V.A.P.E.

Virtual Ape Private Eye

An autonomous on-chain intelligence system for Base and Ethereum/EVM — live investigations, market intelligence, security auditing, and 32 priced offerings settled on-chain.

Live Dashboard · Architecture · Quick Start · x402 Worker · Deployment · @based_vape

Overview

V.A.P.E. is a fully autonomous AI detective for the on-chain ecosystem. It runs continuously on GitHub Actions at zero local compute cost, watching Base, Ethereum, and the wider EVM ecosystem, and publishes exactly what it finds — a token security scan, a market anomaly, a completed investigation — with the evidence attached, never just a conclusion.

It is a verified on-chain identity, not an anonymous script: ERC-8004 agent #59900 (wallet 0x8aAB9a6d28e9AbA2a15a613C90F24f352f0Cce15, basename vapex402.base.eth) — a passport VAPE registered itself, directly against Base's canonical IdentityRegistry contract, with no third-party platform brokering it. Every paid engagement settles in USDC on Base. x402 is VAPE's sole commerce rail — all 32 offerings are hireable instantly with a wallet-signed HTTP payment, no account or escrow step, settled directly against a hosted x402 facilitator on Base mainnet; deeper, async work (full contract audits, wallet forensics, whale-wallet tracing) dispatches a GitHub Actions job and delivers the report inline once it completes. See worker/README.md for the x402 side, or just use the site's Engagement Options section. (VAPE previously also offered an escrow-backed engagement via Virtuals Protocol's Agent Commerce Protocol — ACP — sunset 2026-07-31 as VAPE refocused on Base/Ethereum/EVM + x402; see docs/ACP_PROTOCOL.md for that prior integration.)

X: @based_vape · Live dashboard: juxtaposition1.github.io/V.A.P.E

What it actually does

• Deep investigations — every cycle, VAPE auto-selects the highest-signal live Base target, runs multi-source recon (token-safety and holder/liquidity data, on-chain contract verification, recent-hack technique correlation, and a real web search for public rug/scam mentions), scores it 0-100, and publishes a verdict — PROCEED, CAUTION, or REJECT — to intel/investigations/ and the live dashboard.
• Market intelligence — TVL, gas, Fear & Greed, global market cap, and Base's top protocols and trending pairs, refreshed continuously and shown live on the site.
• Security auditing — a real prompt-injection red-team suite runs against VAPE's own reporting pipeline, plus a static-analysis tier (Slither, Aderyn, Mythril) for smart-contract review offerings.
• Self-improvement — a Builder agent grounded in a shared Memory system proposes and implements real code changes (new tools, bug fixes, skill playbooks), every one gated behind automated security validation and a human-reviewed pull request.
• Commerce — 32 priced offerings (token safety, liquidity checks, rug-pull alerts, market intelligence, full contract audits, general web research, 15 market-data tools, and more), all x402-payable — hireable instantly with a signed wallet payment, no account needed. Results render inline on-site or download as a PDF.

The live site

juxtaposition1.github.io/V.A.P.E is the primary way to see and use VAPE — not a static status page, a working product surface:

• Live network/market data (Base TVL, gas, top protocols, trending pairs, Fear & Greed)
• A wallet-connect portfolio view (injected wallets, Coinbase Wallet, WalletConnect) with real Base holdings, balances, and price history — no fabricated numbers
• Instant x402 payment for any auto-fulfilled offering, signed directly from a connected wallet, with the report rendered inline and downloadable as a letterheaded PDF
• A live, paginated, searchable job ledger of every real x402 settlement VAPE has taken
• A fully linkable investigation archive, indexed automatically from every report VAPE has ever published — nothing curated after the fact
• A free token-security preview tool, open to anyone, no wallet required

Built with zero bundler (plain HTML/CSS/JS under docs/assets/) and a small Cloudflare/Deno Workers backend (worker/) that gates the paid offerings behind x402 and serves the site's live on-chain data panels.

Architecture

Three cooperating pieces:

Full component-level detail, the real data flow, and current status of every piece live in docs/ARCHITECTURE.md — this README stays a map, not a duplicate.

Quick start

git clone https://github.com/jUXTAPOSITION1/V.A.P.E.git
cd V.A.P.E
cp .env.example .env        # fill in at least one LLM provider key
pip install -r agents/requirements.txt

python -m agents.investigate --auto     # run one real investigation
python -m agents.run                    # hourly bounty-cycle pass
python -m agents.run --review-repo      # self-review pass
python -m agents.builder --task "..."   # generate code grounded in Memory

Everything above writes real output: a report in intel/ or reports/, a finding appended to skillforge/memory/findings.jsonl. See docs/DEPLOYMENT.md for running the full system on GitHub Actions (the actual production setup — 24/7, zero local compute) and worker/README.md for the x402 payment backend.

