
agent-godmode
io.github.mdvohra由 GitHub 账号 mdvohra 发布v0.1.9更新于 Oct 24, 2026
Sandboxed workspace MCP: file ops, argv-only run_command, list_files; BYO LLM prompts.
概览
沙箱式工作区 MCP 服务器,让助手在单一根目录内读写编辑文件、列目录,并以 argv 方式执行命令。
- 可以做什么
- 读取 UTF-8 文本文件,可选行范围与字节上限
- 创建、覆盖或追加文本文件,并自动创建父目录
- 在已有文件中做查找替换,默认要求精确匹配一次,也可全部替换
- 列出目录条目,支持递归、glob、深度限制与隐藏文件开关
- 仅以 argv 列表方式运行子进程,可指定根目录内的工作目录
- 适用场景
- 适合让助手在本地代码或文档项目中工作,同时被限制在一个目录内,例如 Cursor 风格的编辑以及运行构建或测试命令。它还附带带版本号的系统提示词和 OpenAI 风格工具定义,便于在 Python 中接入自己的 LLM 循环。
- 运行要求
- 以 stdio 本地进程运行,从 PyPI 安装(pip install agent-godmode)。需要绝对路径的工作区根目录,通过环境变量 AGENT_GODMODE_ROOT 或命令行参数 --root 传入。可选变量可调整读取大小、命令超时、输出截断、列目录上限和允许的命令列表。服务器本身不需要 API 密钥;README 中的 OpenAI 示例需在你自己的应用中设置 OPENAI_API_KEY。
安装
在 SourceWeft 中
- 打开 控制台中的 agent-godmode,将其添加到工作区。
- 为需要使用其工具的对话启用该服务。
Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。
其他 MCP 客户端
参照 仓库 中的启动说明。
README
agent-godmode
Workspace-scoped MCP tools for building Cursor-style agents: read_file, write_file, edit_file, run_command, list_files. Includes strict, versioned system prompts (SYSTEM_PROMPT_V1) and OpenAI-style tool definitions so your app can wire any LLM with one import.
The LLM and API keys stay in your app. This package provides tool execution, sandboxing, and prompts—not a hosted model.
OpenAI + in-process tools: the Tier B section below is self-contained—copy the Python into a script, module, or REPL; no separate artifact is required.
Install
Editable / dev:
Migrating from mcp-agent-tools: uninstall the old package, install agent-godmode, change Python imports from mcp_agent_tools to agent_godmode, the CLI from mcp-agent-tools to agent-godmode, and environment variables from MCP_AGENT_TOOLS_* to AGENT_GODMODE_* (for example AGENT_GODMODE_ROOT).
Tools
All tools are scoped to a single workspace root. Paths are relative to that root (or absolute only if they resolve under it). The same operations are available over MCP (the agent-godmode server) and in-process via AgentWorkspace / WorkspaceTools.
For LLM integrations, tool shapes and descriptions are centralized in OPENAI_TOOL_DEFINITIONS and TOOL_DESCRIPTIONS; agent behavior is guided by SYSTEM_PROMPT_V1.
Tier A — Cursor (or any MCP client)
1. Pick a workspace directory (only paths under this root are allowed).
2. Add a server entry (stdio). Example for a global MCP config (paths use forward slashes on Windows):
Or with an explicit CLI root (overrides env for that process):
3. Paste SYSTEM_PROMPT_V1 (from agent_godmode.prompts or below) into your host’s system prompt if the client does not load server instructions automatically.
Environment variables
Tier B — Python app (in-process + OpenAI)
Design notes
- Workspace root — Examples use
D:\Avi-assignas a placeholder; pointWORK_DIRat any directory you control. - API key policy —
OPENAI_API_KEYis required only for Chat Completions. Imports setclient = OpenAI() if HAS_OPENAI_KEY else None; workspace setup and directedit_filerun without a key. - Model-authored I/O — For
write_file, persist only text returned by the model. Foredit_file, the model must copyold_stringexactly fromread_file(seeSYSTEM_PROMPT_V1).
1. Install dependencies
In a shell or any interactive Python session:
If your environment supports line magics (for example %pip in IPython), you can run the same installs there; do not place shell comments on the same line as %pip.
2. Imports and API key handling
3. Workspace bootstrap and seed file
4. LLM-authored file body (no tool calls)
Requires OPENAI_API_KEY. Skip if you are only exercising tools without the API.
5. Direct edit_file (no Chat Completions)
No API key required. The next lines create ws if you have not run the workspace section yet (same root).
6. Agent loop: model calls write_file
Requires OPENAI_API_KEY.
7. Agent loop: model calls edit_file
Requires OPENAI_API_KEY and the complete function from §6.
8. Optional: custom tool loop without run_agent_loop
Use OPENAI_TOOL_DEFINITIONS, call the Chat Completions API with tools=..., parse tool_calls, and route each call through ws.dispatch(name, json.loads(arguments)) (requires import json). For write_file, the content field should be whatever the model authored; for edit_file, pass old_string, new_string, and replace_all exactly as the model returned.
Optional limits on AgentWorkspace
Lower-level (WorkspaceTools + OPENAI_TOOL_DEFINITIONS)
Same sandbox without AgentWorkspace: use config_from_root(...) and WorkspaceTools. Pass OPENAI_TOOL_DEFINITIONS to your provider as tools= when you implement your own loop instead of run_agent_loop.
Compose the system message:
Imports reference
AgentWorkspace— pass a directory path; useread_file/write_file/edit_file/list_files/run_commandon that tree onlySYSTEM_PROMPT_V1,SYSTEM_PROMPT_CHANGELOG,TOOL_DESCRIPTIONSOPENAI_TOOL_DEFINITIONS— same shapes as MCP tools (fortools=in chat completions)build_server(config)— build aFastMCPapp (stdio viabuild_server(cfg).run())run_agent_loop— minimal multi-turn executor with yourcompletecallable (acceptsWorkspaceConfigorAgentWorkspace)
Safety model
- Python: all paths are resolved under the directory you passed to
AgentWorkspace(...)orconfig_from_root(...). - MCP / CLI: same rule via
AGENT_GODMODE_ROOTor--root(no..escape). read_file,write_file, andedit_fileonly touch UTF-8 text paths under that root;edit_filerequires valid UTF-8 (strict decode).run_commandusesargvonly (no shell). Optional allowlist viaAGENT_GODMODE_ALLOWED_COMMANDS.- Subprocess inherits the current environment; avoid passing secrets you do not want child processes to see.
CLI
Runs the MCP server on stdio (default for Cursor).
Agent loop (conceptual)
- System =
SYSTEM_PROMPT_V1(+ optional suffix). - User message +
OPENAI_TOOL_DEFINITIONS→ your LLM. - For each
tool_call, runWorkspaceTools.dispatch(or MCPcall_tool) — includingread_file,write_file,edit_file,list_files, andrun_commandas defined by the server. - Append tool results; repeat until the model returns text without tools.
run_agent_loop implements steps 2–4 given your complete() function.
License
MIT
来源:README.md,提交 2763967
工具
0版本历史
1- v0.1.9最新Sep 16, 2026
评分
已满足适用检查项的 40%
该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。
- 近 30 天内可连通25 分
请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。
- 已声明认证方式10 分
请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。
- 已列出工具15 分
请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。
- 描述必需15 分
- 图标10 分
- 文档15 分
- 许可证10 分
请在仓库或注册中心条目中声明许可证。
