V.A.P.E. MCP Server — standard Model Context Protocol

mcp_servers/vape_mcp.py

V.A.P.E. now speaks the industry-standard Model Context Protocol (MCP 2024-11-05), so any MCP host — Claude, Cursor, VS Code Copilot, or a custom agent — can discover and call V.A.P.E.'s real capabilities with no bespoke glue.

Why this way

• Zero new dependencies. MCP is JSON-RPC 2.0 over stdio; the server speaks it directly with pure stdlib. Nothing to pip install — keeps VAPE compute-free and Termux/Android-friendly (stdio is the local transport).
• Real code, real data. Every tool calls existing VAPE functions (agents.investigate, agents.token_scan, agents.data_fetchers, agents.defillama, the Memory retriever + SQLite index) directly in-process. The one exception is wallet_trace, which shells out to the real skillforge/tools/recon/wallet_trace.sh (Alchemy-backed, live-verified — see PR #145) rather than re-implementing it. No stubs, no fabricated numbers.
• Read-only / keyless-first. Nothing here signs, spends, or mutates chain state — safe to expose to any host.

Tools exposed (tools/list)

Resources exposed (resources/list)

Run it

# stdio server (what an MCP host launches)
python mcp_servers/vape_mcp.py

# in-process smoke test (no host needed)
python mcp_servers/vape_mcp.py --selftest

Register in an MCP host (stdio)

{
  "mcpServers": {
    "vape-detective": {
      "command": "python",
      "args": ["/abs/path/to/V.A.P.E/mcp_servers/vape_mcp.py"]
    }
  }
}

The host then sees every tool above and two resources — the LLM picks and calls them automatically.

VAPE as an MCP host (consuming the ecosystem)

Beyond serving its own tools, VAPE can now spawn and consume any MCP server — official reference servers and community search/scrape servers — via a pure-stdlib host client.

• skillforge/mcp_client.py — the host. Spawns a server over stdio, does the JSON-RPC handshake, lists/calls tools, tears it down. Servers are launched lazily per call (no daemons = no idle compute). PATH is auto-augmented so npx/uvx resolve even under cron/CI.
• mcp_servers/registry.json — declares every server: command, args, required env keys, keyless flag, and verified package source. Keyed servers activate the instant their env var is set — no code change.

python -m skillforge.mcp_client list                    # registry + live/keyed/needs-runtime
python -m skillforge.mcp_client tools git               # discover a server's tools
python -m skillforge.mcp_client call filesystem list_allowed_directories '{}'

Registered servers

The uvx runtime (uv) is a single static binary: curl -LsSf https://astral.sh/uv/install.sh | sh.

Unified research router

skillforge/research.py gives VAPE one search/scrape API that uses the best available provider and falls back to keyless:

• search: Tavily → Brave → keyless (SearXNG/DDG, best-effort).
• scrape: Firecrawl → Bright Data → Apify → keyless MCP fetch.

python -m skillforge.research providers                 # what's active right now
python -m skillforge.research search "base defi exploit bounty" --max 5
python -m skillforge.research scrape https://docs.base.org/

These are also exposed as MCP tools on VAPE's own server (research_search, research_scrape, mcp_servers), so any host or VAPE agent can call them. Add a provider key and the same call silently upgrades from keyless to Tavily/Firecrawl.

Note: keyless public search is unreliable from datacenter/CI IPs (they get blocked) — that is exactly why the keyed providers exist. Keyless scrape of known URLs works well via the MCP fetch server.

Paid, agent-to-agent MCP surface (remote, x402)

Added 2026-08-07: worker/src/mcpServer.ts exposes the same security suite and market-data catalog already sold at /scan/<name> and /data/<name> (see docs/ACP_PROTOCOL.md) as real MCP tools, reachable over Streamable HTTP at POST /mcp on the live Cloudflare Worker (https://vape-x402.vapex402.workers.dev/mcp) — no bespoke REST integration needed. Any MCP-native agent with an x402-capable client (e.g. @x402/mcp's x402MCPClient/createx402MCPClient) can tools/list for free and then tools/call a priced tool, paying per call the same way the HTTP routes already do — real USDC on Base or Solana, settled through the same CDP facilitator, PAY_TO_ADDRESS, and SOLANA_PAY_TO_ADDRESS as every other paid surface on this Worker. Every MCP tool is a thin wrapper around the exact same handler the equivalent HTTP route calls (handlers.ts::fulfill(), dataHandlers.ts::fulfillData()) — one implementation, two transports, prices/schemas always in sync.

Tools exposed: the 6 security offerings (token_safety_check, liquidity_check, rug_pull_alert, exploit_check, market_intel, dossier_check) and all 14 /data/<name> market-data offerings (see docs/ACP_PROTOCOL.md's tables for current prices — MCP pricing always matches). Not yet wired: the async/bespoke offerings that don't go through fulfill()/fulfillData() (bounty_deep_dive, deep_contract_audit, tx_decode, bulk_safety_bundle, community_intel_broadcast, website_review) — a natural follow-up, not started here.

This is a separate surface from the free stdio server above — that one stays exactly as-is (zero deps, free, local-first, the discovery/goodwill funnel), and requires no wallet to use. The remote MCP surface is the paid one, for agents (not humans in an IDE) that can actually sign and settle an x402 payment.

Relationship to the old MCP layer

skillforge/mcp.py (VAPE's original "Modular Connector Protocol" wrappers for GitHub/social/tool-registry) still runs as-is. This server is the standard MCP surface layered on top — it does not replace the harvest wrappers, it makes VAPE's investigation + memory capabilities interoperable with the wider ecosystem.

Queryable memory (companion)

skillforge/memory/index_db.py projects the append-only JSONL memory into a stdlib-SQLite index (with FTS5 full-text when available). The JSONL files remain the source of truth / audit trail; the DB is a derived, rebuildable projection so agents can ask real questions:

python -m skillforge.memory.index_db build
python -m skillforge.memory.index_db query "honeypot base" --category finding --days 30
python -m skillforge.memory.index_db stats

memory.db is not committed (gitignored) — it is rebuilt from JSONL each cycle by the bounty-cycle workflow, and on demand by the MCP memory_search tool.