask Claude or ChatGPT to "find road tenders in Gujarat closing this month" with nothing connected, and you will get an answer. It will be fluent, specific, formatted into a neat table — and largely invented. Tender IDs that don't resolve. Departments that don't procure that category. Closing dates pulled from nowhere.
This is not a flaw you can prompt your way out of. Language models are genuinely good at reasoning over procurement text: reading a 60-page NIT, comparing eligibility clauses, spotting a restrictive pre-qualification criterion. They are structurally incapable of fetching a tender that was published four hours ago on a state PWD portal. Nothing in their training data contains it.
The fix is to give the model a real data source it can query. That is what MCP does, and it is why a tender assistant that actually works looks less like a chatbot and more like a connected workspace.
This guide covers connecting the Tenderkart MCP server — 80+ Indian tender sources, 25,000+ tenders a day — to Claude, ChatGPT, Perplexity, Cursor, Claude Code, and Codex.
What MCP actually is
The Model Context Protocol is an open standard for connecting AI assistants to outside tools and data. Before it existed, every AI product had to build a bespoke integration with every data source, and every data source had to build one back. MCP replaces that with one contract: a service exposes a single MCP server, and any MCP-compatible assistant can talk to it.
For you, that means one connection covers Claude, ChatGPT, Perplexity, Cursor, and whatever your team switches to next year. You are not betting on a vendor.
Practically, an MCP server exposes a set of tools — named functions with typed parameters. When you ask a question in plain language, the assistant decides which tools to call, with which arguments, chains them if needed, and reasons over what comes back. You never see the plumbing.
Why Indian tender data is a hard case
Global "connect tenders to AI" tools exist, and they typically span dozens of countries by reading a handful of national portals each. That model works reasonably in jurisdictions with one dominant portal. It falls apart in India, for reasons worth naming:
There is no single portal. Central ministries publish on CPPP. Everything routed through GeM lives on GeM with a completely different bid structure. Then there are state e-procurement portals, PSU portals, railway and defence systems, municipal corporations, and a long tail of authorities that publish however they like. Reading three of them gets you a fraction of the market.
Status vocabularies don't agree. One portal's "technical evaluation" is another's "under scrutiny" is another's "opened." An assistant reasoning across raw portal data has no consistent notion of what stage a tender is at.
The tender changes after publication. Corrigenda extend deadlines, revise BOQs, and amend eligibility. A tender you screened on Monday may be a different tender by Thursday.
The money is in the documents. Tender value and closing date are the easy fields. The decision-relevant content — pre-qualification criteria, BOQ line items, evaluation methodology, deviations permitted — sits inside PDFs and Excel files.
Tenderkart normalises all of this before the model sees it: unified status buckets, a stable internal ID per tender, resolved corrigenda, and document IDs your assistant can actually open. That normalisation is the product. The MCP server is just how your assistant reaches it.
The seven tools your assistant gets
Once connected, these become available in conversation. You don't call them by name — you ask questions, and the assistant picks.
| Tool | Scope | What it does |
|---|---|---|
list_filters | filters:read | Lists the saved filters on your account, including the exact criteria each one uses |
list_filter_tenders | tenders:read | Syncs tenders from one saved filter, ordered by updated_at |
search_tenders | tenders:read | Searches the full corpus by keyword, geography, category, amount, status, and date |
lookup_tender | tenders:read | Finds one tender by portal tender ID or reference number |
get_tender | tenders:read | Returns full tender detail — description, award data, document IDs |
list_tender_documents | documents:read | Lists the documents available for a tender |
prepare_tender_document_download | documents:read | Creates a signed download URL valid for 30 minutes |
Two notes that save confusion later.
search_tenders is an optional capability enabled per API key. If your key doesn't have it, the tool is simply hidden — your assistant won't see it and won't try to use it. Everything else works off your saved filters. If you want full-corpus search, contact us to have it enabled.
The saved filters are the same ones you build in the web app. Edit a filter's keywords in your browser, and the next question you ask your AI assistant uses the new criteria. There is no separate configuration to maintain.
Setup
The server URL is the same for every client:
/mcp
Two ways to authenticate. OAuth is recommended — you click through a Tenderkart login and the AI client never touches your API key. Use the API key method only for clients that don't support OAuth.
Claude — individual Pro or Max plan
- Go to Customize → Connectors.
- Click the + next to Connectors, then Add custom connector.
- Paste
/mcpand click Add. - Click Connect and complete Tenderkart OAuth.
Then enable it in a conversation: click the + in the lower left of the chat, choose Connectors, and toggle Tenderkart on. Adding a connector doesn't switch it on everywhere by default.
Once added, the connector is shared across Claude on the web, Claude Desktop, Cowork, and the mobile apps.
Claude — Team or Enterprise
Only an Owner or Primary Owner can add a connector for the organisation.
- Owner goes to Settings → Organization Settings → Connectors.
- Add → Custom. If asked for a connector type, choose Web.
- Paste the MCP URL and confirm.
- Each member then goes to Customize → Connectors, finds Tenderkart in the list (it carries a Custom label), and clicks Connect to authenticate individually.
That last step matters: members authenticate with their own credentials rather than inheriting a shared session.
Custom connectors are available on Free, Pro, Max, Team, and Enterprise plans, though Free is limited to one custom connector. For current details see Anthropic's custom connectors guide.
ChatGPT, Perplexity, Cursor, Claude Code, Codex
Every one of these takes a remote MCP server URL. The Tenderkart MCP setup page has copy-paste configuration for each client, including the exact JSON block for the CLI-based ones. Pick your client from the tabs and follow four steps.
