@bvcc/agent-mcp

@bvcc/agent-mcp

Model Context Protocol server for a BVCC Agent Wallet. It lets MCP-speaking AI runtimes — Claude Code, Cursor, the Claude desktop app — operate a BVCC Agent Wallet on-chain: check balances and limits, plan and simulate swaps, send tokens, swap on Uniswap v3/v4 (including to and from native ETH), provide liquidity on Uniswap v3/v4 (including native-ETH pools), and lend, borrow and unwind positions on Aave v3 — including closing or deleveraging a position in safe steps on its own. It also ships built-in operating guides (listGuides / getGuide, also exposed as prompts) so an agent can learn how to use each area on demand.

Every tool is generated from the @bvcc/agent-sdk capability catalog. There is no per-tool code here: add a capability to the SDK catalog and it appears here automatically.

New here? Start with QUICKSTART.md — the full end-to-end setup (create wallet → authorize the agent on-chain → fund it with gas → configure → verify). It covers the two steps people miss without which the agent does nothing.

Security

This server adds no powers. All limits — spend caps (native + per-token, daily + total), allowed tokens, allowed protocols, recipient whitelist, and a global pause — are enforced on-chain by the Agent Wallet contract. The worst any tool can do is bounded by what you authorized for the agent in the BVCC dashboard.

• The agent's private key is read from the environment, used locally to sign, and never transmitted. BVCC does not receive, store, or custody it.
• Tools are exposed explicitly via the SDK catalog — nothing is auto-discovered.
• Set BVCC_MCP_READONLY=true to expose only read/simulate tools (no writes).

Tools

Generated from the catalog and tagged by class:

Core (wallet, transfers, swaps):

Class	Tools
🟢 read	getAgentStatus, getCapabilities, getNativeBalance, getTokenBalances, getRemaining, needsApproval
🟡 simulate	buildSwapPlan, dryRunSendNative, dryRunSendToken, dryRunSwapV3, dryRunSwapV4
🔴 write	sendNative, sendToken, approve, swapV3, swapV4, swapToNative, swapFromNative

A 0.15% BVCC agent fee is charged automatically on-chain on agent actions, separate from gas and from your spend budget.

Aave v3 lending (aave module — Arbitrum, Ethereum, BNB, Base, Polygon):

Class	Tools
🟢 read	getAavePositions, getAaveMarket
🟡 simulate	dryRunAaveSupply, dryRunAaveBorrow, aavePlanClosePosition, aavePlanDeleverage, aavePlanCollateralSwap, aavePlanDebtSwap
🔴 write	aaveSupply, aaveWithdraw, aaveBorrow, aaveRepay, aaveRepayWithATokens, aaveSetCollateral, aaveSetEMode, aaveClosePosition, aaveDeleverage, aaveCollateralSwap, aaveDebtSwap

Deposits and borrows are always for the wallet itself — no to or onBehalfOf parameter exists, because the wallet's on-chain call policies pin the beneficiary.

⚠️ The four aave* unwinding tools can send several transactions in one call. aaveClosePosition, aaveDeleverage, aaveCollateralSwap and aaveDebtSwap run a chunked loop, re-deriving and re-simulating each step from live state. Call the matching aavePlan* tool first to see what it intends to do. They stop rather than proceed if the health factor would fall below hfFloor (default 1.05), if a swap prices too far from the Aave oracle, if the loop stops making progress, or if the agent's budget would run out.

Uniswap liquidity (lp module — open, collect, close positions on v3 & v4):

Class	Tools
🟢 read	getV3Position, getV4Position
🟡 simulate	dryRunAddLiquidityV3, dryRunRemoveLiquidityV3, dryRunAddLiquidityV4, dryRunRemoveLiquidityV4
🔴 write	addLiquidityV3, removeLiquidityV3, collectFeesV3, burnV3, addLiquidityV4, removeLiquidityV4, collectFeesV4, burnV4

The position NFT is minted to the wallet and every proceed returns to it — pinned on-chain. addLiquidityV3 sizes both sides to the pool's current price (full-range by default, or pass ticks). v4 covers native-ETH pools; its tools need the v4 PositionManager's DEEP validator active on-chain for the agent, else the action fails closed. Note there is no increaseLiquidity on v3 (not owner-gated on the NFPM — add by minting a fresh position).

Guides (meta module — always exposed, help):

Class	Tools
🟢 read	listGuides, getGuide

listGuides names the how-to guides; getGuide({ area }) returns the playbook for one area — the recommended workflow and the gotchas, naming the exact tools — covering getting-started, swaps, lending and liquidity. They stay available even under a module filter, so an agent restricted to one feature can still learn how to use it. The same guides are also exposed as prompts (guide-<area>) so a human can pull one as a slash-command.

Writes carry the MCP destructiveHint annotation so clients can require confirmation. See GUIDE.md for the recommended operating workflow.

Configuration

Recommended: keep the values — above all AGENT_PRIVATE_KEY — in a dedicated env file and point the server at it with BVCC_ENV_FILE, instead of inlining the key in your MCP host's config (which gets shared, synced and screenshotted). chmod 600 that file and keep it outside any cloud-synced folder. You can still inline the variables in the host's env block if you prefer; host env wins over the file. See .env.example.

