Bansun Docs — Features, privacy, AI and API

Bansun is a private AI agent and a toolkit for reading Ethereum before you sign. Open the app without creating an account. Most tools do not require a connected wallet; the optional approval-revoke action asks your browser wallet to connect and confirm a transaction.

1. Open Bansun and choose a tool.
2. Use public addresses, ENS names, transaction hashes or contract data. Never paste a seed phrase or private key.
3. Read the result, its sources, units, time and warnings. A risk score or confident answer is not a safety guarantee.
4. For personal records, choose encrypted Vault storage and export a backup before clearing browser data.

The PRIVATE label means a tool runs on this device; it does not mean the backend is broken. Live market and blockchain lookups need an internet connection. Server AI and browser AI are separate modes.

Agent combines server AI with Ethereum and market tools. Try “What is gas now?” or enter a public token address. Tool receipts show which sources were used and how long requests took. The agent does not automatically read your saved journal, watchlist or vault.

• Chat: messages go to the Bansun backend and its configured AI server, without live tool calls.
• Compare: ask the configured server models the same question. Agreement is a comparison signal, not a correctness test.
• Local AI: after downloading a browser model, prompts and inference stay on this device. Local AI has no live Ethereum or market tools.
• Auto-erase 10 min: clears chat held in app memory. It does not remove copied text, downloaded files, or a separate answer saved to the vault.
• Save to vault / Log prediction: keep an answer encrypted or track a checkable forecast.

Current default server models are Qwen2.5 1.5B and SmolLM2 1.7B. Read numeric values and units against the original tool result. Small models can make mistakes even when their confidence is high.

Open AI agent and chat in Bansun →

There are nine browser choices. Bansun checks for an actual WebGPU adapter, not just the presence of a browser API. Without a usable GPU adapter, the CPU model remains available.

1. Open Ask → Local AI.
2. Choose a supported model and explicitly load it while online. The CPU model downloads about 0.14 GB of model data; GPU models can require much more space and memory.
3. Wait for the model and engine cache to finish. “Loaded” alone is not the same as “offline reload ready.”
4. After the offline-ready state, reopen the app on the same origin and browser. Clearing storage, private browsing, or changing domains can remove or isolate the cache.

SmolLM2 135M is a small English-chat CPU model, not a financial reasoning expert. GPU models require compatible hardware and sufficient memory; availability is not a claim that every model has been tested on your device.

Review each model's original license before commercial use. The Qwen2.5 3B browser choice uses a research-only license. The default Qwen2.5 1.5B and SmolLM2 1.7B server sources use Apache 2.0. Offline inference does not make live prices, balances, approvals or gas available offline.

View Ethereum gas estimates and a token watchlist. Gas results distinguish base fees and transaction-cost estimates; watchlist entries show available prices, changes and charts.

1. Add a symbol or Ethereum token address to the watchlist.
2. Choose device or encrypted Vault storage for your saved list.
3. Use the token-check action to inspect a watched contract.

Prices and network fees change, providers can be delayed, and an estimate is not a promised execution price. Always check units such as USD, ETH and gwei. Do not treat a model's rewritten number as more authoritative than the source receipt.

Open Market and gas in Bansun →

Add up to ten public addresses or ENS names with optional labels. View supported balances, allocation, value and locally saved snapshots. Add extra token contracts if an asset is not on the built-in list.

Device storage is local but unencrypted. Vault storage is encrypted and requires an unlocked vault. The saved labels stay on your device; addresses selected for a balance or ENS lookup are sent to the backend and its data providers. This is watch-only monitoring, not custody or automatic trading. Provider coverage can omit assets.

Open Watch-only portfolio in Bansun →

Record open and closed trades with entry, stop, size, fees, targets and notes. Review available performance measures, including expectancy and drawdown.

• The position-size calculator uses your account size, risk percentage, entry, stop and fees. Check every input before relying on its output.
• “Log as open trade” copies a plan into the journal; it does not place an order.
• Edit or delete a trade after reviewing the confirmation.
• Export CSV for a portable copy, or import compatible CSV. CSV exports are readable, not encrypted.
• Choose Vault storage if you want encrypted records.

Slippage, liquidity, liquidation, gaps and execution fees can make real losses larger than a planned stop. Bansun is not a brokerage or investment adviser.

