一个 Model Context Protocol (MCP) 服务器，封装了本地 Wolfram Engine，使 AI 助手（Claude、ChatGPT 等）能够通过 Wolfram 语言执行符号数学、数值分析和数据可视化。
免责声明： 这是一个非官方、独立的个人项目。 它不隶属于 Wolfram Research, Inc.，也不受其赞助、认可或认证。“Wolfram”、“Wolfram Language”、“Wolfram Engine”、“Mathematica”及相关标志是 Wolfram Research 的商标。
本软件不包含任何 Wolfram Engine / Mathematica 二进制文件、激活密钥、许可证文件或其他专有材料。用户必须根据 Wolfram 的许可条款独立获取并合法授权其自己的 Wolfram Engine 或 Mathematica 副本。
本项目的唯一目的是允许已获得许可的个人在 AI 助手的辅助下，在其自己的机器上调用其本地安装的 Wolfram 内核，且在许可允许的范围内使用。 将 Wolfram Engine 的访问权限重新分发给第三方并非本项目的预期用途，且可能违反 Wolfram 的许可条款。
MCP 工具：evaluate（文本）和 evaluate_image（PNG，实验性）——通过两个通用工具实现所有 Wolfram 语言功能
传输方式：stdio（本地）和流式 HTTP
安全性：内核预处理表达式过滤，支持黑名单/白名单模式及 29 个功能组
客户端 RBAC：基于客户端的凭据，以及基于角色的工具和安全策略控制——用于在同一台机器上隔离不同的 AI 客户端
OAuth 2.1：用于基于 Web 的 MCP 客户端（Claude.ai、ChatGPT）的授权服务器
配置驱动：单个 TOML 文件控制所有行为
Related MCP server: MCP Mathematics
Python 3.11+
Wolfram Engine 或 Mathematica（已获得合法授权）
uv 包管理器
# Clone and install
git clone https://github.com/siqiliu-tsinghua/mma-mcp.git
cd mma-mcp
uv sync
# Graphics export dependencies (headless servers only — desktops already have these)
sudo apt-get install -y libfontconfig1 libgl1 libasound2t64 libxkbcommon0 libegl1
# Generate default config
uv run mma-mcp init
# Generate security group files (requires Wolfram kernel, ~1 min)
uv run mma-mcp setup
# Start server (stdio, for local MCP clients)
uv run mma-mcp serve添加到你的 .mcp.json：
{
"mcpServers": {
"mma-mcp": {
"command": "uv",
"args": ["--directory", "/path/to/mma-mcp", "run", "mma-mcp"]
}
}
}添加到你的 claude_desktop_config.json（设置 -> 开发者 -> 编辑配置）：
{
"mcpServers": {
"mma-mcp": {
"command": "/path/to/mma-mcp/.venv/bin/mma-mcp"
}
}
}在 macOS/Linux 上，配置文件位于
~/Library/Application Support/Claude/claude_desktop_config.json或~/.config/Claude/claude_desktop_config.json。
uv run mma-mcp serve --transport http --host 127.0.0.1 --port 8000所有设置均位于 mma_mcp.toml（或 pyproject.toml 中的 [tool.mma-mcp] 部分）。
uv run mma-mcp init # generates mma_mcp.toml with comments关键部分：
部分 | 描述 |
| Wolfram 内核路径、超时时间、输出格式 |
| 传输模式、主机、端口 |
| 黑名单/白名单模式、功能组 |
| 要公开的 MCP 工具 |
| 用于 HTTPS 的域名和 DNS 提供商 (Caddy) |
| 客户端身份和基于角色的访问控制 |
表达式在到达 Wolfram 内核之前会被过滤。符号通过正则表达式提取，并根据活动策略进行检查。
黑名单模式（默认）：阻止危险组（system_exec、文件 I/O、网络、动态评估）。
白名单模式：仅允许来自明确启用组的符号。
29 个功能组（22 个安全 + 7 个危险）涵盖了约 6000 个 Wolfram 语言符号。可从本地内核重新生成：
uv run mma-mcp setup # required after cloning (generates from your local kernel)
uv run mma-mcp setup --force # force regeneration (e.g., after Wolfram Engine upgrade)使用 HTTP 传输时，你可以配置每个客户端的凭据和角色，以隔离连接到同一内核的不同 AI 客户端（例如 Claude 和 ChatGPT）：
# Generate password hash
uv run mma-mcp hash-password
# Generate TOML snippet for a new client
uv run mma-mcp add-client alice --role admin每个客户端都绑定到一个角色，该角色控制其可以访问的工具、可以使用的 Wolfram 符号以及资源限制（超时、结果大小）。并发客户端通过内核工作池进行隔离——每个工具调用都在具有临时 WL 上下文的独立内核进程中运行。
有关配置详情，请参阅 mma_mcp.toml 中的 [auth] 部分。
# Run tests
uv run pytest tests/ -v
# Inspect MCP tools interactively
uv run mcp dev src/mma_mcp/server.py命令 | 描述 |
| 启动 MCP 服务器（默认） |
| 生成默认的 |
| 从本地内核生成安全组 JSON |
| 为 HTTPS 生成 Caddyfile |
| 为配置生成密码哈希 |
| 为新的 AI 客户端生成 TOML 片段 |
客户端 | 长时间计算 | 备注 |
Claude.ai | ✔ 支持 | 发送 |
ChatGPT | ✘ 可能会超时 | 不发送 |
Claude Desktop / Claude Code | 未测试 | 本地 stdio 传输 |
MIT — 仅适用于本仓库中的代码。Wolfram Engine / Mathematica 的使用受 Wolfram Research 自身的许可条款约束。
Evaluate a Wolfram Language expression and return the result as text.
