mcp-server-code-execution-mode

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

MCP Server Code Execution Mode

MCP Server代码执行模式

Skill by ara.so — MCP Skills collection.

This MCP server implements Anthropic's "Code Execution with MCP" pattern, exposing a single

run_python

tool instead of hundreds of individual tools. The agent writes Python code to discover, call, and compose MCP tools dynamically, reducing context overhead from ~30,000 tokens to ~200 tokens.

Key capabilities:

Execute Python in rootless Podman/Docker containers
Proxy any stdio MCP server into the sandbox
Discover tools at runtime (no context preloading)
Fuzzy search across all connected servers
Persistent sessions (variables/state retained)
Security: no network, read-only filesystem, dropped capabilities

由ara.so提供的Skill — MCP Skills合集。

该MCP服务器实现了Anthropic的「基于MCP的代码执行」模式，仅暴露一个

run_python

工具而非数百个独立工具。Agent可编写Python代码来动态发现、调用和组合MCP工具，将上下文开销从约30,000 tokens降低至约200 tokens。

核心功能：

在无根Podman/Docker容器中执行Python代码
将任何标准输入输出的MCP服务器代理到沙箱环境
运行时发现工具（无需预加载上下文）
在所有已连接服务器中进行模糊搜索
持久化会话（保留变量/状态）
安全性：无网络访问、只读文件系统、移除权限

Installation

安装

Prerequisites

前置要求

Install a container runtime (rootless mode):

bash

undefined

安装容器运行时（无根模式）：

bash

undefined

macOS (Podman Desktop recommended)

macOS（推荐使用Podman Desktop）

brew install podman podman machine init podman machine start

Linux

sudo apt install podman # or dnf/yum/pacman podman system migrate # enable rootless

sudo apt install podman # 或使用dnf/yum/pacman podman system migrate # 启用无根模式

Windows (WSL2 + Podman Desktop)

Windows（WSL2 + Podman Desktop）

Download from https://podman-desktop.io/

从https://podman-desktop.io/下载

undefined

undefined

Install the Bridge

安装桥接工具

bash

undefined

bash

undefined

Via pip

通过pip安装

pip install mcp-code-execution

Via uv (recommended)

通过uv安装（推荐）

uv pip install mcp-code-execution

From source

从源码安装

git clone https://github.com/elusznik/mcp-server-code-execution-mode.git cd mcp-server-code-execution-mode uv pip install -e .

undefined

git clone https://github.com/elusznik/mcp-server-code-execution-mode.git cd mcp-server-code-execution-mode uv pip install -e .

undefined

Pull the Container Image

拉取容器镜像

bash

undefined

bash

undefined

Pre-built image (recommended)

预构建镜像（推荐）

podman pull ghcr.io/elusznik/mcp-code-execution:latest

Or build custom image

或构建自定义镜像

podman build -t mcp-code-execution:custom -f Dockerfile .

undefined

podman build -t mcp-code-execution:custom -f Dockerfile .

undefined

Configuration

配置

Claude Desktop

Claude桌面端

Add to

~/Library/Application Support/Claude/claude_desktop_config.json

(macOS) or

%APPDATA%\Claude\claude_desktop_config.json

(Windows):

json

{
  "mcpServers": {
    "code-execution": {
      "command": "uv",
      "args": [
        "--directory",
        "/path/to/mcp-server-code-execution-mode",
        "run",
        "mcp-code-execution"
      ],
      "env": {
        "MCP_BRIDGE_RUNTIME": "podman",
        "MCP_BRIDGE_IMAGE": "ghcr.io/elusznik/mcp-code-execution:latest",
        "MCP_BRIDGE_OUTPUT_MODE": "compact"
      }
    }
  }
}

添加至

~/Library/Application Support/Claude/claude_desktop_config.json

（macOS）或

%APPDATA%\Claude\claude_desktop_config.json

（Windows）：

json

{
  "mcpServers": {
    "code-execution": {
      "command": "uv",
      "args": [
        "--directory",
        "/path/to/mcp-server-code-execution-mode",
        "run",
        "mcp-code-execution"
      ],
      "env": {
        "MCP_BRIDGE_RUNTIME": "podman",
        "MCP_BRIDGE_IMAGE": "ghcr.io/elusznik/mcp-code-execution:latest",
        "MCP_BRIDGE_OUTPUT_MODE": "compact"
      }
    }
  }
}