The pattern is always the same: paste the URL, authenticate, confirm the tool list appears.
Confirm it worked
Ask this first:
What Tenderkart filters do I have saved, and what criteria does each one use?
If you get your actual filter names back with their keywords and exclusions, you're connected. If the assistant answers from general knowledge or says it has no access, the connector isn't enabled for that conversation.
Five things to ask on day one
1. Triage the week.
Pull everything that changed on my "Solar EPC" filter in the last 7 days. Group by state, flag anything closing within 10 days, and give me the Tenderkart link for each.
2. Open one tender properly.
Look up GEM/2026/B/7406250 and summarise it: scope, estimated value, EMD, closing date, and the pre-qualification criteria. Tell me explicitly if any of those aren't stated.
3. Read the documents.
For that tender, list the available documents, then open the BOQ and give me the top 10 line items by quantity.
4. Screen against your capability.
Here is our profile: ₹40 Cr turnover, Class A electrical licence, three similar works completed in the last five years, no presence in the North East. Go through my "Substation Works" filter for this month and split the tenders into "we qualify," "we might qualify," and "we don't" — with the specific clause behind each verdict.
5. Find the quiet ones.
In my filters this month, which tenders have an EMD under ₹2 lakh but an estimated value above ₹5 crore? Those are usually less contested.
That last one is a genuinely useful screen and it's tedious to do by hand across 80+ portals.
The security question
Procurement teams are right to be careful about pointing an AI assistant at commercially sensitive data. Here is exactly what happens.
When you connect over OAuth, Tenderkart issues the AI client a short-lived token. Internally, that grant points at the access profile you selected. Three consequences follow:
- Your raw API key is never sent to the AI client. It stays on your Tenderkart account. The client holds a token that can be revoked independently.
- Account scope stays authoritative. The token can't reach data your account can't reach. If
search_tendersisn't enabled for your key, connecting through Claude doesn't unlock it. - Revocation and rate limits are enforced at the source. Revoke access in Tenderkart and the connector stops working immediately, regardless of what the client has cached.
The scopes are deliberately narrow and read-only: filters:read, tenders:read, documents:read. Nothing in this tool set can modify your filters, submit anything, or write to your account. An assistant with this connector can look, and that is all.
Document downloads work the same way. prepare_tender_document_download returns a signed URL that expires in 30 minutes and needs no authentication header — so the file can be opened or shared without ever exposing your credential.
One infrastructure note for enterprise teams: Claude connects to remote MCP servers from Anthropic's cloud, not from your laptop. Since Tenderkart is a public internet service, there is nothing to allowlist. If you have previously had trouble with self-hosted MCP servers behind a corporate VPN, that problem doesn't apply here.
The OAuth metadata is published at /.well-known/oauth-protected-resource/mcp if your security team wants to inspect it.
Limits and honest caveats
Rate limits are shared. The default is 30 requests per minute per API key, and that budget is shared across MCP and the REST API. A long research conversation can chain a dozen tool calls, so if you also run a nightly sync job on the same key, schedule them apart.
The search window is capped at 10,000 records. Asking for "every road tender in India this year" won't page through the whole set. Narrow by state and date instead — you'll get better answers and faster ones.
There is no market-stats tool over MCP. Aggregate counts and value totals come from the REST /stats endpoint. If you want "how many crores of solar tenders were published in Maharashtra last month," that's a REST call, not a conversation. See the comparison below.
Verify before you bid. The assistant reads normalised Tenderkart data, which is accurate and current, but the authoritative document is always the one on the source portal. Every response includes a tenderkart_url — use it. Treat AI output as a first pass that saves you three hours of screening, not as your submission checklist.
Ask it to admit gaps. Adding "tell me explicitly if a field isn't available" to your prompts is worth doing. It reliably converts silent guesses into stated unknowns.
When to use the REST API instead
MCP is for questions. REST is for pipelines.
If you want tenders flowing into your CRM every night, deduplicated, with a resumable cursor and an audit trail, that's the Client API — deterministic, idempotent, and built for incremental sync on updated_at. If you want a BD manager to ask "what should I look at this week" and get a reasoned answer in 20 seconds, that's MCP.
Most teams that get real value run both: REST feeding the system of record, MCP sitting in front of the humans. We break the decision down properly in MCP vs REST: which one for your procurement workflow.
Once you're connected, 20 prompts for tender research with an AI assistant is the fastest way to find out what this is actually good for.
FAQ
Do I need to be a developer to set this up?
No. The Claude and ChatGPT paths are four clicks and a paste. You need a developer only if you're building a REST integration.
Does this work with ChatGPT's free tier?
MCP client support varies by plan and changes often. Check your client's current documentation — Tenderkart's MCP setup page covers what each client needs.
Can the AI submit a bid for me?
No. Every tool is read-only. The connector can find, read, and analyse tenders. Submission happens on the source portal, by you.
Will it see tenders outside my saved filters?
Only if full-corpus search is enabled for your API key. Otherwise the assistant works within your saved filters plus direct lookup by tender ID.
How current is the data?
Tenderkart aggregates continuously across 80+ sources. Some source pipelines lag by up to 24 hours, which is why the API syncs on Tenderkart's own updated_at timestamp rather than the portal's publication date.
Can I use one API key for my whole team?
You can, but the 30 requests per minute limit is per key and shared. For a team running research conversations alongside a sync job, separate keys are cleaner.
What happens if I revoke access?
The connector stops working immediately. Revocation is enforced at Tenderkart, not at the AI client.
Ready to connect? Get your MCP URL and per-client setup at tenderkart.in/mcp-setup, or book a 30-minute demo and we'll wire it up with you.