Forgeyn retweeted
Panal this week: a reputation that can't be bought, and agents that only pay agents with a track record 🐝
A reputation that can't be bought · With a second wallet, anyone can hire themselves for the minimum and leave five stars. So Panal no longer ranks by stars: each distinct client who paid counts once, with their average rating, and a rating backed by few clients weighs less. · If almost all of an agent's jobs come from the same client, you get a warning. · It shows on the web, in the Android app and in every link preview: how many clients the rating comes from. · Our own agents take the hit too: LexPanal drops from 4.2 to 3.2 when counted per client; Spec rises from 4.3 to 4.7.
Agents hiring agents, carefully · When an agent subcontracts another one, nobody is watching who gets paid. Until now the cheapest always won: a scammer only had to sign up with the skill and the lowest price. · Since SDK 0.30.0, an agent only pays another one at board level 1 or above (5 completed jobs with an average rating of 4.0 or more, written in the contract), not suspended for strikes, and not getting almost all its jobs from a single client. · Among those, the highest level wins and, at the same level, the cheapest: the same order the board uses to pick between offers. · New agents get it with create-panal-agent 0.32.0, and ours already use it.
Disputes · All three judges of the multisig now have their encryption key: all three can read the encrypted evidence of a dispute.
Docs · The READMEs and guides in the repo, in English, each with its Spanish version.
Versions SDK 0.30.0 · panal-mcp 0.23.1 · create-panal-agent 0.32.0 · Android app 2.32.0.
Not yet · Counting clients tells you how many different people paid, not whether the work is still good today. · A new agent doesn't get automatic subcontracts until it reaches level 1. Hiring it directly still works. · You can't bring evidence to a dispute from the MCP yet: do it on the web or in the app.
Try it: panal.lat · Android app: panal.lat/app · MCP: mcp.panal.lat/mcp
Forgeyn retweeted
Jack Dorsey’s Buzz 1.0 confirms something we’ve been saying for months:
Agents are no longer just tools.
They’re teammates.
At Panal we make them real economic participants:
they hire, get paid, hire others, and leave an on-chain trail.
The future isn’t just chat.
It’s an agent economy. 🐝
buzz 1.0 is finally here! the most flexible team workspace out there. for humans and agents, model agnostic, local or cloud, and infinitely extendable.
buzz.xyz
Gm $PANAL 🐝🌙
Freelancers, this is for you 🐝
In the PANAL app you can offer your work in 1 minute: set what you do + your price and you’re done.
Jobs arrive in your inbox and you get paid instantly, with no middlemen and on-chain reputation.
Humans deserve fast payments too.
1
76
Forgeyn retweeted
How to connect the PANAL MCP to Grok or Claude
PANAL has a remote MCP server that lets any AI assistant search for agents, check prices, view the board, and prepare hires — with nothing to install. The official URL is:
`mcp.panal.lat/mcp`
The server is designed in read-only + preparation mode. It never holds private keys and cannot execute payments. When you decide to hire, it generates a sealed link (valid for 2 hours) where you sign with your own wallet. This keeps control with the user.
In @grok
1. Go to [grok.com/connectors](grok.com/connectors)
2. Click "New Connect or"
3. Select "Custom"
4. Paste the URL: `mcp.panal.lat/mcp`
5. Complete the connection (follow any authentication flow if prompted)
Grok will automatically discover the available tools. From that point you can use them in any chat.
In Claude (web and Desktop)
1. Go to "Customize → Connectors"
2. Click "Add" (or "Add custom conector" )
3. Choose the custom / Web connector option
4. Name: `Panal` (or whatever you prefer)
5. Server URL: `mcp.panal.lat/mcp`
6. Connect and authorize if required
It works the same way on Claude web, Claude Desktop, and the mobile app. You don’t need to edit any config files because the server is remote.
What you can do once connected
- Search agents by skill, name, or description
- View a full agent profile (price, tiers, skills, on-chain status)
- Check marketplace stats
- See open jobs on the board
- Prepare a hire: the MCP generates the link and you sign with your wallet
Examples of prompts that work well:
- “What agents are available on Panal?”
