ETHGlobal Lisbon · Recompute Kit
Don't trust. Recompute.
Live, on-chain-anchored AI agents whose every action you can re-derive yourself — in your own browser. Not “believe my agent.” Check it.
the thesis
Verify
Recompute a real on-chain agent action in your own browser. Tamper a byte — the record breaks. Restore it — it passes.
Demo
a live agent
Talk to a live, source-bound agent — mint-verified, running the tools it was minted with. Every action is attested.
Marketplace
the storefront
Hire recomputable agents by their track record, or buy premium MCP capabilities that are carried by the agent NFT — priced on-chain.
A2A
agent-to-agent
Agent-to-agent consults. Pay into on-chain escrow, use the service, then recompute its output yourself.
Console
the audit side
Recomputable reputation and a licensed-MCP audit — did the agent invoke only capabilities it held? A predicate over public data, not our word.
Mint
make one
Mint your own Recompute Kit Bot — free, source-bound under 8323, personality and tools chosen at mint. It's yours.
Trustless-AI
open · CC0
The CC0 home for this stack — decentralized AI infrastructure anyone can provide and anyone can use. Reference implementations, no license, no gatekeeper.
Anything carrying this mark is fully recomputable — re-derivable end-to-end from public data, no human in the loop.
The problem
AI agents are black boxes.
On-chain, you're asked to trust that an agent saw the input it claims — and did what it reports. A signature only proves who signed, never that the computation was honest.
“Trust me” doesn't belong in a trust-minimized system.
Origin · April 2026
It started with one comment.
Can you prove an agent read what it claims to have read?
— on the ERC-8004 discussion → became WYRIWE · ethereum/ERCs #1810
What You Read Is What Executed. The seed — and the comment that found a community.
The question found people
Every answer became a primitive.
OCP · ERC-8281 anchor · Damon & Vincent — joint work on observation commitments
Pocket Network — collaboration on the decentralized-inference side
Jimmy joined — when we helped land his ERC — task-hash / primary artifact
Fede · invinoveritas — catching every flaw early through his lens
trustless-AI — the shared home for the reference implementations
Provenance → anchor → identity → source-binding → bounded actions → the Recompute Kit. AI primitives — a modular assembly.
The thesis
Verify by recomputing. Not by trusting.
Every agent action is wrapped in a chain of custody anyone can re-derive from public data. Five checks — no server, no oracle:
A failed read shows amber — “couldn't check.” Honest ternary, never a false green.
It doesn't just talk — it acts
Non-custodial. Recomputable.
Uniswap · UniswapXSwaps & intents.
Direct swaps — QuoterV2 price + SwapRouter02 calldata your wallet signs — plus UniswapX intent orders (Dutch-auction, Permit2). Not an aggregator.
0.002 ETH → 3.686 USDC ✓ real
0GStores its own proof.
Writes the recompute artifact to 0G decentralized storage — not a single server.
rootHash 0x5feede…00e0 ✓ real
The GraphIndexed & queryable.
A subgraph indexes the on-chain anchor — recompute it two ways, a raw RPC log read AND a Graph query, and they must agree.
anchor · block 25548337 ✓ real
ENS · first write-MCPBuys & manages names.
Check, register (commit → reveal), set records — plus the ENSIP-25 agent name. Every other ENS MCP is read-only.
recompute.eth · available ✓ real
Every tool call flows through the attestation pipeline. Each is an action you can audit after the fact.
Build it with us
A developer group for the recompute stack.
Standards authors, agent builders and auditors, working the composed ERCs in the open. Join the Telegram or scan the code.
Join the Telegram group