Environment Variables

环境变量

Variable	Default	Description
`MCP_BRIDGE_RUNTIME`	`podman`	Container runtime ( `podman` or `docker` )
`MCP_BRIDGE_IMAGE`	`ghcr.io/elusznik/mcp-code-execution:latest`	Container image to use
`MCP_BRIDGE_TIMEOUT`	`300`	Execution timeout (seconds)
`MCP_BRIDGE_OUTPUT_MODE`	`compact`	Output format ( `compact` , `toon` , or `json` )
`MCP_BRIDGE_MEMORY_LIMIT`	`512m`	Container memory limit
`MCP_BRIDGE_PIDS_LIMIT`	`128`	Max processes in container
`MCP_BRIDGE_STARTUP_TIMEOUT`	`60`	Server startup timeout (seconds)

变量	默认值	描述
`MCP_BRIDGE_RUNTIME`	`podman`	容器运行时（ `podman` 或 `docker` ）
`MCP_BRIDGE_IMAGE`	`ghcr.io/elusznik/mcp-code-execution:latest`	使用的容器镜像
`MCP_BRIDGE_TIMEOUT`	`300`	执行超时时间（秒）
`MCP_BRIDGE_OUTPUT_MODE`	`compact`	输出格式（ `compact` 、 `toon` 或 `json` ）
`MCP_BRIDGE_MEMORY_LIMIT`	`512m`	容器内存限制
`MCP_BRIDGE_PIDS_LIMIT`	`128`	容器内最大进程数
`MCP_BRIDGE_STARTUP_TIMEOUT`	`60`	服务器启动超时时间（秒）

Proxying Other MCP Servers

代理其他MCP服务器

Create a

mcp_bridge_config.json

in your working directory:

json

{
  "servers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"],
      "env": {}
    },
    "github": {
      "command": "uvx",
      "args": ["mcp-server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "${GITHUB_TOKEN}"
      }
    },
    "postgres": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-e", "POSTGRES_CONNECTION_STRING",
        "mcp/postgres"
      ],
      "env": {
        "POSTGRES_CONNECTION_STRING": "${DATABASE_URL}"
      }
    }
  }
}

The bridge auto-discovers configs from:

```
./mcp_bridge_config.json
```
```
~/.config/mcp-bridge/config.json
```
```
MCP_BRIDGE_CONFIG
```
env var (JSON string)
Claude Desktop config (proxies sibling servers)

在工作目录创建

mcp_bridge_config.json

：

json

{
  "servers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"],
      "env": {}
    },
    "github": {
      "command": "uvx",
      "args": ["mcp-server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "${GITHUB_TOKEN}"
      }
    },
    "postgres": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-e", "POSTGRES_CONNECTION_STRING",
        "mcp/postgres"
      ],
      "env": {
        "POSTGRES_CONNECTION_STRING": "${DATABASE_URL}"
      }
    }
  }
}

桥接工具会自动从以下位置发现配置：

```
./mcp_bridge_config.json
```
```
~/.config/mcp-bridge/config.json
```
```
MCP_BRIDGE_CONFIG
```
环境变量（JSON字符串）
Claude桌面端配置（代理同级服务器）

Usage Patterns

使用模式

Basic Code Execution

基础代码执行

python

undefined

python

undefined

Simple calculation

简单计算

result = 42 * 1.5 print(f"Answer: {result}")


```python

result = 42 * 1.5 print(f"Answer: {result}")


```python

Data analysis

数据分析

import pandas as pd import numpy as np

data = pd.DataFrame({ 'x': np.random.randn(100), 'y': np.random.randn(100) })

correlation = data['x'].corr(data['y']) print(f"Correlation: {correlation:.3f}")

undefined

import pandas as pd import numpy as np

data = pd.DataFrame({ 'x': np.random.randn(100), 'y': np.random.randn(100) })

correlation = data['x'].corr(data['y']) print(f"Correlation: {correlation:.3f}")

undefined

Discovering MCP Servers

发现MCP服务器

python

from mcp import runtime

python

from mcp import runtime

List all available servers

列出所有可用服务器

servers = await runtime.discovered_servers() for server in servers: print(f"- {server}")

undefined

servers = await runtime.discovered_servers() for server in servers: print(f"- {server}")