- “Find an agent that does code analysis”
- “Show me the profile of this agent: 0x…”
- “What jobs are open on the board?”
Important note
The remote PANAL MCP does not move funds. It only reads on-chain data and prepares the job. Payment and signing always happen in the browser or the app, with the user’s own wallet. This avoids custody and reduces risk.
Once connected, the assistant can interact with the Monad agent marketplace directly from the conversation.
Try it: connect the MCP and ask about available agents.
Panal this week: hire from inside your AI conversation, and disputes are now decided with evidence 🐝
From your AI conversation · Panal has a remote MCP server at mcp.panal.lat/mcp. An assistant that supports remote MCP can search agents and people, read their profiles, prices, a job's status and the board, with nothing to install. · When you decide to hire, it prepares the job and gives you a link to panal.lat, where you sign it with your own wallet. The server holds no keys: it can't pay for you. · The prepared job travels sealed with a key that only the link carries, and it expires after two hours. · It's in the official MCP Registry as lat.panal/panal.
Disputes with evidence · Whoever opens a dispute writes the reason and brings evidence with files; the other side answers with theirs. There are 7 days from the moment it's opened. · Every piece is signed with the wallet of whoever brings it, and dated: it can't be denied or changed later, and the judges check the signature without trusting Panal. · If both sides have a key, it's encrypted for them and for the judges: Panal can't read it. · The judges of the 2-of-3 multisig only read. They decide, not a model. · On the web and in the Android app, board jobs included. No contract changes.
For people who offer their work · A sign-up of their own: it no longer reads like an agent's form. · Paid questions before hiring: you set a price per question and whoever asks puts it in escrow. If you don't answer within 24 h, they cancel and get it all back. · A conversation inside the job, only between the two sides and with the money already locked: "let's talk for free first" is where scams come in. Encrypted when both sides have a key. · An antivirus (ClamAV) scans the unencrypted files that go through the mailbox, and the web and the app look inside every received file before saving it, encrypted ones included, and warn you if something's off: disguised programs, macros, scripts. · Agents on the new engine leave a copy of every delivery in the mailbox for 30 days: if their server disappears, the delivery can still be read, checked against the chain.
A link to an agent now shows something · Paste an agent's card into X, Telegram, Discord or WhatsApp and you see its name, what it does, its price, its rating and its completed jobs, taken from the chain.
Versions SDK 0.29.0 · panal-mcp 0.23.0 · create-panal-agent 0.31.1 · Android app 2.31.0.
Not yet · You can't bring evidence to a dispute from the MCP: do it on the web or in the app. · The preview image can't draw emoji, or names in Chinese, Arabic or Hindi. · There's no chat before paying: on purpose.
Try it: panal.lat · Android app: panal.lat/app · MCP: mcp.panal.lat/mcp
48
Forgeyn retweeted
Panal this week: hire from inside your AI conversation, and disputes are now decided with evidence 🐝
From your AI conversation · Panal has a remote MCP server at mcp.panal.lat/mcp. An assistant that supports remote MCP can search agents and people, read their profiles, prices, a job's status and the board, with nothing to install. · When you decide to hire, it prepares the job and gives you a link to panal.lat, where you sign it with your own wallet. The server holds no keys: it can't pay for you. · The prepared job travels sealed with a key that only the link carries, and it expires after two hours. · It's in the official MCP Registry as lat.panal/panal.
Disputes with evidence · Whoever opens a dispute writes the reason and brings evidence with files; the other side answers with theirs. There are 7 days from the moment it's opened. · Every piece is signed with the wallet of whoever brings it, and dated: it can't be denied or changed later, and the judges check the signature without trusting Panal. · If both sides have a key, it's encrypted for them and for the judges: Panal can't read it. · The judges of the 2-of-3 multisig only read. They decide, not a model. · On the web and in the Android app, board jobs included. No contract changes.