Memory, Builder, and MCP

• Central Memory (skillforge/memory/) — an append-only, searchable brain every component reads from and writes to: findings.jsonl (security discoveries), lessons.jsonl (build outcomes), skills.jsonl (learned playbooks), build_log.jsonl (instructional build patterns — see BUILD_LEDGER.md), and social-events.jsonl. See docs/MEMORY.md.
• Builder (agents/builder.py) — grounds every code-generation task in Memory, validates the output against a hard-block/soft-warn security policy, and auto-appends results back to Memory. See docs/BUILDER.md.
• MCP — VAPE speaks the standard Model Context Protocol both ways: a real MCP server (mcp_servers/vape_mcp.py) exposes VAPE's own tools to any MCP host, and VAPE itself hosts several registered MCP servers for research and tool discovery. See docs/MCP.md and docs/MCP_SERVER.md.

Security and audit

• Code-generation gate — Builder-generated code is checked against a hard-block list (eval, exec, shell execution, unsafe deserialization) before it can be appended or opened as a PR. See docs/SECURITY_PROTOCOL.md for the full threat model and CI regression guard.
• Human review gate — every autonomous PR (self-improvement, build requests, skill synthesis) requires human review before merge; nothing reaches main unreviewed.
• Adversarial testing — a real prompt-injection red-team suite (agents/redteam.py) runs against VAPE's own report-generation pipeline daily, scored PASS/FAIL from actual model output, not a hypothetical scenario.
• Append-only Memory — skillforge/memory/ is write-once; findings, lessons, and skill records are never silently deleted or rewritten.
• Payment-path integrity — x402 settlements verify against a live facilitator (VAPOR/CDP hybrid routing) before a job is billed; a failed deliverable never charges the caller (see worker/src/index.ts's non-2xx-skips-settlement contract).
• Dependency and secret hygiene — automated dependency auditing across agents/, worker/, and scripts/, plus a static security-lint regression guard (scripts/security_lint.py) run in CI on every PR.

Repository layout

V.A.P.E/
├── agents/              Python engine — investigations, market sweeps, builder, red-team
├── skillforge/          Self-growing skill/tool ecosystem + shared Memory
├── intel/               Real-data audit trail: reports, broadcasts, investigations, bounty radar
├── docs/                Site (docs/index.html) + architecture/deployment/protocol docs
├── worker/              Cloudflare/Deno Workers backend — x402 payments, portfolio data
├── mcp_servers/         Standard MCP server exposing VAPE's tools
├── scripts/acp-monitor/ ACP job listener/fulfillment daemons (sunset 2026-07-31, kept for reference)
└── reports/             Timestamped bounty-cycle and self-review output

Status

Live and running: Featured Investigations every 30 minutes, hourly bounty/market sweeps, the SKILLFORGE tool ecosystem (16 tools registered, 16 verified), x402 job fulfillment on Base mainnet across all 31 of VAPE's live offerings (21 auto-fulfilled with zero manual work end-to-end), wallet-connect portfolio, a live searchable job ledger, the Central Memory system, the standard MCP server, and the Builder self-improvement pipeline.

Sunset (2026-07-31): VAPE's ACP (Agent Commerce Protocol) escrow-engagement rail and its dedicated Virtuals Protocol tracking/reporting sweep — VAPE is refocusing on Base/all-EVM/Ethereum + x402 as its sole commerce rail rather than the Virtuals ecosystem specifically. The underlying code (scripts/acp-monitor/, agents/acp_fulfill.py, agents/virtuals_sweep.py) is left in the repo, just no longer scheduled/advertised; see docs/ACP_PROTOCOL.md for the prior integration.

In progress: a dashboard reputation loop, richer social-sentiment intel beyond the current keyless aggregate. app.py is an unconnected placeholder stub (it doesn't call any real VAPE agent code) that exists only to satisfy this repo's synced Hugging Face Space's Gradio entry point — the actual product is the live GitHub Pages site above.

Documentation

Environment variables

Copy .env.example to .env. At minimum, set one LLM key:

GROQ_API_KEY=       # free tier — enough to run everything

xAI's Grok 4.1 Fast is the primary model for reports/investigations/expert assessment when XAI_API_KEY_1 is set (see agents/llm.py's FRONTIER_ORDER), but every path has a free fallback — nothing in VAPE requires a paid key to run. Everything else (additional recon/market-data keys, LLM fallbacks, MCP research providers, the worker's optional data keys) is optional — VAPE degrades gracefully to keyless data sources when a key isn't set. See .env.example for the full list and docs/DEPLOYMENT.md for which secrets the live site's CI actually needs.

Contributing

1. Read docs/ARCHITECTURE.md and the relevant subsystem doc before changing it.
2. New code must pass the Builder's security validation (no eval/exec, no unrestricted shell/OS access) and include real error handling — no silent failures.
3. A data source that's unreachable should degrade honestly (report the gap), not fall back to an approximated or invented value.
4. Autonomous agents open pull requests for review; nothing is designed to write directly to main unreviewed.

License

MIT — see LICENSE.