Telegram

Audit from chat — no terminal, no install. @tryannealbot gives you a command bot and a Mini App, both backed by the same engine and on-chain registry.

Mini App

Tap the bot’s menu button (“Open TryAnneal”) to launch the Mini App — a mobile-native safety oracle. Pick an example or paste a code hash; read a live verdict straight from Mantle, inside Telegram.

Commands

Long audits send a “⏳ Auditing…” message first, then edit it with the result. Hard timeout 60s.

Multilingual reports

Append a language code to /audit — e.g. /audit 0x… zh — to get the verdict and findings translated. The audit always runs in English; the finished report is then translated by Tencent Hunyuan (its Hunyuan-MT model on Tencent Cloud TokenHub). Supported languages include zh, es, ja, ko, fr, pt, de, ru, it, ar, hi, vi, th, and tr. On the web /try page, language chips under each result translate it in one click.

The cascade behind it

Every /audit runs the full critic cascade by default (thorough, not a quick pre-screen-only pass): ChainGPT pre-screens, then two architecturally-distinct critics — Groq Llama-3.3-70B and OpenAI GPT-OSS-120B — run as independent Stage-2 critics that cross-validate each other (Gemini 2.5 Pro is an optional third critic, off by default), alongside Slither + Aderyn + 16 custom detectors and a 98-pattern / $7.1B corpus for static cross-validation. The cascade is resilient — a ChainGPT pre-screen failure is non-fatal and the critics still run. If nothing could analyze a contract (for example a single .sol file with unresolved imports that won’t compile and no model response), the verdict is flagged analysisIncomplete and is never reported as “safe” or “100/100” — it says it could not complete the audit.

AI audits are usually non-deterministic — TryAnneal’s verdict is reproducible. The same contract always returns the same verdict: every model decodes at temperature 0 (greedy, seeded), a corroboration rule requires any reported finding to have ≥2 independent sources (≥2 models, or a model + Slither, when the full panel runs) so single-model hunches don’t drive the verdict, scoring is confidence-weighted, and results are memoized by code hash (keccak/sha3 of the source) on the bot — identical source returns the identical audit.

On-chain attestation

Verdicts are posted on-chain to AnnealValidation as ERC-8004 agent #131 — idempotently, and for both verified-address and GitHub-source audits (for GitHub sources the codeHash = keccak(source)). The bot resolves an address across Mantle, Ethereum, Base, Arbitrum, Optimism, BNB, Polygon, and Avalanche using eth_getCode as the ground truth for where a contract is actually deployed (not the explorer’s response); verified source is then fetched via the Etherscan V2 multichain API.

Channel-native UX

• →No terminal, no wallet pop-ups to read a verdict — Web2 users get a real security answer in a chat.
• →The Mini App is the same web safety-oracle, themed for Telegram and launched in one tap.