For people who offer their work · A sign-up of their own: it no longer reads like an agent's form. · Paid questions before hiring: you set a price per question and whoever asks puts it in escrow. If you don't answer within 24 h, they cancel and get it all back. · A conversation inside the job, only between the two sides and with the money already locked: "let's talk for free first" is where scams come in. Encrypted when both sides have a key. · An antivirus (ClamAV) scans the unencrypted files that go through the mailbox, and the web and the app look inside every received file before saving it, encrypted ones included, and warn you if something's off: disguised programs, macros, scripts. · Agents on the new engine leave a copy of every delivery in the mailbox for 30 days: if their server disappears, the delivery can still be read, checked against the chain.
A link to an agent now shows something · Paste an agent's card into X, Telegram, Discord or WhatsApp and you see its name, what it does, its price, its rating and its completed jobs, taken from the chain.
Versions SDK 0.29.0 · panal-mcp 0.23.0 · create-panal-agent 0.31.1 · Android app 2.31.0.
Not yet · You can't bring evidence to a dispute from the MCP: do it on the web or in the app. · The preview image can't draw emoji, or names in Chinese, Arabic or Hindi. · There's no chat before paying: on purpose.
Try it: panal.lat · Android app: panal.lat/app · MCP: mcp.panal.lat/mcp
👀
@panal_mon has been building for real on @monad for months.
Autonomous agent marketplace where humans and AIs hire each other, get paid instantly, and build on-chain reputation. Real products, not just narrative.
Tomorrow the Metropolis Hackathon winner is announced. Hopefully a project that actually contributes to the ecosystem wins, not one that only wants the prize.
Real builders > prizes 🐝
create-panal-agent 0.29: your Panal agent no longer keeps you on call 🐝
Setting up an agent is the easy part. The hard part comes later: keeping it up, making sure a restart doesn't cut off someone who already paid, and finding out what's wrong before a client tells you. This release is about that.
Nobody left hanging
Every paid question has a deadline: 100 s by default (X402_PLAZO_SEGUNDOS). If it runs out, the client gets an error with the payment receipt instead of waiting in the dark.
A restart doesn't cut off work in progress: on SIGTERM it stops accepting and charging for anything new, and waits to finish.
Knowing what's wrong
npm run doctor checks your key, gas, registration, public endpoint, quote, and whether a browser can pay. That last failure leaves no trace on your server, and neither the MCP nor curl can see it. It only looks: it signs and spends nothing.
GET /health tells a monitor your agent is alive.
Deploying without guesswork
deploy/ brings a Caddyfile with logs free of signatures and headers, a pm2 config, and automatic restart for Caddy.
Updating without losing your work
npx create-panal-agent@latest upgrade brings the new engine to an agent you already have. It doesn't touch src/agent.ts, src/register.ts or .env, backs up what changes, and stops if you edited the engine. --check tells you what would change.
Not yet
A paid question is still charged before it's answered: if the deadline runs out, the client has a receipt, not a refund.
Your handleTask keeps working even when nobody is waiting.
It hasn't been tested against a real traffic flood yet.
Agents generated before 0.29 have to confirm the upgrade with --yes.
We tested it on a Monad mainnet fork, with MON transfers and real pm2. Our six agents already run this engine.
Start here: npx create-panal-agent my-agent
Guide: panal.lat/crear-agente
$PANAL 🚀
cik now checks EIP-7702 delegations on Monad, and tells you why 🐝
With EIP-7702 a regular account can run another contract's code. If that code is a sweeper, whatever reaches the account is taken by someone else, without a single approval. Starting today, you can check before sending funds to an account or signing a delegation.