Open Trade journal and position sizing in Bansun →

Decode transaction calldata or an EIP-712 signature request before authorizing it in another wallet. Enter the destination, ETH value and data, or paste the typed-data JSON.

• Inspect approvals, Permit / Permit2 requests, spenders, recipients, deadlines and unlimited allowances.
• Check whether a signature or listing transfers rights you did not intend to grant.
• Look up a past transaction hash online to inspect its mined status and transfers.
• Local decoding works offline; live contract or transaction checks need the backend.

Warnings are explanations, not a formal audit or a complete simulation. A green result does not prove that an action is safe. Decoding does not sign or broadcast the transaction.

Open Before you sign in Bansun →

Enter an Ethereum token address or ENS target. Bansun checks available contract behavior, ownership, proxy or upgrade controls, liquidity and market information, then explains the observed risk signals.

Possible signals include minting, blacklisting, pausing, changeable taxes, transfer restrictions and administrator powers. Public-provider coverage is incomplete, bytecode patterns are not a full audit, and a low score means fewer detected signals—not “safe to buy.” Verify the exact contract address with independent sources. Never use a ticker alone to identify a token.

Open Token risk checks in Bansun →

Enter a public wallet address or ENS name to inspect supported token and operator approvals. Deep scanning also looks at recent events and may take longer on public RPC services. Results can flag unlimited allowances and unknown spenders.

Optional on-chain action: clicking Revoke first asks for confirmation, then requests a connection to your installed browser wallet. That wallet must use the scanned account on Ethereum mainnet. It presents the transaction for you to approve; gas fees apply. Bansun does not receive your private key. If no browser wallet is available, copy the revoke transaction data and review it in your own wallet.

Re-scan after confirmation. A pending transaction is not a completed revoke, and revoking one allowance does not remove unrelated approvals or undo a compromised key.

Open Approval audits and revocation in Bansun →

Check a message, email, URL or supposed support request for seed-phrase requests, fake claims, impersonation, urgency, recovery scams and suspicious domains.

The initial rules-based analysis runs on your device. An optional AI explanation sends the submitted text to the configured server AI. Remove personal details first. A low score does not verify a sender or prove a link is safe; do not open suspicious links just to test them.

Open Scam-message checks in Bansun →

Paste a whitepaper section, exchange rule, staking condition or other text and choose everyday language, steps, a table or a five-point summary. This uses server AI, so the pasted text is sent to the backend.

Check important figures, dates, fees, lock-ups and slashing conditions against the original document. Summaries can omit qualifications. Do not paste confidential contracts, credentials or sensitive personal records unless you are authorized and accept the processing involved.

Open Simplify documents in Bansun →

Write a specific, checkable prediction with a probability, deadline and source. After the outcome is known, resolve it as happened or did not happen. Brier scores compare the accuracy of recorded probabilities; lower is better.

A score evaluates past entries you resolved, not future performance. The ledger is local; records are not automatically published or resolved from an oracle. Avoid ambiguous claims and keep supporting evidence. Predictions from AI answers can be logged through the app's prediction action.

Open Prediction ledger in Bansun →

Create a vault with a strong passphrase, unlock it to read or edit notes, and lock it when you are finished. Vault records use AES-256-GCM encryption; the passphrase protects the key through PBKDF2-SHA256.

• Use a long, unique passphrase. Bansun cannot reset or recover it for you.
• Choose Vault storage separately in tools that support it. Ordinary device storage is not automatically encrypted.
• Download an encrypted backup and keep it separately from your passphrase and recovery shares.
• Importing a backup can replace the vault on this device; read the confirmation first.
• Deleting a note or clearing local data cannot remove copies you exported or shared.

Encryption does not protect an unlocked device, a malicious browser extension, malware or secrets copied into another service. Storage belongs to the browser and origin, not a cloud account.

Open Encrypted vault and backups in Bansun →

Create multiple recovery shares and choose how many must be combined—for example, three out of five. Give each share to a different trusted person and keep the encrypted vault backup.

To recover, import the matching encrypted backup if necessary, collect enough matching shares, and set a new passphrase. A share is sensitive; anyone with the required threshold and the vault backup can recover access. Test your recovery process before relying on it. Bansun does not hold your shares or provide an account-recovery desk.