undefined

Querying Tool Schemas

查询工具Schema

python

from mcp import runtime

python

from mcp import runtime

Get all tools from a specific server

获取特定服务器的所有工具

github_tools = await runtime.query_tool_docs("github")

for tool_name, schema in github_tools.items(): print(f"{tool_name}: {schema.get('description', 'No description')}")

undefined

github_tools = await runtime.query_tool_docs("github")

for tool_name, schema in github_tools.items(): print(f"{tool_name}: {schema.get('description', 'No description')}")

undefined

Fuzzy Tool Search

模糊工具搜索

python

from mcp import runtime

python

from mcp import runtime

Search across all servers

在所有服务器中搜索

matches = await runtime.search_tool_docs("create issue", limit=5)

for hit in matches: print(f"{hit['server']}.{hit['tool']}: {hit.get('description', '')}") print(f" Score: {hit.get('score', 0):.2f}")

undefined

matches = await runtime.search_tool_docs("create issue", limit=5)

for hit in matches: print(f"{hit['server']}.{hit['tool']}: {hit.get('description', '')}") print(f" Score: {hit.get('score', 0):.2f}")

undefined

Calling MCP Tools Directly

直接调用MCP工具

python

undefined

python

undefined

Dynamic lookup

动态查找

result = await mcp_servers["github"].call_tool( "create_issue", { "owner": "elusznik", "repo": "mcp-server-code-execution-mode", "title": "Add feature X", "body": "Description here" } ) print(result)


```python

result = await mcp_servers["github"].call_tool( "create_issue", { "owner": "elusznik", "repo": "mcp-server-code-execution-mode", "title": "Add feature X", "body": "Description here" } ) print(result)


```python

Attribute access

属性访问

result = await mcp_github.create_issue( owner="elusznik", repo="mcp-server-code-execution-mode", title="Bug report", body="Steps to reproduce..." )


```python

result = await mcp_github.create_issue( owner="elusznik", repo="mcp-server-code-execution-mode", title="Bug report", body="Steps to reproduce..." )


```python

Module import pattern

模块导入模式

from mcp.servers.github import create_issue

issue = await create_issue( owner="myorg", repo="myrepo", title="Task", body="Details" )

undefined

from mcp.servers.github import create_issue

issue = await create_issue( owner="myorg", repo="myrepo", title="Task", body="Details" )

undefined

Composing Multiple Tools

组合多个工具

python

from mcp import runtime

python

from mcp import runtime

1. Search for calendar tools

1. 搜索日历工具

cal_tools = await runtime.search_tool_docs("calendar events", limit=3) calendar_server = cal_tools[0]["server"] if cal_tools else None

if calendar_server: # 2. Get today's events events = await mcp_servers[calendar_server].call_tool( "list_events", {"date": "2025-01-15"} )

# 3. Create GitHub issues for each event
for event in events:
    await mcp_github.create_issue(
        owner="myorg",
        repo="tasks",
        title=f"Follow-up: {event['title']}",
        body=f"From calendar: {event['description']}"
    )

print(f"Created {len(events)} issues")

undefined

cal_tools = await runtime.search_tool_docs("calendar events", limit=3) calendar_server = cal_tools[0]["server"] if cal_tools else None

if calendar_server: # 2. 获取今日事件 events = await mcp_servers[calendar_server].call_tool( "list_events", {"date": "2025-01-15"} )

# 3. 为每个事件创建GitHub Issue
for event in events:
    await mcp_github.create_issue(
        owner="myorg",
        repo="tasks",
        title=f"Follow-up: {event['title']}",
        body=f"From calendar: {event['description']}"
    )

print(f"Created {len(events)} issues")

undefined

Error Handling

错误处理

python

from mcp import runtime

try:
    result = await mcp_servers["github"].call_tool(
        "get_issue",
        {"owner": "invalid", "repo": "repo", "issue_number": 999}
    )
except Exception as e:
    print(f"Tool call failed: {e}")
    
    # Fallback: search for alternative tools
    alternatives = await runtime.search_tool_docs("get issue")
    print(f"Found {len(alternatives)} alternative tools")

python

from mcp import runtime

try:
    result = await mcp_servers["github"].call_tool(
        "get_issue",
        {"owner": "invalid", "repo": "repo", "issue_number": 999}
    )
except Exception as e:
    print(f"Tool call failed: {e}")
    