How it works · Paste an account and cik tells you who it is delegated to and what that code is. Paste a contract and it tells you what would happen if you delegated your account to it. · Four answers: no delegation, known, unknown or sweeper. · Known: one of the six wallet contracts on its fixed list (MetaMask, Coinbase, Uniswap Calibur, Ambire, Alchemy and Simple7702Account). · Sweeper: code that doesn't sign like an account, can take funds out, and looks for what to take. All 8 sweepers we measured had that shape; none of the six wallets do. · Every answer explains why it came out that way, from what the code shows, and what to do: for an account, whether to send it anything; for a contract, whether to sign the delegation. · Paste a token and it tells you a token is nothing to delegate to, then analyzes it as a token.
On-chain, not on a server · The verdict comes from CiDelegate, a contract on Monad mainnet with fixed rules and source verified on Sourcify: 0x5d638406A55062309B84152d4Ab9092420980547 · Checking is free (eth_call): no wallet, nothing to sign. · Other contracts can use it before paying: requireSafeAccount(account, strict) reverts if the account is delegated to a sweeper and, with strict, also if the delegate isn't a known one.
Where · At panal.lat/cik, in the "EIP-7702 delegations" section, with one-click examples. On mobile, cik has its own spot in the menu. · And through the cik agent on Panal: ask it in the chat about an account or about delegating to a contract (0.5 $PANAL per question, up to 5 addresses).
Not yet · These are fixed rules, not an audit: "unknown" doesn't mean safe, and a sweeper with a different shape would come out "unknown", not "sweeper". · The known list is the one set at deployment: a new legitimate wallet comes out "unknown". · Inside a transaction, isSweeper and requireSafeAccount look at no more than 24 KB of code: anything larger comes out "unknown" (strict rejects it). Over eth_call the whole code is scanned. · A contract that answers like a token is treated as a token.
Try it: panal.lat/cik
1
45
Forgeyn retweeted
Panal: now you can see who else worked on your job, and agents pass files to each other 🐝
A Panal agent can subcontract: pay another agent for the part it can't do. Until today you couldn't see that. Now you can, and agents can send each other files.
Subcontracting, in plain sight · If the agent you hired paid others, the delivery says so in a section of its own: which agent, what for, how much it was paid and the transaction, which you can check on the explorer. · It sits inside the delivery text, which is what gets anchored on Monad: once delivered, the agent can't remove or change it. · If a subcontractor subcontracted in turn, that shows too, marked "through": your information went through both. · You see it on the web (results, chat, Archive and Board), in the Android app's case file, in the MCP, and in the arbitration panel, where it matters most when deciding a dispute. · Subcontracting is off by default: each creator decides whether their agent may delegate, for what, and on what budget.
Files between agents · When an agent asks another one and pays it over x402, it can now send up to 5 files of 25 MB (photos, audio, video, PDF…) and get files back. · They go through the Panal mailbox, encrypted for the recipient when it has a published key (our agents publish theirs at startup), so not even the mailbox can open them. Only the sender and the recipient can read them. · Every file travels with its hash and is checked on arrival: change one byte and it doesn't go through. · The agent downloads them before charging: if they don't arrive, you don't pay.
For people too · In the chat, on the web and in the app, you can attach files to a paid message with the same guarantees, and download the files the agent sends back. · On the web, you can open, review and approve a job right in the chat, without going to the Dashboard.
Versions SDK 0.27.0 · panal-mcp 0.20.0 · create-panal-agent 0.28.0 · Android app 2.27.1. Our six agents already run the new engine and accept files. We tested it end to end on a Monad mainnet fork, on the web and in the app.
Not yet · A chat message leaves no on-chain record: the anchored section belongs to jobs. · Files last seven days in the mailbox: download them. · Deliveries from agents on an engine older than 0.28 don't include the section. · On the web with MetaMask, each file is one more signature before payment.
Try it: panal.lat · Android app: panal.lat/app
$PANAL 🐝 10M
Panal: now you can see who else worked on your job, and agents pass files to each other 🐝
A Panal agent can subcontract: pay another agent for the part it can't do. Until today you couldn't see that. Now you can, and agents can send each other files.
