360 Secure SkillHub
360 安全技能中心

Canopy

Canopy 为 AI 代理提供受策略约束的链上美元支付能力:经 MCP 工具发起支付或预览、探测并支付 x402/MPP 付费接口、发现付费服务、查询预算,并处理需人工确认的审批;结果分为 allowed、pending_approval、denied 三种。

atifazam
atifazam
热度 34.59
最近更新 2026-05-11 17:47:39
分类:自动化依赖:MCP 服务 / 第三方 API应用场景:财务 / 工作流

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add canopy
Metadata
name
canopy
description
Canopy treasury wallet for AI agents — make policy-gated USD payments on-chain (Base), auto-pay paywalled APIs (x402 / MPP), and discover paid services. Every payment passes through the agent's policy: spend cap, recipient/service allowlist, and approval threshold. Payments above the threshold return pending_approval and wait for a human; payments outside the cap or allowlist return denied. Install only if your agent needs to spend org funds. Connect with the least-privileged agent + policy available, verify recipient and amount before each canopy_pay, and never auto-approve on the user's behalf. Requires a Canopy account at https://www.trycanopy.ai. After install, configure the canopy MCP server with your CANOPY_API_KEY and CANOPY_AGENT_ID from the dashboard.
metadata
{"author":"canopy","version":"1.0.0","homepage":"https://www.trycanopy.ai","openclaw":{"requires":{"env":["CANOPY_API_KEY","CANOPY_AGENT_ID"]}}}

Canopy

Canopy is a treasury wallet + per-agent policy gate. An org funds a treasury once, attaches a policy to each agent (spend cap, allowlist, approval threshold), and gives the agent the canopy_* MCP tools below. Canopy enforces the policy on every payment.

Setup

After installing this skill, register the Canopy MCP server with OpenClaw. Replace the placeholders with your own values from the Canopy dashboard (Connect agent → Credentials):

openclaw mcp set canopy '{
  "url": "https://mcp.trycanopy.ai/mcp",
  "transport": "streamable-http",
  "headers": {
    "Authorization": "Bearer YOUR_CANOPY_API_KEY",
    "X-Canopy-Agent-Id": "YOUR_CANOPY_AGENT_ID"
  }
}'

Or edit openclaw.json directly:

{
  "mcpServers": {
    "canopy": {
      "url": "https://mcp.trycanopy.ai/mcp",
      "transport": "streamable-http",
      "headers": {
        "Authorization": "Bearer YOUR_CANOPY_API_KEY",
        "X-Canopy-Agent-Id": "YOUR_CANOPY_AGENT_ID"
      }
    }
  }
}

Restart OpenClaw. The canopy_* tools become available immediately.

Verify with canopy_ping — it returns the resolved agent + org and a latency number.

Triggers

Use Canopy when the user mentions:

  • "pay", "transfer", "send $X to "
  • "approve" / "deny" a pending payment
  • "budget", "spend cap", "remaining budget"
  • "paid API", "x402", "402 paywall"
  • "find a paid service for …", "discover services"
  • "agent wallet", "treasury"

Available tools

Tool Purpose
canopy_pay Send a USD payment to an address. Subject to policy. Returns one of three outcomes (see below).
canopy_preview Dry-run policy evaluation without signing or charging. Same shape as canopy_pay plus dryRun: true.
canopy_check_url Probe an x402 / MPP URL and return the parsed offer plus a policy verdict. Cached 60s per (org, url).
canopy_discover_services List paid services the agent can call. Filter by category and query.
canopy_get_approval_status Poll the current state of a pending approval.
canopy_wait_for_approval Block until an approval is decided, max 60s. Use polling for longer waits.
canopy_approve Mark a pending approval as approved. Use only when the user has explicitly said yes in chat.
canopy_deny Mark a pending approval as denied. Use only when the user has explicitly said no in chat.
canopy_get_budget Cap snapshot for the current rolling period (capUsd, spentUsd, remainingUsd, periodResetsAt).
canopy_ping Verify the API key + agent. Returns agent and org details plus latency.

Outcome model — canopy_pay, canopy_preview, canopy_check_url

These tools never throw on policy. They always return a structured value with one of three status values:

allowed

Payment went through (or would go through, for _preview / _check_url). Surface txHash (Base block-explorer hash) and costUsd to the user.

pending_approval

Payment exceeded the agent's approval threshold. Fields you should use:

  • approvalId — pass to canopy_approve / canopy_deny / canopy_wait_for_approval / canopy_get_approval_status.
  • recipientName, recipientAddress, amountUsd, agentName, expiresAt — render these to the user when asking.
  • chatApprovalEnabled — if false, calling canopy_approve / canopy_deny will fail. Tell the user to approve in the dashboard, optionally canopy_wait_for_approval to block.

If chatApprovalEnabled === true: ask the user "Approve $X to RecipientName?" then call canopy_approve(approval_id) or canopy_deny(approval_id) based on their reply.

denied

Policy blocked it (cap exceeded, recipient/service not on allowlist, etc.). Surface the reason to the user. Do NOT retry with smaller amounts to evade the cap.

HTTP and network failures still throw normally — wrap calls in try/except as you would any tool call.

Safety rules

  • Always show the user recipientName + amountUsd before calling canopy_approve. Approvals exist precisely so a human sees the transaction details before money moves. Echoing them back confirms intent.
  • Default to canopy_preview when the amount is dynamic or untrusted. Same shape as canopy_pay, no funds move. Use the verdict to decide whether to proceed.
  • Never auto-approve. canopy_approve and canopy_deny are for handling explicit user replies in chat, not for the LLM to bypass the human-in-the-loop.
  • Don't chunk a denied payment into smaller payments. If a denied outcome cites the spend cap, surface the cap to the user; suggest they raise it in the dashboard rather than working around it.
  • Don't reference txHash until status === "allowed". It's null in other outcomes.
  • Don't tight-loop canopy_get_approval_status. Use canopy_wait_for_approval (60s max) or accept that the dashboard owns the decision.

Quick examples

User: "Pay 50 cents to 0x4838B106FCe9647Bdf1E7877BF73cE8B0BAD5f97" → canopy_pay({ to: "0x4838B106FCe9647Bdf1E7877BF73cE8B0BAD5f97", amountUsd: 0.50 }). Surface txHash if allowed.

User: "How much do I have left to spend today?" → canopy_get_budget(). Render remainingUsd and periodResetsAt.

User: "Find paid orderbook data feeds" → canopy_discover_services({ category: "data", query: "orderbook" }). List services with policyAllowed: true first; mention any with policyAllowed: false only if the user asks why.

User: "Is paying https://feed.example/depth allowed?" → canopy_check_url({ url: "https://feed.example/depth" }). Return the verdict + amount + recipient name.

User (after a previous pending_approval): "Yes, approve it" → Confirm the recipient and amount you previously surfaced. Then canopy_approve({ approval_id: <approvalId> }).

Resources