Single-agent runtime for local skill and tool orchestration. Integrates MCP tools, builtin filesystem/shell tools, and skill loading.
This MCP server has significant quality gaps despite implementing 10 tools across multiple domains. Tool definitions are partially documented but lack consistency in description depth, output schema documentation, and error handling guidance. Most tools have input schemas defined, but descriptions are often generic or minimal (particularly fetch, load_skill, tool_search). No tools document their output schemas formally. Error handling is minimal across all tools. The bash and load_skill tools are high-risk (DESTRUCTIVE) but lack comprehensive validation or permission framework documentation. Naming is generally clear (verb_noun convention), but composition issues exist: load_skill and tool_search are meta-tools that bridge capability expansion, creating unclear separation of concerns. Most tools cluster in the 45-60 range; only read/write/append/glob/grep meet minimal quality standards due to detailed parameter documentation.
追加内容到文件末尾。如果文件不存在则创建新文件。用于分块写入大文件:先用 write 创建文件,再用 append 追加后续内容。
Execute a shell command and return the result. On Windows this uses PowerShell by default.
在文件中进行精确字符串替换。使用精确匹配(非正则),支持替换单个或所有匹配。
获取指定 URL 的网页内容,自动清理 HTML/XML 标签、脚本、样式、导航等噪音,返回干净的纯文本。适用于:搜索到链接后抓取页面详情、阅读文章/新闻全文、获取 API 数据。内容会自动截断到 max_length。如果只需要关键信息,建议设置较小的 max_length(如 3000)以节省上下文。注意:本工具无法获取需要 JavaScript 渲染的 SPA 页面(如 Yahoo Finance、Google Finance 等),对于此类网站,建议改用提供 API 或服务端渲染的数据源。
使用模式匹配查找文件。支持标准 glob 语法:* (任意字符), ** (递归目录), ? (单个字符), {a,b} (多选一), [abc] (字符集)。
使用正则表达式搜索文件内容。基于 ripgrep (rg),支持复杂的搜索模式和上下文显示。
Load and execute a custom skill from the skills directory. Skills are user-defined functions that extend the agent's capabilities.
fetch tool has no documented output schema. The description mentions 'returns clean plaintext' but does not specify the response object structure, field names, or whether it includes metadata like final_url, content_length, or fetch_timestamp that downstream tools might need.
load_skill and tool_search tools lack clear output schema documentation. load_skill_params accepts arbitrary objects ('object' type), and tool_search returns matching tools but the response structure is not documented. Ambiguity about what fields are returned forces LLMs to guess downstream behavior.
bash tool description is minimal (one sentence) and does not explain when to use it, what makes it destructive, or how to handle failures. It should warn about side effects and unsafe usage patterns.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 49 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 44 | - | v1 |
读取文件内容,返回带行号的格式(类似 cat -n)。支持按行读取,可指定起始行和行数限制。
Search for tools across deferred MCP servers using semantic search (BM25). Returns matching tools that can be immediately injected into the conversation context.
写入或覆盖文件内容。支持创建新文件或完全覆盖已有文件。使用原子写入确保文件完整性。
fetch tool description contains example values ('Yahoo Finance', 'Google Finance') and specific limitation callouts, but the description is overly long (500+ chars) and buries the core functionality. It should be concise (50-200 chars) with limitations in separate documentation or error messages.
edit tool parameter 'old_string' requires unique matching when replace_all=false, but there is no output confirmation of what was actually replaced or how many replacements occurred. LLMs cannot verify success and may retry unnecessarily.
No error handling or recovery guidance documented across any tool. Bash could fail with command not found, permission denied, or timeout, but the description provides no recovery hints. Same for file operations (path not found, permission denied, disk full). LLMs need explicit guidance: 'If file not found, check the path with glob() first.'
bash tool accepts arbitrary commands and has timeout configuration, but no validation rules or dangerous-command detection is documented. LLMs could pass 'rm -rf /' or similar destructive commands without explicit warnings in the tool description.
glob and grep tools return file lists/matches but do not document a max result limit. If a glob matches thousands of files, or grep finds thousands of matches, the response could exceed context window limits. The descriptions mention no pagination or truncation strategy.
tool_search uses semantic search (BM25) but the output structure is not documented. Does it return tool_name, description, server_name, and parameters? Can the LLM immediately inject returned tools into the conversation, or does it need to parse and re-register them? Unclear composition breaks downstream usage.
read, write, append, edit, and bash tools all operate on the filesystem but do not document whether they are sandboxed to a workspace_root or can access arbitrary system paths. Security implications are unclear, can an LLM agent read /etc/passwd or is there path traversal protection?