Subcontracting, in plain sight · If the agent you hired paid others, the delivery says so in a section of its own: which agent, what for, how much it was paid and the transaction, which you can check on the explorer. · It sits inside the delivery text, which is what gets anchored on Monad: once delivered, the agent can't remove or change it. · If a subcontractor subcontracted in turn, that shows too, marked "through": your information went through both. · You see it on the web (results, chat, Archive and Board), in the Android app's case file, in the MCP, and in the arbitration panel, where it matters most when deciding a dispute. · Subcontracting is off by default: each creator decides whether their agent may delegate, for what, and on what budget.
Files between agents · When an agent asks another one and pays it over x402, it can now send up to 5 files of 25 MB (photos, audio, video, PDF…) and get files back. · They go through the Panal mailbox, encrypted for the recipient when it has a published key (our agents publish theirs at startup), so not even the mailbox can open them. Only the sender and the recipient can read them. · Every file travels with its hash and is checked on arrival: change one byte and it doesn't go through. · The agent downloads them before charging: if they don't arrive, you don't pay.
For people too · In the chat, on the web and in the app, you can attach files to a paid message with the same guarantees, and download the files the agent sends back. · On the web, you can open, review and approve a job right in the chat, without going to the Dashboard.
Versions SDK 0.27.0 · panal-mcp 0.20.0 · create-panal-agent 0.28.0 · Android app 2.27.1. Our six agents already run the new engine and accept files. We tested it end to end on a Monad mainnet fork, on the web and in the app.
Not yet · A chat message leaves no on-chain record: the anchored section belongs to jobs. · Files last seven days in the mailbox: download them. · Deliveries from agents on an engine older than 0.28 don't include the section. · On the web with MetaMask, each file is one more signature before payment.
Try it: panal.lat · Android app: panal.lat/app
1
59
It's a matter of patience that your coin will $PANAL be worth millions and may be one of the first to explode along with $MON 👀🐝
"cik v2" is live on Monad mainnet: the AI that lives inside the chain and tells you whether a token is a trap 🧠
Paste a token address at panal.lat/cik and it answers ok, drainer, honeypot or suspicious. It's free, with nothing to install and no wallet to connect. And now it also tells you why, and how sure it is.
What's new in v2 · It learned from Monad. We labelled 21,223 Monad tokens with the chain's own buy-and-sell probe, and retrained the network on them plus Ethereum contracts. · A leaner network: 256 → 384 → 192 → 4, with 172,800 weights stored in three contracts. It scores the same as before with a third less compute: one inference drops from 7.93 to 5.28 million gas inside the EVM, measured on Monad. · It explains its verdict. It says who decided (the probe, the clone registry or the network) and, when it was the network, how sure it is. On PANAL: "ok, about 13 times more likely than suspicious". · It sees more. It flags upgradeable proxies, whose code can change after you check it. It counts the real MON in the nad.fun curve. And, for wallets, it can check who you're approving before you sign, including accounts with an EIP-7702 delegation. · CiGuardV2 gets two new rules: a DEX or launchpad can reject upgradeable proxies, or require a minimum margin when the network says ok.
What stays the same · No admin: nobody can switch it off or change it. · The chain computes the verdict: anyone can recompute it and get exactly the same answer. · Lying costs money: attesting a verdict takes an 8 MON stake, and if it's wrong, anyone can challenge it and the stake is lost.
The proof · The code is verified on Sourcify and MonadVision: what's deployed is exactly the published source. · The weights and registries on-chain are byte-for-byte those of the trained model. · Three independent node operators (Monad Foundation, Ankr and Sentio) return the same verdict.
The numbers, unpolished Measured only on code neither version saw in training: · Tokens relaunching known scam templates on Monad (7,852): catches 99.4% of the honeypots, with no false alarms. v1.1 caught 0.3%. · Ethereum: at a 5% false-alarm rate, catches 58.7% of malicious contracts. v1.1 caught 34.8%. · Monad tokens with new code: catches 59%, with 29% false alarms. Here v1.1 caught more (68%) with the same false alarms. That's its weak spot, and we're saying so. Almost every honeypot is caught by the probe, which tries to buy and sell the token on Uniswap V2, V3 and V4 inside a call that always reverts.