Example agent.env (path passed via BVCC_ENV_FILE):

AGENT_PRIVATE_KEY=0xYOUR_AGENT_KEY
WALLET_ADDRESS=0xYOUR_WALLET
CHAIN_ID=42161

Variable	Required	Description
AGENT_PRIVATE_KEY	yes	Agent EOA private key (0x + 64 hex). Used locally only.
WALLET_ADDRESS	yes	The BVCC Agent Wallet this agent operates.
CHAIN_ID	yes	Default chain: 42161 Arbitrum One · 56 BNB · 1 Ethereum · 8453 Base · 137 Polygon · 421614 Arbitrum Sepolia.
RPC_URL	no	Custom RPC for the default chain (otherwise a public default).
RPC_URL_<chainId>	no	Per-chain RPC override, e.g. RPC_URL_56.
BVCC_ENV_FILE	no	Path to a dedicated env file to load (keeps the key out of the host config). Host env wins over it.
BVCC_MCP_READONLY	no	true exposes only read/simulate tools.
BVCC_MCP_MODULES	no	Comma-separated feature groups to expose: core, aave, lp. Unset = all.

Narrowing what the agent can reach. BVCC_MCP_READONLY and BVCC_MCP_MODULES are independent and combine. If an agent will never touch lending, leaving those tools out is one less thing it can get wrong:

BVCC_MCP_MODULES=core              # wallet + swaps    (18 + 2 guides = 20)
BVCC_MCP_MODULES=aave              # Aave lending      (19 + 2 = 21)
BVCC_MCP_MODULES=lp                # Uniswap liquidity (14 + 2 = 16)
BVCC_MCP_MODULES=core,aave,lp      # everything (= unset)  (51 + 2 = 53)
BVCC_MCP_MODULES=core
BVCC_MCP_READONLY=true             # read-only wallet  (11 + 2 = 13)

The two guide tools (listGuides, getGuide, module meta) are always exposed on top of these — that is the + 2 — so an agent restricted to one feature can still read how to use it.

An unrecognised module name is ignored, so a typo yields a smaller surface, never a larger one. Neither switch is the security boundary — the contract is — but a smaller surface is fewer ways for a confused agent to act.

Multi-network: one server operates the agent on any supported chain. Every tool takes an optional network (chain id or name: ethereum, bsc, arbitrum, base, polygon, arbitrum-sepolia), defaulting to CHAIN_ID — so you can say "swap on bsc" without restarting. The wallet address is the same on every chain (CREATE2); the agent must be authorized on each chain you use.

Install & build

npm install
npm run build
npm test          # builds + stdio smoke test (no chain calls)

Connect to a client

Claude Code

claude mcp add bvcc-agent-wallet \
  --env BVCC_ENV_FILE=/secure/agent.env \
  -- npx -y @bvcc/agent-mcp

Cursor / Claude app (mcp.json)

{
  "mcpServers": {
    "bvcc-agent-wallet": {
      "command": "npx",
      "args": ["-y", "@bvcc/agent-mcp"],
      "env": { "BVCC_ENV_FILE": "/secure/agent.env" }
    }
  }
}

The key lives in agent.env, not in the config above. If you'd rather inline it, replace the env block with AGENT_PRIVATE_KEY / WALLET_ADDRESS / CHAIN_ID directly (less safe — the key sits in the host config). Pin a version for reproducibility, e.g. @bvcc/agent-mcp@0.2.1 (see Upgrading).

Upgrading

The SDK is bundled into this package, so updating the MCP is all you need to get new capabilities (Aave lending shipped this way in 0.2.0) — you never install or update @bvcc/agent-sdk separately.

• Pin the version in your config for reproducibility:

  "args": ["-y", "@bvcc/agent-mcp@0.2.1"]
  

  npx caches, so an unpinned @bvcc/agent-mcp can keep running an old build. To upgrade, bump the number (e.g. @0.2.1) — or use @latest if you prefer. Global installs: npm i -g @bvcc/agent-mcp@latest.
• Restart your MCP client. New tools from the catalog appear automatically; nothing else in the config changes. The startup banner prints the running version ([bvcc-agent-mcp vX.Y.Z]).
• ⚠️ Authorize any new protocol on-chain. A release that adds a new protocol (e.g. Aave) exposes its tools immediately, but the agent must have that protocol's contract in its allowedProtocols — otherwise the action reverts with ProtocolNotAllowed. Authorize it from the dashboard, same as a router.

See CHANGELOG.md for what each version changes. Versioning follows SemVer: patch = fix, minor = new compatible feature, and 0.x means the API may still change.

How it works

@bvcc/agent-sdk  ──catalog──►  @bvcc/agent-mcp  ──MCP──►  Claude Code / Cursor / Claude
   (on-chain limits live in the Agent Wallet contract, not here)

The server loads the catalog, registers one MCP tool per capability (Zod schema → tool input schema, kind → tool annotations), and routes each call to the SDK, which signs with the agent key and submits executeAsAgent.

License

MIT © BlockVenture Chain Capital (BVCC)