Args: expression: A valid Wolfram Language expression string. form: Output format — TeXForm (default), OutputForm, InputForm, StandardForm, or TraditionalForm.
| Name | Required | Description | Default |
|---|---|---|---|
| expression | Yes | ||
| form | No |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries full burden. It discloses the core behavior (evaluates expressions, returns text results) and mentions format options, but doesn't cover important behavioral aspects like error handling, computational limits, authentication requirements, or rate limits. The description adds value beyond what would be in annotations but leaves significant gaps.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is perfectly structured and front-loaded: the first sentence states the core purpose, followed by a clean parameter section. Every sentence earns its place, with zero wasted words. The formatting with 'Args:' section enhances readability without verbosity.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the tool's moderate complexity (expression evaluation), no annotations, and no output schema, the description does well but has gaps. It thoroughly documents parameters and purpose, but doesn't describe return value format beyond 'text' or potential error conditions. For a computational tool with no structured safety hints, more behavioral context would be beneficial.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
With 0% schema description coverage, the description fully compensates by providing comprehensive parameter documentation. It clearly explains both parameters: 'expression' as 'a valid Wolfram Language expression string' and 'form' with its five possible values and default. This adds substantial meaning beyond the bare schema.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool's purpose with specific verb ('evaluate') and resource ('Wolfram Language expression'), and distinguishes it from sibling tool 'evaluate_image' by specifying it returns text rather than image results. The phrase 'return the result as text' explicitly differentiates it from the image-focused sibling.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides clear context about when to use this tool (for evaluating Wolfram Language expressions to get text results) and implies when not to use it (when image results are needed, suggesting 'evaluate_image' as an alternative). However, it doesn't explicitly state exclusion criteria or name the alternative tool directly in the main description.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Evaluate a Wolfram Language expression and return the result as a PNG image.
Useful for Plot, Graphics, or any expression with visual output.
Args: expression: A valid Wolfram Language expression string.
| Name | Required | Description | Default |
|---|---|---|---|
| expression | Yes |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
No annotations are provided, so the description carries the full burden of behavioral disclosure. While it mentions the tool evaluates expressions and returns PNG images, it lacks critical behavioral details such as error handling, performance characteristics, rate limits, authentication requirements, or what happens with invalid expressions. This leaves significant gaps in understanding how the tool behaves in practice.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is perfectly concise and well-structured. It starts with the core purpose, provides usage guidelines, then clearly documents the parameter. Every sentence earns its place with no redundant information, making it easy to scan and understand quickly.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the tool's moderate complexity (evaluating Wolfram Language expressions with visual output), no annotations, no output schema, and 0% schema description coverage, the description provides adequate basics but lacks completeness. It covers the purpose and parameter semantics well, but misses important behavioral context about how the tool operates, what errors might occur, and what the PNG output contains.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The description adds meaningful context for the single parameter: 'expression: A valid Wolfram Language expression string.' This clarifies what type of input is expected beyond the schema's basic 'string' type. Since schema description coverage is 0% and there's only one parameter, the description adequately compensates by specifying the parameter's purpose and format requirements.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool's purpose: 'Evaluate a Wolfram Language expression and return the result as a PNG image.' It specifies both the action (evaluate) and the resource (Wolfram Language expression), and distinguishes it from the sibling tool 'evaluate' by mentioning the visual output format (PNG image).
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides clear context on when to use this tool: 'Useful for Plot, Graphics, or any expression with visual output.' This gives guidance on appropriate use cases. However, it doesn't explicitly state when NOT to use it or name alternatives (like the sibling 'evaluate' tool for non-visual results), which prevents a perfect score.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
v0.1.0evaluateevaluate_imageScored across 2 tools
The two tools have clearly distinct purposes: evaluate returns text output in various formats, while evaluate_image returns PNG images for visual output. There is no overlap or ambiguity between them, as each targets a different output type for Wolfram Language expressions.
Both tools follow a consistent verb_noun pattern with 'evaluate' as the verb and descriptive suffixes ('_image') to differentiate them. The naming is predictable and aligned, making it easy to understand their relationship and functionality.
With only two tools, this server feels too thin for its apparent scope of evaluating Wolfram Language expressions. It covers text and image output but lacks other essential operations like querying Wolfram Alpha, handling errors, or managing sessions, which limits its utility in broader workflows.
The tool surface is severely incomplete for a Wolfram Language server. It only provides evaluation with text or image output, missing critical functionality such as data import/export, symbolic computation queries, step-by-step solutions, or integration with Wolfram Cloud services, leading to significant gaps in agent capabilities.
Hosted MCP server connecting AI assistants to 9,000+ apps and 40,000+ actions via Zapier.
MCP server for secureFlows: token-free URL builders and integration-linting tools for AI agents.
MCP server for building and testing AI agents with multi-model experimentation and insights.
A comprehensive Model Context Protocol (MCP) server that enables AI assistants to interact with yo…