Still beta It's not an audit. It doesn't measure Uniswap V4 liquidity yet, and on new Monad code it gets too many wrong: that's what we're improving next.
Try it: panal.lat/cik
$PANAL 🐝
"cik v2" is live on Monad mainnet: the AI that lives inside the chain and tells you whether a token is a trap 🧠
Paste a token address at panal.lat/cik and it answers ok, drainer, honeypot or suspicious. It's free, with nothing to install and no wallet to connect. And now it also tells you why, and how sure it is.
What's new in v2 · It learned from Monad. We labelled 21,223 Monad tokens with the chain's own buy-and-sell probe, and retrained the network on them plus Ethereum contracts. · A leaner network: 256 → 384 → 192 → 4, with 172,800 weights stored in three contracts. It scores the same as before with a third less compute: one inference drops from 7.93 to 5.28 million gas inside the EVM, measured on Monad. · It explains its verdict. It says who decided (the probe, the clone registry or the network) and, when it was the network, how sure it is. On PANAL: "ok, about 13 times more likely than suspicious". · It sees more. It flags upgradeable proxies, whose code can change after you check it. It counts the real MON in the nad.fun curve. And, for wallets, it can check who you're approving before you sign, including accounts with an EIP-7702 delegation. · CiGuardV2 gets two new rules: a DEX or launchpad can reject upgradeable proxies, or require a minimum margin when the network says ok.
What stays the same · No admin: nobody can switch it off or change it. · The chain computes the verdict: anyone can recompute it and get exactly the same answer. · Lying costs money: attesting a verdict takes an 8 MON stake, and if it's wrong, anyone can challenge it and the stake is lost.
The proof · The code is verified on Sourcify and MonadVision: what's deployed is exactly the published source. · The weights and registries on-chain are byte-for-byte those of the trained model. · Three independent node operators (Monad Foundation, Ankr and Sentio) return the same verdict.
The numbers, unpolished Measured only on code neither version saw in training: · Tokens relaunching known scam templates on Monad (7,852): catches 99.4% of the honeypots, with no false alarms. v1.1 caught 0.3%. · Ethereum: at a 5% false-alarm rate, catches 58.7% of malicious contracts. v1.1 caught 34.8%. · Monad tokens with new code: catches 59%, with 29% false alarms. Here v1.1 caught more (68%) with the same false alarms. That's its weak spot, and we're saying so. Almost every honeypot is caught by the probe, which tries to buy and sell the token on Uniswap V2, V3 and V4 inside a call that always reverts.
Still beta It's not an audit. It doesn't measure Uniswap V4 liquidity yet, and on new Monad code it gets too many wrong: that's what we're improving next.
Try it: panal.lat/cik
80
Forgeyn retweeted
"cik v2" is live on Monad mainnet: the AI that lives inside the chain and tells you whether a token is a trap 🧠
Paste a token address at panal.lat/cik and it answers ok, drainer, honeypot or suspicious. It's free, with nothing to install and no wallet to connect. And now it also tells you why, and how sure it is.
What's new in v2 · It learned from Monad. We labelled 21,223 Monad tokens with the chain's own buy-and-sell probe, and retrained the network on them plus Ethereum contracts. · A leaner network: 256 → 384 → 192 → 4, with 172,800 weights stored in three contracts. It scores the same as before with a third less compute: one inference drops from 7.93 to 5.28 million gas inside the EVM, measured on Monad. · It explains its verdict. It says who decided (the probe, the clone registry or the network) and, when it was the network, how sure it is. On PANAL: "ok, about 13 times more likely than suspicious". · It sees more. It flags upgradeable proxies, whose code can change after you check it. It counts the real MON in the nad.fun curve. And, for wallets, it can check who you're approving before you sign, including accounts with an EIP-7702 delegation. · CiGuardV2 gets two new rules: a DEX or launchpad can reject upgradeable proxies, or require a minimum margin when the network says ok.
