Pinned post
Ava is the control plane for agents that move capital.
Give any AI agent a wallet it cannot drain. Bounded on-chain execution and chain-checked receipts for agents that do DeFi.
Your coding agent can write code. It cannot safely move money.
Ava is the missing layer: a wallet the model never holds, a mandate it cannot widen, and a receipt that only counts when the chain confirms it.
@claudeai , @cursor_ai , @openclaw , codex - they connect over MCP.
They never hold the key.
Nothing spends outside a mandate.
A receipt without chain-confirmed standing is not a fill.
→ Website: getava.xyz
→ MCP: getava.xyz/mcp
THE PROBLEM?
An agent can pick the wrong venue, misread an amount, or say “done” when nothing settled.
A tx hash proves submission or inclusion.
It does not prove the agent stayed inside what you authorised.
Most agent tooling logs what it meant to do.
By the time anyone checks, claims and evidence look identical.
WHAT AVA IS?
Ava is DeFi execution for AI agents - with user-defined limits and receipts checked against chain data.
1. Agent (or you) opens a session
2. Ava provisions a Turnkey wallet - the agent never sees the key
3. You file a mandate: asset, chain, amount, allowed kinds
4. Agent proposes an action in plain English
5. Ava checks policy, mandate, and route capability
6. First execute always refuses - returns a preview + hash
7. Only a re-send with that exact hash can proceed
8. Turnkey signs. Chain settles. Ava re-reads the effect.
9. Receipt gets a standing: chain-confirmed, unconfirmed, or no proof
Only chain-confirmed counts as proven.
ONE SENTENCE → ONE SETTLED PATH
“Supply 5 solana:EPjFWdd5AufqSSqeM2qN1xzybapC8G4wEGGkZwyTDt1v to @Morpho on @base"
Same-chain Morpho USDC supply on Base, under a mandate, then a chain-confirmed receipt.
One route, proven, is the claim. n=1 is first-class.
A MANDATE IS NOT A SPEND LIMIT
A spend limit answers one question: how much can leave.
It has no opinion about whether the agent is doing the job you hired it for.
An Ava mandate fixes:
• capital - asset, amount, chain, scoped to one agent instance
• allowedKinds - a swap under a lend mandate → MANDATE_KIND_NOT_ALLOWED
• maxDrawdownPct / notional ceilings
• maxSlippageBps
• conditions - named bounds (unattended ticking is the wedge shipping next)
• mandateHash - hashed at creation, carried into the artifact the signer commits to
Nobody can widen it at runtime.
Not the model. Not a user message. Not a tool call.
THE FIRST CALL TO SPEND MONEY IS ALWAYS A REFUSAL
An MCP call is not a human saying yes.
First execute returns POLICY_HUMAN_CONFIRMATION_REQUIRED with the exact artifact + previewHash = sha256(RFC8785(preImage)).
Nothing signed. Nothing moved.
Re-send with that hash → proceeds.
Send a different one → refused again, without asking the signer for anything.
You can recompute the digest yourself from the preImage.
The LLM is not the signer, the policy, or the verifier.
RECEIPTS THAT GRADE THEMSELVES
Three standings. Never collapsed into one.
• chain-confirmed - public RPC confirmed the tx. Only this counts as proven.
• unconfirmed - an identifier exists; nothing independent checked it yet.
• no proof - nothing was submitted. The receipt says so rather than going quiet.
Refusals are counted next to wins.
An agent that never refuses anything is not being held to anything.
Typed refusal codes your agent can branch on:
MANDATE_KIND_NOT_ALLOWED
MANDATE_NOTIONAL_EXCEEDED
MANDATE_DRAWDOWN_BREACH
MANDATE_PAUSED
MANDATE_CANCELLED
MANDATE_COMPLETED
Verify any receipt with no credential:
GET /v1/receipts/:executionId/verify
ONE WALLET PER AGENT, NOT PER HUMAN
Three Claude Code windows, a Cursor tab, a Codex run = separate actors.
Each agent instance gets its own Turnkey wallet and its own mandate.
A limit set for the billing refactor does not travel to the inbox reader.
When one spends, the receipt names which instance did it.
Revoke one → that agent stops. The others keep running.