Open Threshold recovery in Bansun →

The device creates and blinds credit tokens; a compatible server signs the blinded values. Redeeming a valid, unused token proves entitlement without revealing the token during issuance. This does not hide your network connection or guarantee complete anonymity.

Credit enforcement is currently disabled in this app. The credit interface is not a checkout, token purchase or paid subscription. Server keys and credit records use filesystem storage and are not guaranteed to survive a redeploy. Paid enforcement should not be enabled until durable storage and backup procedures are configured.

Open Anonymous credits in Bansun →

Review browser permissions, storage information, Global Privacy Control and a device-safety checklist. These checks run locally; a result is not a security audit of the operating system.

Request persistent browser storage to reduce automatic eviction, although the browser decides whether to grant it. “Erase all Bansun data on this device” is destructive: export anything you want to keep before confirming. Erasure does not remove exported files, records on other devices or data already sent to third-party services. Read the Privacy Policy for the complete processing overview.

Open Privacy check and local deletion in Bansun →

$BANSUN is not launched. There is no official token contract, presale, airdrop claim or staking service in this release.

The token page describes a plan, draft allocation, possible utility and risks—not an available investment product. Do not purchase or connect a wallet to a claim based on the plan. Verify updates through the official Bansun X account. Any launch would require separate published terms and verified contract details.

Open $BANSUN token status in Bansun →

Device storage, encrypted vault storage, the browser AI cache and server state are different things. An encrypted vault does not encrypt every local preference or CSV export.

• On device: saved lists, journal entries, ledger records, preferences, vault ciphertext and browser model caches. Available export formats vary by tool.
• In transit: server AI receives the messages you submit; live chain and price requests send the addresses or tokens needed for the lookup.
• On server: credit signing keys, spent-credit records and operational state. These are not a cloud backup of your private records.

A new domain, browser profile or device has a separate storage area. Export supported records before moving from a preview or Replit address to bansun.xyz. There is no automatic synchronization. Do not clear a device until you have verified the backup you need.

Requests use the same app origin and JSON bodies. Never put credentials, seed phrases or private keys in a request. Responses use endpoint-specific fields; errors include a code and a readable message. Invalid input, rate limiting and AI-capacity limits can produce 400, 429 or 503 responses.

Read the detailed API payload reference. Use the live health endpoint rather than assuming a model is ready because its files exist. A warming, downloading or busy runtime is not ready for a new request.

The native Node.js backend manages a loopback-only Ollama process. It downloads and verifies the primary model with real inference before reporting readiness. Additional configured models load on demand rather than repeatedly warming in the background. The service binds to its configured PORT.

Defaults adapt to the detected memory limit: below 3 GB, Qwen2.5 0.5B; at 3 GB or more, Qwen2.5 1.5B and SmolLM2 1.7B. Below 8 GB, only one model can be resident; larger servers allow two. Models have a five-minute idle lifetime. Explicit LLM_MODELS configuration takes precedence. Smaller models are less capable and do not replace verified financial data. A server too small for Qwen2.5 0.5B needs more resources rather than an unusable automatic fallback.

/api/health and /api/models report the memory profile, actual inference-verification flags and safe failure codes. Installed files and a reachable model engine alone do not prove inference works. Failed primary startup is retried every five minutes without restarting the web app. A memory profile cannot guarantee sufficient free memory or compatible hardware.

On Replit, use a Reserved VM; four CPU cores and 8 GB RAM remain recommended for comfortable use of the larger models. Browser AI uses device resources instead. Server AI always needs internet from the browser; installing Ollama on the server does not make phone chat work offline. VM availability does not guarantee filesystem durability across redeploys.

Replit readiness notes · Architecture reference. Keep credit enforcement disabled until persistent storage is configured. Workspace changes do not update the public site until the changed version is published.

Bansun does not guarantee the accuracy of AI, prices, contract analysis, provider coverage or forecasts. It is not investment, tax or legal advice, and it is not a substitute for a wallet's own transaction review or a professional contract audit.

Read the Terms, Privacy Policy, Security guidance and Licenses. For troubleshooting, see Support. The official website is bansun.xyz and the official X profile is @bansun_ai. Do not trust look-alike support accounts or anyone asking for a recovery share, seed phrase or passphrase.