    # 备选方案：搜索替代工具
    alternatives = await runtime.search_tool_docs("get issue")
    print(f"Found {len(alternatives)} alternative tools")

Persistent State

持久化状态

Variables and imports persist across calls in the same session:

python

undefined

同一会话中，变量和导入内容会在多次调用间保留：

python

undefined

First call

第一次调用

import pandas as pd df = pd.DataFrame({'a': [1, 2, 3], 'b': [4, 5, 6]})


```python

import pandas as pd df = pd.DataFrame({'a': [1, 2, 3], 'b': [4, 5, 6]})


```python

Second call (same session)

第二次调用（同一会话）

df and pandas are still available

df和pandas仍可用

print(df.describe())

undefined

print(df.describe())

undefined

Advanced Patterns

进阶模式

Building a Tool Catalog

构建工具目录

python

from mcp import runtime
import json

catalog = {}

python

from mcp import runtime
import json

catalog = {}

Discover all servers

发现所有服务器

servers = await runtime.discovered_servers()

for server_name in servers: # Get all tools for this server tools = await runtime.query_tool_docs(server_name) catalog[server_name] = { tool_name: { "description": schema.get("description", ""), "parameters": schema.get("inputSchema", {}).get("properties", {}) } for tool_name, schema in tools.items() }

servers = await runtime.discovered_servers()

for server_name in servers: # 获取该服务器的所有工具 tools = await runtime.query_tool_docs(server_name) catalog[server_name] = { tool_name: { "description": schema.get("description", ""), "parameters": schema.get("inputSchema", {}).get("properties", {}) } for tool_name, schema in tools.items() }

Save catalog

保存目录

with open("/tmp/tool_catalog.json", "w") as f: json.dump(catalog, f, indent=2)

print(f"Cataloged {len(catalog)} servers")

undefined

with open("/tmp/tool_catalog.json", "w") as f: json.dump(catalog, f, indent=2)

print(f"Cataloged {len(catalog)} servers")

undefined

Conditional Tool Execution

条件工具执行

python

from mcp import runtime

python

from mcp import runtime

Search for weather tools

搜索天气工具

weather_tools = await runtime.search_tool_docs("weather forecast", limit=1)

if weather_tools: server = weather_tools[0]["server"] tool = weather_tools[0]["tool"]

forecast = await mcp_servers[server].call_tool(
    tool,
    {"location": "San Francisco"}
)

# If rain predicted, create calendar reminder
if "rain" in forecast.lower():
    cal_tools = await runtime.search_tool_docs("create event")
    if cal_tools:
        await mcp_servers[cal_tools[0]["server"]].call_tool(
            cal_tools[0]["tool"],
            {
                "title": "Bring umbrella",
                "date": "2025-01-16",
                "time": "08:00"
            }
        )

undefined

weather_tools = await runtime.search_tool_docs("weather forecast", limit=1)

if weather_tools: server = weather_tools[0]["server"] tool = weather_tools[0]["tool"]

forecast = await mcp_servers[server].call_tool(
    tool,
    {"location": "San Francisco"}
)

# 如果预测有雨，创建日历提醒
if "rain" in forecast.lower():
    cal_tools = await runtime.search_tool_docs("create event")
    if cal_tools:
        await mcp_servers[cal_tools[0]["server"]].call_tool(
            cal_tools[0]["tool"],
            {
                "title": "Bring umbrella",
                "date": "2025-01-16",
                "time": "08:00"
            }
        )

undefined

Batch Processing

批量处理

python

import asyncio

repos = ["repo1", "repo2", "repo3"]
results = []

for repo in repos:
    try:
        issues = await mcp_github.list_issues(
            owner="myorg",
            repo=repo,
            state="open"
        )
        results.append({"repo": repo, "count": len(issues)})
    except Exception as e:
        results.append({"repo": repo, "error": str(e)})

for r in results:
    if "error" in r:
        print(f"{r['repo']}: ERROR - {r['error']}")
    else:
        print(f"{r['repo']}: {r['count']} open issues")

python

import asyncio

repos = ["repo1", "repo2", "repo3"]
results = []

for repo in repos:
    try:
        issues = await mcp_github.list_issues(
            owner="myorg",
            repo=repo,
            state="open"
        )
        results.append({"repo": repo, "count": len(issues)})
    except Exception as e:
        results.append({"repo": repo, "error": str(e)})

for r in results:
    if "error" in r:
        print(f"{r['repo']}: ERROR - {r['error']}")
    else:
        print(f"{r['repo']}: {r['count']} open issues")