WHAT WORKS TODAY (LIVE REGISTRY)
Callable lending routes:
• Morpho Blue on @base
• Aave v3 on @avax
• Aave v3 on @BNBCHAIN
• Aave v3 on @monad
• Aave v3 on @arbitrum (wired; no settlement recorded yet)
Settled on mainnet (re-read from each chain’s node before we claim it):
• Base - Morpho supply + withdraw
• Avalanche - Aave v3 supply
• BNB Chain - Aave v3 supply
• Monad - Aave v3 supply
@SushiSwap via LI.FI settled once through a direct script - no adapter in the repo.
That’s the point: binf the calldata, not the venue list.
Generic aggregator path is prepare-only until the API can execute it honestly.
HOW TO CONNECT
Claude / Cursor / Codex:
npx @getava-xyz/connect
OpenClaw:
openclaw skills install @kamalbuilds/ava
Or point any MCP client at:
getava.xyz/mcp
claude mcp add --transport http ava api.getava.xyz/mcp
No credential. No waitlist.
ava_session → token.
15 tools by default, 27 with ?toolsets=all
Try to make it lie. Tell us if it does.
WHO THIS IS FOR?
• Agent developers who need an execution API / MCP for DeFi
• Crypto-native users who want an agent inside explicit limits
• Vault / treasury teams that need to explain what an agent was allowed to do and what the chain shows happened
• Protocols and chains that want agent activity with a verifiable execution record
THE WHOLE ARGUMENT
One-click deposit is not the hard part. Bounding it is.
Caps the model cannot widen.
A preview the agent cannot skip.
A receipt that grades itself.
A wallet per agent, not per human.
Fail closed - never invent a fill.
Give your agent a wallet it cannot misuse.
One session. One mandate. One receipt.
No key to hand over. No signup. No wallet popup.
getava.xyz
curl -s getava.xyz/health
→ { "ok": true, "productionMode": true }
Backers: @sozuhaus , @avax , @monad , @Mantle_Official , @elizaOS , @ricematcha , @ai16zeliza , @Atoma_Network , @navi_protocol , @bluefinapp , @SuiNetwork , @goat_sdk , @Zircuit , @datafdn , @hedera , @wormhole , @flow_blockchain , @BrianknowsAI , @modenetwork
Hackathon Wins:
- Cookie DeFAI Hackathon Winner (@avax Track): x.com/cookiedotfun/s…
- Monad Founder Fellowship Winner (@monad): x.com/monad_dev/stat…
- ETHDenver 2025 Winner (@EthereumDenver):
devfolio.co/projects/ava-t…
- Best Plugins to @ai16zeliza on @flow_blockchain
- First Prize: Build Multichain DeFi using the @wormhole TypeScript SDK
x.com/PauChan23/stat…
x.com/PauChan23/stat…
- @hedera Explorers: Imagine, Create, Hello Future - Prize: hedera.com/blog/announcin…
x.com/hedera/status/…
- @datafdn: Third Prize: AI Agents
x.com/Data_Ecosystem…
- @Zircuit Second Prize: Best DeFAI Project on Zircuit: zircuit.com/en/blog/zircui…
x.com/Zircuit/status…
- Unfold 2024 Winner (@avax prize #2):
devfolio.co/projects/ava-t…
x.com/avax/status/18…
- AI Winter is Coming Hackathon winner:
Main track winner: x.com/BrianknowsAI/s…
x.com/BrianknowsAI/s…
Mode Track Winner (@modenetwork):
x.com/BrianknowsAI/s…
Functor Network Track Winner (@FunctorNetwork):
x.com/BrianknowsAI/s…
Start here
- Website: getava.xyz
- Litepaper: getava.xyz/litepaper
- Docs: getava.xyz/docs
- GitHub: github.com/getava-xyz/ava…
- SDK: getava.xyz/install
- MCP: getava.xyz/mcp
- Demo Video: x.com/avatheaiagent/…
- Telegram: t.me/avatheaiagent
- Article: x.com/avatheaiagent/…
- X: x.com/avatheaiagent
- LinkedIn: linkedin.com/company/avathe…
- Discord: discord.gg/SyQsE8UgA
- Pitch Deck: getava.xyz/pitch
AVA is now live on solana:So11111111111111111111111111111111111111112 via @Pumpfun
CA: JBRRNazDJAAKezojZDTDoCistm4uqEbkG1ziPVApump
Give any AI agent a wallet it cannot drain. Bounded on-chain execution and chain-checked receipts for agents that do DeFi.
