ChainAware Behavioral Prediction MCP
Plug predictive Web3 risk intelligence into your AI agent — flag fraud and rug pulls before they happen, score wallets, and rank token communities.
Positive: SKILL.md explicitly declares key sourcing from env var only, no logging, external endpoint data_sent/data_NOT_sent disclosure (no off-chain PII, no private keys), privacy policy link, MIT license. Deducted: total dependence on a single closed-source third-party backend, retention governed only by an external policy, no isolation/rollback/explicit-confirmation mechanisms, unverified publisher; static review cannot confirm the data-flow claims are true.
Positive: tool schemas, batch pipeline (job_id+signature, status machine, expires_at) and input validation checkpoints are internally consistent with documented failure states. Deducted: documentation-only skill with no tests or executable key-path reproduction evidence; failure feedback depends on a remote service; static cap of 10 applies, scored 8.
Positive: exhaustive trigger lists, an explicit 'When NOT to Use' section with alternatives, and clear per-tool network coverage tables. Deducted: no Chinese-language support evidence; core function depends entirely on overseas services (chainaware.ai endpoints) with no mainland-China reachability disclosure; some trigger phrases are overly broad, risking false invocation.
Positive: well-layered docs (SKILL.md + references + CLAUDE.md), version, license, stable naming, parameter tables, validation checkpoints. Deducted: heavy marketing content (awards, endorsement links) interleaved with functional documentation, no changelog, maintenance ownership/update path only implicit, and consistency between SKILL.md and the 34 subagents unverified (SKILL.md covers only a subset of tools).
Positive: complete workflows, risk threshold tables and output field docs that could yield directly usable verdicts. Deducted: accuracy claims (98%/90.1%) rest on self-published backtest links unverifiable statically; actual value depends entirely on a paid remote API; static cap of 7 applies, scored 5.
Positive: external references (CB Insights, BNB Chain, privacy policy, backtest pages) and structured self-described schemas; primary material auditable. Deducted: all key claims (accuracy, 14M+ wallets, 1.3B data points) are publisher self-assertions with no independent reproduction, no committed test suite or verified CI covering the skill's key paths; static cap of 5 applies, scored 4.
- The skill depends entirely on ChainAware's closed-source paid remote API (x402 supported); accuracy claims (98%/90.1%) are publisher self-assertions and unverified;
- Data retention is governed only by an external privacy policy — assess compliance before sending wallet addresses to the third-party service;
- Core endpoints are overseas with no mainland-China reachability disclosure and no Chinese-language support; Chinese users should test endpoint connectivity first;
- This is a static documentation review (confidence: low); no tool calls were executed and output quality/accuracy have not been independently reproduced.
What does this skill do, and when should you use it?
ChainAware Behavioral Prediction MCP is a remote MCP server exposing a behavioral intelligence layer built from 14M+ wallet profiles across 8 blockchains and 1.3B+ predictive data points. It offers fourteen tools: fraud detection (claimed ~98% backtesting accuracy on ETH), rug pull prediction (claimed 90.1%), batch screening, behavioral profiling, a 1–9 crypto credit score, token holder-community ranking, a multi-module token audit, and ERC-8004 AI-agent trust scores. Unlike tools that describe past transactions, its pitch is predicting what is about to happen. It requires a CHAINAWARE_API_KEY (subscription) or an x402-compatible client for pay-per-call access.
Connects over SSE to https://prediction.mcp.chainaware.ai/sse and calls: predictive_fraud / predictive_fraud_batch (with AML checks and forensic indicators for laundering, mixers, sanctions); predictive_behaviour / predictive_behaviour_batch (intent probabilities, experience score, behavioral categories, protocol usage, recommendations); predictive_rug_pull (contract/LP rug probability and risk indicators); credit_score (1–9 rating, ETH only); token_rank_list / token_rank_single (rank tokens by holder quality on ETH/BNB/BASE/SOLANA); run_token_audit / get_token_audit_result (async 8-module contract audit with 0–100 aggregate score); agents_trust_score_list / agents_trust_score_single (0–1000 ERC-8004 agent trust scores); plus check_job_status / get_job_results for batch job management. Chain support varies per tool — see the compatibility table.
- DeFi lending platforms use credit_score and fraud scores to set collateral ratios and interest rates per borrower (ETH only)
- Airdrop teams batch-screen address lists and filter Sybil/bot wallets via batch fraud or behavioural tools
- Launch reviewers check new pools and contracts for rug risk with predictive_rug_pull and the token audit pipeline
- AI-agent developers integrate ERC-8004 agent trust scores to decide whether an agent wallet can be trusted
- Compliance teams AML/sanction-screen user addresses before onboarding
- Marketing teams segment wallets by behavior and personalize DeFi product recommendations
What are this skill's strengths and limitations?
- Fourteen tools cover the full risk workflow: fraud, behaviour, rug pull, credit, token ranking, contract audit, agent trust
- Claims backtesting-verified accuracy (~98% fraud on ETH, 90.1% rug pull) with public verification pages
- Async batch jobs support screening large address lists
- x402 pay-per-call support means no subscription required
- Clear privacy boundary: only wallet addresses and network are sent — no names, emails, keys, or raw transactions
- Third-party recognition from CB Insights, BNB Chain, a Google Cloud grant, and AWS Fintech Accelerator
- Fully dependent on a remote API — no offline use; accuracy figures are self-reported backtest results, not independently audited
- credit_score is ETH-only; behaviour analysis covers only ETH/BNB/BASE/HAQQ/SOLANA — inconsistent chain support across tools is error-prone
- Claude Web/ChatGPT integrations require the API key in the URL, an inherent leak surface
- SKILL.md references ~35 .claude/agents/ subagents — Claude Code-specific assets that other platforms cannot use
- Paid subscription or x402 payment required; no free tier documented; no local test suite provided
- Wallet addresses may constitute personal data in some jurisdictions (GDPR/MiCA); operators must do their own data protection assessment
How do you install this skill?
No local install needed — the server runs remotely at https://prediction.mcp.chainaware.ai/sse. 1) Get an API key at https://chainaware.ai/pricing and set it as the CHAINAWARE_API_KEY env var; 2) Claude Code: run claude mcp add --transport sse chainaware-behavioural-prediction-mcp-server https://prediction.mcp.chainaware.ai/sse --header "X-API-Key: your-key-here"; 3) Cursor: add the URL and X-API-Key header in mcp.; 4) Claude Web/Desktop or ChatGPT Connectors: add the URL with ?apiKey= in integration settings — these platforms don't support custom SSE headers, so the key appears in the browser network tab; use a restricted-scope key.
How do you use this skill?
Ask in natural language, e.g. "Is it safe to interact with 0xABC... on Ethereum?", "Will this new DeFi pool rug pull?", "What will this wallet do next?", "What are the top AI tokens on Ethereum?". Batch workflow: call predictive_fraud_batch or predictive_behaviour_batch to get job_id and signature (store both), poll check_job_status until status is completed or partial, then call get_job_results; never fetch while status is pending/processing. For full due diligence, combine fraud + behaviour + rug pull outputs into one verdict. Confirm the network before each call — tool chain support differs.
How does this skill compare with similar options?
SKILL.md positions itself as 'predictive' versus forensic blockchain tools that describe the past, and notes that block explorers suit balances/history while code-auditing tools suit bug hunting — i.e., complementary rather than a replacement.