I’ve been using the Klever Connect MCP — the AI knowledge server for the @klever/connect JS/TS SDK — and wanted to share a small demo, because the result was cleaner than I expected.
Briefly, what it is: it’s not for writing smart contracts. It’s for any app that talks to Klever Blockchain — Node services, React frontends, browser dApps. It hands your AI assistant the full Connect SDK knowledge base (code examples, React hooks, wallet patterns, provider config, error handling, API reference) so it writes correct code on the first try instead of guessing method names. It’s read-only by design: it doesn’t hold keys, doesn’t sign anything, doesn’t touch the chain. That’s the SDK’s and your wallet’s job.
The demo
I started with a plain React app — just a state counter, nothing Klever-related. Goal: turn it into an address-balance lookup.
I used Claude Code for this (npm install -g @anthropic-ai/claude-code if you don’t have it). Setup is one command, and there are two paths:
Hosted (what I used — always current, nothing to maintain):
claude mcp add -t http klever-connect https://ai.klever.org/sdk-js
Local via npx (offline / full control, but you keep it current yourself):
claude mcp add klever-connect -- npx -y @klever/klever-connect-mcp
Verify with claude mcp list — you should see klever-connect in the output.
Then a single prompt:
Based on this React app, update the main page to have an input that, when you click search, gets the balance of that account in KLV. Use the Klever Connect MCP to know how to query this kind of data.
Then I let it run, hands off. On its own it:
- noticed the SDK wasn’t installed and ran
npm install @klever/connect - queried the MCP for the right balance APIs
- wrote a
fetchBalancethat strips stray characters from the input and validates it withisValidAddress - wired up
KLVProvideron mainnet - used
formatKLVto convert raw on-chain integer units into a human-readable number - updated the page with the input + button + formatted result
- ran
npm run buildto confirm it actually compiled
I tested it with a real address pulled off KleverScan, and the app returned 43M KLV — matching KleverScan to the decimal. One prompt, working integration.
What I didn’t have to figure out
These are the bits people usually get wrong the first time, and exactly what generic AI training data tends to hallucinate:
- Package install —
@klever/connect, added without me asking - Address validation —
isValidAddress, not a half-remembered regex - Provider selection —
KLVProvideron mainnet (testnet would’ve been just as easy to ask for) - Decimal formatting —
formatKLVto turn raw units into readable KLV - Imports + types — TypeScript-friendly out of the box
The MCP routed Claude through the SDK’s own examples instead of letting it guess.
Going further
The demo only touches the most basic surface. The same MCP also knows how to build and sign transactions, call smart contracts through the SDK, use Klever Wallet’s React hooks for connect/disconnect flows, handle multi-asset balances (KLV, KFI, custom KDAs), and deal with error patterns like user-rejected signatures, nonce mismatches, and broadcast failures. The workflow is the same each time: ask in natural language, let it pull from the MCP, iterate.
Links
- SDK docs — docs.klever.org
- MCP landing page (install snippets per AI client) — ai.klever.org
- Launch video — YouTube walkthrough
If you’d rather watch than read, the video covers the whole thing end to end.
Curious what you’d build with it
If anyone here has already tried the MCP, did you run into anything weird — multi-asset balances, the hosted vs. local setup, error handling on signatures? And if you haven’t tried it, what integration would you want to test first? Happy to answer questions in the thread.