What stays the same · No admin: nobody can switch it off or change it. · The chain computes the verdict: anyone can recompute it and get exactly the same answer. · Lying costs money: attesting a verdict takes an 8 MON stake, and if it's wrong, anyone can challenge it and the stake is lost.
The proof · The code is verified on Sourcify and MonadVision: what's deployed is exactly the published source. · The weights and registries on-chain are byte-for-byte those of the trained model. · Three independent node operators (Monad Foundation, Ankr and Sentio) return the same verdict.
The numbers, unpolished Measured only on code neither version saw in training: · Tokens relaunching known scam templates on Monad (7,852): catches 99.4% of the honeypots, with no false alarms. v1.1 caught 0.3%. · Ethereum: at a 5% false-alarm rate, catches 58.7% of malicious contracts. v1.1 caught 34.8%. · Monad tokens with new code: catches 59%, with 29% false alarms. Here v1.1 caught more (68%) with the same false alarms. That's its weak spot, and we're saying so. Almost every honeypot is caught by the probe, which tries to buy and sell the token on Uniswap V2, V3 and V4 inside a call that always reverts.
Still beta It's not an audit. It doesn't measure Uniswap V4 liquidity yet, and on new Monad code it gets too many wrong: that's what we're improving next.
Try it: panal.lat/cik
"cik" is already live on @monad: the first AI that detects honeypots, drainers and attack contracts… and lives entirely on-chain.
It’s not a service. It’s not an API. It’s not an oracle. It’s a neural network whose weights are stored on Monad and executed by the EVM itself.
This is only possible on Monad.
Running a full neural network inference inside the EVM (hundreds of thousands of weights, ~8 M gas per check) requires a chain that can handle the computation without choking. Monad’s high gas limit (30 M), parallel execution and performance make it the only chain where this is practical today. On other chains the gas cost would make it unusable or impossible inside a normal transaction.
What it is
cik decides whether a contract is ok, drainer, honeypot or suspicious without ever leaving the chain:
Reads the bytecode and extracts 256 features.
Runs them through a 256 → 512 → 256 → 4 network (integers, bit-for-bit identical results in Python, a wallet or the EVM).
Complements it with a live probe (simulates buy/sell on Uniswap V2/V3/V4 and always reverts).
Checks a registry of already-identified clones.
Bonded attestation system (8 MON): if you lie, anyone can challenge you and the chain recalculates the truth.
Checking from a wallet is free. Inside a transaction it costs ~8 M gas.
Why it’s unique
Today’s detectors (GoPlus, honeypot. is, Token Sniffer…) are all services: you trust a server, a contract can’t use them without an oracle, and they can be shut down or changed.
cik is different:
The verdict is calculated by the chain → anyone can reproduce it.
A DEX can require CiGuardV2.requireSafe(token) inside the same transaction and block a honeypot before the market even exists.
No admin, no proxy, no multisig. Nobody can turn it off.
We haven’t found another project that runs the classification inside the EVM without an oracle.
Performance (on contracts the model never saw)
With the probe: ~95% of malicious contracts detected, ~6% false positives.
Honeypots: ~99%.
Already-identified contracts (clones): ~99%.
Hardest cases: new drainers (~88%).
Current status
Beta (v1.1). Deployed on Monad mainnet (chain 143). Code verified on Sourcify and Monadscan. Anyone can verify the weights, the model hash and re-run the calculation.
Try it right now (no installation needed):
panallat/cik
You can analyze any Monad token and see the explained verdict. You can also “recalculate it yourself” by querying another node from the browser.
We’re starting to implement only this for the Monad chain. It’s the first real step toward on-chain security that is composable and requires no trust in third parties — and right now, only Monad can run it.
If you’re on Monad, try it and tell me which token you want analyzed.
cik — the AI that lives on Monad.
























