Llm Context.py

What cyberchitta/llm-context.py does

The cyberchitta/llm-context.py MCP server curates code context from a local project for an agent, a chat prompt, or a dispatched sub-agent. It turns file-selection decisions into reusable rules instead of requiring a developer to copy files manually for every task.

A generated context pack can contain complete contents for selected files, structural excerpts for supporting files, a filtered listing of included and excluded paths, and a timestamp. This lets a consumer see what it received and gives follow-up commands a reference point for resolving missing or changed files.

The package also supports clipboard-based workflows. Without MCP, a developer can select and pack context for pasting into a chat. With MCP, the model can request additional context and check whether source files changed after the pack was created.

How it works

After initialization, the project receives an lc-curate-context skill under .claude/skills/, along with the .llm-context/ working area. The skill guides an agent through identifying needed files, writing a rule, validating the selection, and producing a pack. Claude Code loads the skill automatically; other environments can read the installed files directly.

Rules use YAML frontmatter and can compose filters, excerpting behavior, full-file patterns, and explanatory text. Files expected to be edited can be included in full, while supporting code can be represented through signatures and definitions. Selection is also affected by .gitignore files, so the resulting listing is intentionally limited to paths visible to the rule and repository filters.

The command-line workflow can send a pack to stdout for a pipe or redirect. An optional mode includes shell fetch instructions that a child agent can use to request omitted files from the repository root.

Setup and configuration

Install the package with uv tool install "llm-context>=0.6.0", change to the project root, and run lc-init. The MCP configuration runs lc-mcp through uvx:

{
  "mcpServers": {
    "llm-context": {
      "command": "uvx",
      "args": ["--from", "llm-context", "lc-mcp"]
    }
  }
}

The cyberchitta/llm-context.py MCP server does not document required environment variables. Upgrades use uv tool upgrade llm-context; subsequent lc-* commands refresh the installed skill, rules, and templates.

Tools and capabilities

The MCP workflow exposes these documented tools:

• lc_missing: identifies files not included in the available context.
• lc_changed: checks for files changed since the pack timestamp.
• lc_outlines: provides structural information about code.
• lc_preview: reports selected files, context size, and related definitions found through the code graph.

The related CLI includes lc-select, lc-context, lc-preview, and lc-init. lc-context can copy a pack to the clipboard or write it to stdout when requested. lc-preview is useful for checking whether a rule omitted a definition required by the selected files.

The cyberchitta/llm-context.py MCP server is best suited to local repositories where an agent needs focused, inspectable context rather than an indiscriminate dump of the project. It does not replace repository access: packs are partial, and the generated listing can omit files excluded by repository filters or the rule.