Your coding agent can write code. It cannot safely move money.
Ava is the missing layer: a wallet the model never holds, a mandate it cannot widen, and a receipt that only counts when the chain confirms it.
@claudeai , @cursor_ai , @openclaw , codex - they connect over MCP.
They never hold the key.
Nothing spends outside a mandate.
A receipt without chain-confirmed standing is not a fill.
→ Website: getava.xyz
→ MCP: getava.xyz/mcp
THE PROBLEM?
An agent can pick the wrong venue, misread an amount, or say “done” when nothing settled.
A tx hash proves submission or inclusion.
It does not prove the agent stayed inside what you authorised.
Most agent tooling logs what it meant to do.
By the time anyone checks, claims and evidence look identical.
WHAT AVA IS?
Ava is DeFi execution for AI agents - with user-defined limits and receipts checked against chain data.
1. Agent (or you) opens a session
2. Ava provisions a Turnkey wallet - the agent never sees the key
3. You file a mandate: asset, chain, amount, allowed kinds
4. Agent proposes an action in plain English
5. Ava checks policy, mandate, and route capability
6. First execute always refuses - returns a preview + hash
7. Only a re-send with that exact hash can proceed
8. Turnkey signs. Chain settles. Ava re-reads the effect.
9. Receipt gets a standing: chain-confirmed, unconfirmed, or no proof
Only chain-confirmed counts as proven.
ONE SENTENCE → ONE SETTLED PATH
“Supply 5 solana:EPjFWdd5AufqSSqeM2qN1xzybapC8G4wEGGkZwyTDt1v to @Morpho on @base"
Same-chain Morpho USDC supply on Base, under a mandate, then a chain-confirmed receipt.
One route, proven, is the claim. n=1 is first-class.
A MANDATE IS NOT A SPEND LIMIT
A spend limit answers one question: how much can leave.
It has no opinion about whether the agent is doing the job you hired it for.
An Ava mandate fixes:
• capital - asset, amount, chain, scoped to one agent instance
• allowedKinds - a swap under a lend mandate → MANDATE_KIND_NOT_ALLOWED
• maxDrawdownPct / notional ceilings
• maxSlippageBps
• conditions - named bounds (unattended ticking is the wedge shipping next)
• mandateHash - hashed at creation, carried into the artifact the signer commits to
Nobody can widen it at runtime.
Not the model. Not a user message. Not a tool call.
THE FIRST CALL TO SPEND MONEY IS ALWAYS A REFUSAL
An MCP call is not a human saying yes.
First execute returns POLICY_HUMAN_CONFIRMATION_REQUIRED with the exact artifact + previewHash = sha256(RFC8785(preImage)).
Nothing signed. Nothing moved.
Re-send with that hash → proceeds.
Send a different one → refused again, without asking the signer for anything.
You can recompute the digest yourself from the preImage.
The LLM is not the signer, the policy, or the verifier.
RECEIPTS THAT GRADE THEMSELVES
Three standings. Never collapsed into one.
• chain-confirmed - public RPC confirmed the tx. Only this counts as proven.
• unconfirmed - an identifier exists; nothing independent checked it yet.
• no proof - nothing was submitted. The receipt says so rather than going quiet.
Refusals are counted next to wins.
An agent that never refuses anything is not being held to anything.
Typed refusal codes your agent can branch on:
MANDATE_KIND_NOT_ALLOWED
MANDATE_NOTIONAL_EXCEEDED
MANDATE_DRAWDOWN_BREACH
MANDATE_PAUSED
MANDATE_CANCELLED
MANDATE_COMPLETED
Verify any receipt with no credential:
GET /v1/receipts/:executionId/verify
ONE WALLET PER AGENT, NOT PER HUMAN
Three Claude Code windows, a Cursor tab, a Codex run = separate actors.
Each agent instance gets its own Turnkey wallet and its own mandate.
A limit set for the billing refactor does not travel to the inbox reader.
When one spends, the receipt names which instance did it.
Revoke one → that agent stops. The others keep running.
