volcano-agent-sdk MCP — 使用場景, 安裝 & 即時演示

為什麼要用

核心特性

100+ LLM models across major providers
Automatic tool selection from multiple MCP endpoints
Multi-agent crews with task delegation
Parallel / branching / loop execution patterns
OpenTelemetry traces, native OAuth for MCPs, connection pooling

即時演示

實際使用效果

volcano-agent-sdk.replay ▶ 就緒

0/0

安裝

選擇你的客戶端

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "volcano-agent-sdk": {
      "command": "npx",
      "args": [
        "-y",
        "volcano-agent-sdk"
      ],
      "_inferred": true
    }
  }
}

開啟 Claude Desktop → Settings → Developer → Edit Config。儲存後重啟應用。

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "volcano-agent-sdk": {
      "command": "npx",
      "args": [
        "-y",
        "volcano-agent-sdk"
      ],
      "_inferred": true
    }
  }
}

Cursor 使用與 Claude Desktop 相同的 mcpServers 格式。專案級設定優先於全域。

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "volcano-agent-sdk": {
      "command": "npx",
      "args": [
        "-y",
        "volcano-agent-sdk"
      ],
      "_inferred": true
    }
  }
}

點擊 Cline 側欄中的 MCP Servers 圖示，然後選 "Edit Configuration"。

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "volcano-agent-sdk": {
      "command": "npx",
      "args": [
        "-y",
        "volcano-agent-sdk"
      ],
      "_inferred": true
    }
  }
}

格式與 Claude Desktop 相同。重啟 Windsurf 生效。

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "volcano-agent-sdk",
      "command": "npx",
      "args": [
        "-y",
        "volcano-agent-sdk"
      ]
    }
  ]
}

Continue 使用伺服器物件陣列，而非映射。

~/.config/zed/settings.json

{
  "context_servers": {
    "volcano-agent-sdk": {
      "command": {
        "path": "npx",
        "args": [
          "-y",
          "volcano-agent-sdk"
        ]
      }
    }
  }
}

加入 context_servers。Zed 儲存後熱重載。

claude mcp add volcano-agent-sdk -- npx -y volcano-agent-sdk

一行命令搞定。用 claude mcp list 驗證，claude mcp remove 移除。

使用場景

實戰用法： volcano-agent-sdk

Build a coding agent that uses GitHub + Sentry MCPs

👤 TS devs building internal automation ⏱ ~60 min advanced

何時使用： You want a programmable agent, not a chat session.

前置條件

Node 20+ — Standard
MCP endpoints for the tools you want (github, sentry, etc.) — Either existing public or your own deploys

步驟

Install + scaffold

npm install @volcano.dev/agent and write a minimal agent that connects to github + sentry MCPs with Anthropic as the model.✓ 已複製

→ Running TS project
Write the task

The agent's task: every 15 min, find new Sentry errors, correlate to GitHub commits via MCP, draft revert PRs for the obvious ones.✓ 已複製

→ Agent executes task autonomously
Instrument

Enable OTel traces and pipe to Honeycomb/Grafana.✓ 已複製

→ Spans visible

結果： A production-ready automation agent with observability.

注意事項

Agent hallucinates tool calls that don't exist — Restrict the MCP set passed to the agent; fewer, well-documented tools > more
Retries amplify transient upstream issues — Tune retry policy and add exponential backoff

搭配使用： github · sentry

Compose a multi-agent crew for research tasks

👤 Developers exploring agent patterns ⏱ ~45 min advanced

何時使用： A task benefits from specialization (researcher + writer + reviewer).

步驟

Define agents

Create agents: Researcher (web search MCP), Writer (drafts), Reviewer (fact-checks Researcher's sources).✓ 已複製

→ Three typed agent instances
Delegate

Run the crew: topic 'state of MCP in 2026'. Have Researcher gather, Writer draft, Reviewer verify claims.✓ 已複製

→ Coordinated output

結果： Higher-quality output than a single-pass agent on complex tasks.

搭配使用： omnisearch

Build a streaming chatbot with tool access

👤 Product devs integrating AI into a TS app ⏱ ~40 min advanced

何時使用： User-facing feature that needs real-time streaming + MCP tool calls.

步驟

Wire streaming

Build a chat endpoint that streams tokens to the client and calls tools mid-stream as the model requests.✓ 已複製

→ Working streaming endpoint
Explainability

After each response, expose agent.summary() so the UI can show which tools were used.✓ 已複製

→ Tool trail visible

結果： A production chat UI with transparent tool use.

組合

與其他 MCP 搭配，撬動十倍槓桿

volcano-agent-sdk + github + sentry

Autonomous triage agent across incident + code

Build an agent that, given a Sentry alert, fetches the stack, finds the offending commit via GitHub, and opens a PR with a minimal fix.✓ 已複製

volcano-agent-sdk + vurb-ts

Vurb builds the server side; Volcano builds the agent side

Expose my business data via a Vurb MCP; build a Volcano agent that uses it to answer user questions.✓ 已複製

工具

此 MCP 暴露的能力

工具	輸入參數	何時呼叫	成本
(SDK)	You write TS; SDK picks tools from configured MCPs automatically	N/A — Volcano Agent SDK is a library you build with, not an MCP server you call	n/a — depends on model + tools

成本與限制

運行它的成本

API 配額: LLM provider limits apply; MCP upstream limits apply
每次呼叫 Token 數: Depends on model + conversation length
費用: SDK free; LLM usage billed by provider
提示: Use cheaper models (Haiku/GPT-4o-mini) for routing and reserve expensive models for reasoning; Volcano supports per-step model choice.

安全

權限、密鑰、影響範圍

憑證儲存： LLM provider keys + MCP credentials in env; SDK injects them

資料出站： LLM provider APIs + configured MCP endpoints

切勿授予： LLM keys to untrusted code paths in the same process

Auto tool selection means the agent can call anything you give it — scope MCP surface narrowly.
OTel traces may include prompt/response content — review before exporting to third-party APM.

故障排查

常見錯誤與修復

MCP connection fails on startup

Verify the MCP endpoint URL and auth. SDK logs full error when --debug is set.

驗證： Curl the MCP endpoint directly

Model refuses to use available tools

Tool descriptions may be poorly phrased; rewrite for clarity or force via agent config.

驗證： Inspect tools via agent.listTools()

High token cost on simple tasks

Check that system prompt isn't dragging MCP tool defs into every call; use lazy tool-load mode.

驗證： agent.summary() shows token breakdown

替代方案

volcano-agent-sdk 對比其他方案

替代方案	何時用它替代	權衡
LangChain / LangGraph (TS)	You want the largest ecosystem of integrations	Heavier abstraction; slower cold path
Vercel AI SDK	You want tight Next.js integration	Less focus on multi-agent patterns
Anthropic SDK raw	You only need Anthropic and minimal abstraction	You reimplement tool routing, retries, multi-provider

volcano-agent-sdk