Data Science Workflow

数据科学工作流

python

import pandas as pd
import matplotlib.pyplot as plt
from io import StringIO

python

import pandas as pd
import matplotlib.pyplot as plt
from io import StringIO

Fetch data from MCP tool

从MCP工具获取数据

csv_data = await mcp_filesystem.read_file(path="/tmp/sales.csv")

Process

处理数据

df = pd.read_csv(StringIO(csv_data)) summary = df.groupby("region")["sales"].sum().sort_values(ascending=False)

Generate chart

生成图表

fig, ax = plt.subplots() summary.plot(kind="bar", ax=ax) plt.title("Sales by Region") plt.tight_layout() plt.savefig("/tmp/sales_chart.png")

Save results

保存结果

report = f""" Sales Analysis

Total Sales: ${df['sales'].sum():,.2f} Top Region: {summary.index[0]} (${summary.iloc[0]:,.2f})

Chart saved to /tmp/sales_chart.png """

await mcp_filesystem.write_file( path="/tmp/sales_report.txt", content=report )

print(report)

undefined

report = f""" Sales Analysis

Total Sales: ${df['sales'].sum():,.2f} Top Region: {summary.index[0]} (${summary.iloc[0]:,.2f})

Chart saved to /tmp/sales_chart.png """

await mcp_filesystem.write_file( path="/tmp/sales_report.txt", content=report )

print(report)

undefined

Troubleshooting

故障排除

Container Runtime Not Found

未找到容器运行时

Error:

RuntimeNotFoundError: Neither podman nor docker found

Solution:

bash

undefined

错误：

RuntimeNotFoundError: Neither podman nor docker found

解决方案：

bash

undefined

Install Podman

安装Podman

brew install podman # macOS sudo apt install podman # Linux

Or install Docker

或安装Docker

brew install docker # macOS (with Docker Desktop)

undefined

brew install docker # macOS（需搭配Docker Desktop）

undefined

Permission Denied (Rootless)

权限拒绝（无根模式）

Error:

Error: creating container: mkdir /run/user/1000: permission denied

Solution:

bash

undefined

错误：

Error: creating container: mkdir /run/user/1000: permission denied

解决方案：

bash

undefined

Enable rootless mode

启用无根模式

podman system migrate

Or use Docker rootless

或使用Docker无根模式

dockerd-rootless-setuptool.sh install

undefined

dockerd-rootless-setuptool.sh install

undefined

Image Pull Timeout

镜像拉取超时

Error:

TimeoutError: Container image pull exceeded 60s

Solution:

bash

undefined

错误：

TimeoutError: Container image pull exceeded 60s

解决方案：

bash

undefined

Pre-pull image manually

手动预拉取镜像

podman pull ghcr.io/elusznik/mcp-code-execution:latest

Or increase timeout

或增加超时时间

export MCP_BRIDGE_STARTUP_TIMEOUT=180

undefined

export MCP_BRIDGE_STARTUP_TIMEOUT=180

undefined

Server Discovery Fails

服务器发现失败

Error: No servers found via

discovered_servers()

Solution:

Check config file exists:
bash
```
cat ~/.config/mcp-bridge/config.json
```

Validate JSON syntax:

python

import json
with open("mcp_bridge_config.json") as f:
    json.load(f)  # Should not raise

Test server manually:

bash

npx -y @modelcontextprotocol/server-filesystem /tmp
# Should output JSON-RPC messages

错误：

discovered_servers()

未找到任何服务器

解决方案：

检查配置文件是否存在：
bash
```
cat ~/.config/mcp-bridge/config.json
```

验证JSON语法：

python

import json
with open("mcp_bridge_config.json") as f:
    json.load(f)  # 不应抛出异常

手动测试服务器：

bash

npx -y @modelcontextprotocol/server-filesystem /tmp
# 应输出JSON-RPC消息

Tool Call Hangs

工具调用挂起

Error: Tool execution never completes

Solution:

Check timeout settings:

bash

export MCP_BRIDGE_TIMEOUT=600  # Increase to 10 minutes

Debug in container:

bash

podman run -it --rm ghcr.io/elusznik/mcp-code-execution:latest /bin/bash
# Test commands manually

错误： 工具执行从未完成

解决方案：

检查超时设置：

bash

export MCP_BRIDGE_TIMEOUT=600  # 增加至10分钟

在容器中调试：

bash

podman run -it --rm ghcr.io/elusznik/mcp-code-execution:latest /bin/bash
# 手动测试命令

Memory Limit Exceeded

内存限制超出

Error:

OOMKilled

or memory allocation errors

Solution:

bash

export MCP_BRIDGE_MEMORY_LIMIT=2g  # Increase to 2GB

错误：

OOMKilled

或内存分配错误

解决方案：

bash

export MCP_BRIDGE_MEMORY_LIMIT=2g  # 增加至2GB

Volume Mount Issues (macOS)

卷挂载问题（macOS）

Error: Files not accessible in container

Solution:

bash

undefined

错误： 容器内无法访问文件

解决方案：

bash

undefined

Ensure Podman machine has volume sharing enabled

确保Podman机器已启用卷共享

podman machine stop podman machine set --rootful=false --volume /Users:/Users podman machine start

undefined

podman machine stop podman machine set --rootful=false --volume /Users:/Users podman machine start

undefined

Output Format Issues

输出格式问题

Problem: Responses too verbose or unstructured

Solution:

bash

undefined

问题： 响应过于冗长或非结构化

解决方案：

bash

undefined

Use compact mode (default)

使用紧凑模式（默认）

export MCP_BRIDGE_OUTPUT_MODE=compact

Or TOON for deterministic tokens

或使用TOON模式获取确定性tokens

export MCP_BRIDGE_OUTPUT_MODE=toon

undefined

export MCP_BRIDGE_OUTPUT_MODE=toon

undefined

Security Best Practices

安全最佳实践

Never expose the bridge directly to untrusted users — the agent can execute arbitrary Python.

Use environment variables for secrets:

json

{
  "servers": {
    "github": {
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "${GITHUB_TOKEN}"
      }
    }
  }
}

Restrict filesystem access for proxied servers:

json

{
  "servers": {
    "filesystem": {
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/safe/path"]
    }
  }
}

Monitor resource usage:

bash

podman stats $(podman ps -q --filter ancestor=ghcr.io/elusznik/mcp-code-execution)

Audit logs: Check container logs for suspicious activity:
bash
```
podman logs <container-id>
```

切勿将桥接工具直接暴露给不可信用户 — Agent可执行任意Python代码。

使用环境变量存储密钥：

json

{
  "servers": {
    "github": {
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "${GITHUB_TOKEN}"
      }
    }
  }
}

限制代理服务器的文件系统访问：

json

{
  "servers": {
    "filesystem": {
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/safe/path"]
    }
  }
}

监控资源使用：

bash

podman stats $(podman ps -q --filter ancestor=ghcr.io/elusznik/mcp-code-execution)

审计日志： 检查容器日志是否存在可疑活动：
bash
```
podman logs <container-id>
```

Performance Tips

性能优化建议

Warm up servers by calling
```
discovered_servers()
```
early in the session.

Cache tool schemas instead of querying repeatedly:

python

from mcp import runtime

# Cache at session start
all_tools = {}
for server in await runtime.discovered_servers():
    all_tools[server] = await runtime.query_tool_docs(server)

Use attribute access (slightly faster than dynamic lookup):

python

# Preferred
await mcp_github.create_issue(...)

# Slower
await mcp_servers["github"].call_tool("create_issue", ...)

Pre-pull images to avoid startup delays:

bash

podman pull ghcr.io/elusznik/mcp-code-execution:latest

提前预热服务器：在会话初期调用
```
discovered_servers()
```
。

缓存工具Schema：避免重复查询：

python

from mcp import runtime

# 在会话开始时缓存
all_tools = {}
for server in await runtime.discovered_servers():
    all_tools[server] = await runtime.query_tool_docs(server)

使用属性访问（比动态查找略快）：

python

# 推荐方式
await mcp_github.create_issue(...)

# 较慢方式
await mcp_servers["github"].call_tool("create_issue", ...)

预拉取镜像：避免启动延迟：

bash

podman pull ghcr.io/elusznik/mcp-code-execution:latest