I built a human approval layer for AI agent payments — looking for MCP testers

I wanted an agent to handle routine actions without giving it unlimited authority.

AskGrokWallet evaluates each proposed action against a human-written policy:

- allow — routine action can continue

- ask — pause and wait for a human

- deny — do not execute

Every outcome gets an Ed25519-signed receipt. The receipt records the agent, action, amount, destination, verdict, human

decision and, after settlement, the transaction hash. Receipt events are hash-chained, with the chain head anchored on Ethereum

Sepolia.

The guarded testnet flow now works end to end:

Grok Bot / MCP request → policy → approval inbox → Sepolia vault execution → independently verifiable receipt

Watch the walkthrough and try the no-signup demo:

Verify an anchored example yourself:

I’m looking for five Grok Bot or Cursor builders to test the MCP workflow. The feedback I need is not “looks cool,” but:

- Which real action would you put behind this gate?

- What context would be missing when you approve it?

- Which part of the receipt would you not trust?

- Where does the MCP workflow become confusing or cumbersome?

Important limitation: this is an unaudited developer preview. The guarded end-to-end settlement has been exercised on Ethereum

Sepolia with mock assets; it has not protected production funds.

1 Like

Congrats on getting this out Richard!

I tried the demo and ran your receipt verifier. The sample checked out, and changing the signed fields was caught. I found one interesting edge case: a simulated pending anchor response (blockNumber: null) still gets marked “fixed onchain” at block 0. Should that stay pending until it lands in a block? Happy to share the small repro.