WHAT WORKS TODAY (LIVE REGISTRY)
Callable lending routes:
• Morpho Blue on @base
• Aave v3 on @avax
• Aave v3 on @BNBCHAIN
• Aave v3 on @monad
• Aave v3 on @arbitrum (wired; no settlement recorded yet)
Settled on mainnet (re-read from each chain’s node before we claim it):
• Base - Morpho supply + withdraw
• Avalanche - Aave v3 supply
• BNB Chain - Aave v3 supply
• Monad - Aave v3 supply
@SushiSwap via LI.FI settled once through a direct script - no adapter in the repo.
That’s the point: binf the calldata, not the venue list.
Generic aggregator path is prepare-only until the API can execute it honestly.
HOW TO CONNECT
Claude / Cursor / Codex:
npx @getava-xyz/connect
OpenClaw:
openclaw skills install @kamalbuilds/ava
Or point any MCP client at:
getava.xyz/mcp
claude mcp add --transport http ava api.getava.xyz/mcp
No credential. No waitlist.
ava_session → token.
15 tools by default, 27 with ?toolsets=all
Try to make it lie. Tell us if it does.
WHO THIS IS FOR?
• Agent developers who need an execution API / MCP for DeFi
• Crypto-native users who want an agent inside explicit limits
• Vault / treasury teams that need to explain what an agent was allowed to do and what the chain shows happened
• Protocols and chains that want agent activity with a verifiable execution record
THE WHOLE ARGUMENT
One-click deposit is not the hard part. Bounding it is.
Caps the model cannot widen.
A preview the agent cannot skip.
A receipt that grades itself.
A wallet per agent, not per human.
Fail closed - never invent a fill.
Give your agent a wallet it cannot misuse.
One session. One mandate. One receipt.
No key to hand over. No signup. No wallet popup.
getava.xyz
curl -s getava.xyz/health
→ { "ok": true, "productionMode": true }
Backers: @sozuhaus , @avax , @monad , @Mantle_Official , @elizaOS , @ricematcha , @ai16zeliza , @Atoma_Network , @navi_protocol , @bluefinapp , @SuiNetwork , @goat_sdk , @Zircuit , @datafdn , @hedera , @wormhole , @flow_blockchain , @BrianknowsAI , @modenetwork
Hackathon Wins:
- Cookie DeFAI Hackathon Winner (@avax Track): x.com/cookiedotfun/s…
- Monad Founder Fellowship Winner (@monad): x.com/monad_dev/stat…
- ETHDenver 2025 Winner (@EthereumDenver):
devfolio.co/projects/ava-t…
- Best Plugins to @ai16zeliza on @flow_blockchain
- First Prize: Build Multichain DeFi using the @wormhole TypeScript SDK
x.com/PauChan23/stat…
x.com/PauChan23/stat…
- @hedera Explorers: Imagine, Create, Hello Future - Prize: hedera.com/blog/announcin…
x.com/hedera/status/…
- @datafdn: Third Prize: AI Agents
x.com/Data_Ecosystem…
- @Zircuit Second Prize: Best DeFAI Project on Zircuit: zircuit.com/en/blog/zircui…
x.com/Zircuit/status…
- Unfold 2024 Winner (@avax prize #2):
devfolio.co/projects/ava-t…
x.com/avax/status/18…
- AI Winter is Coming Hackathon winner:
Main track winner: x.com/BrianknowsAI/s…
x.com/BrianknowsAI/s…
Mode Track Winner (@modenetwork):
x.com/BrianknowsAI/s…
Functor Network Track Winner (@FunctorNetwork):
x.com/BrianknowsAI/s…
Start here
- Website: getava.xyz
- Litepaper: getava.xyz/litepaper
- Docs: getava.xyz/docs
- GitHub: github.com/getava-xyz/ava…
- SDK: getava.xyz/install
- MCP: getava.xyz/mcp
- Demo Video: x.com/avatheaiagent/…
- Telegram: t.me/avatheaiagent
- Article: x.com/avatheaiagent/…
- X: x.com/avatheaiagent
- LinkedIn: linkedin.com/company/avathe…
- Discord: discord.gg/SyQsE8UgA
- Pitch Deck: getava.xyz/pitch
AVA is now live on solana:So11111111111111111111111111111111111111112 via @Pumpfun
CA: JBRRNazDJAAKezojZDTDoCistm4uqEbkG1ziPVApump
View quoted post on X