llm-context.py MCP — 使用场景, 安装 & 实时演示

为什么要用

核心特性

Rule-based file selection (filter rules) — per-task context, not one-size-fits-all
Multiple rule types: prompt / filter / instruction / style / excerpt
lc_outlines compresses big codebases into structural outlines
lc_missing: LLM can fetch specific files mid-conversation
Clipboard-only mode for LLM clients that don't support MCP

实时演示

实际使用效果

llm-context-py.replay ▶ 就绪

0/0

安装

选择你的客户端

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

{
  "mcpServers": {
    "llm-context-py": {
      "command": "uvx",
      "args": [
        "llm-context.py"
      ],
      "_inferred": true
    }
  }
}

打开 Claude Desktop → Settings → Developer → Edit Config。保存后重启应用。

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "llm-context-py": {
      "command": "uvx",
      "args": [
        "llm-context.py"
      ],
      "_inferred": true
    }
  }
}

Cursor 使用与 Claude Desktop 相同的 mcpServers 格式。项目级配置优先于全局。

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "llm-context-py": {
      "command": "uvx",
      "args": [
        "llm-context.py"
      ],
      "_inferred": true
    }
  }
}

点击 Cline 侧栏中的 MCP Servers 图标，然后选 "Edit Configuration"。

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "llm-context-py": {
      "command": "uvx",
      "args": [
        "llm-context.py"
      ],
      "_inferred": true
    }
  }
}

格式与 Claude Desktop 相同。重启 Windsurf 生效。

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "llm-context-py",
      "command": "uvx",
      "args": [
        "llm-context.py"
      ]
    }
  ]
}

Continue 使用服务器对象数组，而非映射。

~/.config/zed/settings.json

{
  "context_servers": {
    "llm-context-py": {
      "command": {
        "path": "uvx",
        "args": [
          "llm-context.py"
        ]
      }
    }
  }
}

加入 context_servers。Zed 保存后热重载。

claude mcp add llm-context-py -- uvx llm-context.py

一行命令搞定。用 claude mcp list 验证，claude mcp remove 卸载。

使用场景

实战用法： llm-context.py

Ship a focused code-review context to your LLM

👤 Developers tired of pasting 3000 lines ⏱ ~15 min intermediate

何时使用： You're about to ask 'review my auth module' and don't want the whole repo in context.

前置条件

llm-context.py installed — uv tool install 'llm-context>=0.6.0'
Initialized in your repo — lc-init in your repo root

步骤

Create a filter rule for auth

Create an lc filter rule 'flt-auth' that includes src/auth/** and src/middleware/auth*.ts.✓ 已复制

→ Rule file created under .lc/
Preview what the rule selects

Run lc_preview on flt-auth — show me which files will be sent and total token count.✓ 已复制

→ File list + token count
Share context with the LLM via MCP

Using the flt-auth rule, review the module for security issues. If you need to see a specific file not included, ask via lc_missing.✓ 已复制

→ Review with targeted file requests

结果： A code review that fits in context and can still explore — no manual paste.

注意事项

Rule too narrow → LLM can't understand callers of your module — Include interface / types of neighboring modules, use outlines for the rest
Rule too broad → token budget blown — Start broad, watch lc_preview, tighten until under your client's limits

搭配使用： filesystem

Give an LLM a structural map of a huge codebase

👤 Engineers onboarding to unfamiliar monorepos ⏱ ~20 min intermediate

何时使用： 100k+ LOC repo; you need orientation, not all the code.

步骤

Generate outline

lc_outlines for the whole repo — classes, top-level functions, exports per file. Skip bodies.✓ 已复制

→ Skeletal outline with a few thousand lines
Ask orientation questions

Given the outline, where's the entry point, where's routing defined, and which files hold the data-layer abstractions?✓ 已复制

→ Architectural answers
Drill in on one area

Show me the full contents of the three files that define routing via lc_missing.✓ 已复制

→ Specific files expand into context

结果： A guided tour without ever pasting the whole repo.

Attach project style rules to every prompt

👤 Teams with specific conventions ⏱ ~10 min beginner

何时使用： You want Claude to always know 'we use pytest, not unittest' without repeating.

步骤

Write a style rule

Create sty-python rule: 'pytest only, type hints required, black formatting'.✓ 已复制

→ Rule saved
Apply automatically

Use prm-default which composes flt-current + sty-python + ins-standards for every context.✓ 已复制

→ Rule auto-attached

结果： Per-task conventions enforced without manual boilerplate.

搭配使用： drift

组合

与其他 MCP 搭配，撬动十倍杠杆

llm-context-py + drift

drift records conventions; llm-context pushes only relevant files + style rules per task

Load drift conventions for this repo, then use lc with rule flt-auth + sty-ts to review the auth module.✓ 已复制

llm-context-py + filesystem

After reviewing, filesystem applies edits

Based on the llm-context review, use filesystem to apply the suggested edits to src/auth/.✓ 已复制

工具

此 MCP 暴露的能力

工具	输入参数	何时调用	成本
lc_outlines	rule?: str	Orient on large codebases	free (local)
lc_preview	rule: str	Before shipping context, verify scope	free
lc_missing	path: str	LLM calls this mid-conversation to request a file	free

成本与限制

运行它的成本

API 配额: None
每次调用 Token 数: Depends entirely on your rule scope — that's the point
费用: Free, open source
提示: Always run lc_preview before lc_outlines/shipping context — one second of preview saves thousands of tokens.

安全

权限、密钥、影响范围

凭据存储： None

数据出站： Only to whichever LLM client/provider you pipe context into

Review rule globs before running — a stray **/* will happily include node_modules.
lc_missing is powerful; constrain its allowed paths in rules if you store secrets in the repo.

故障排查

常见错误与修复

No rules found

Run lc-init in repo root to create .lc/ scaffolding.

验证： ls .lc/

lc_preview token count surprisingly high

Your filter is too loose or includes generated files. Narrow globs and add ignore patterns.

验证： lc_preview again

MCP tool not available

Use uvx --from llm-context lc-mcp in your MCP server config.

验证： claude mcp list

替代方案

llm-context.py 对比其他方案

替代方案	何时用它替代	权衡
repomix / ai-digest	You want a single-file dump, not MCP tools	No interactive lc_missing; static snapshot
filesystem MCP	You want raw file access	No rule-based selection or outlines
drift	You want persistent convention memory, not per-task file bundling	Different problem entirely

llm-context.py