Skip to content

llms-install.md

Latest commit

 

History

History
129 lines (96 loc) · 4.13 KB

llms-install.md

File metadata and controls

129 lines (96 loc) · 4.13 KB

codesteward-mcp

Codesteward MCP server — parses a source repository into a structural code graph and exposes it as queryable tools. Answers questions like "what calls function X?", "is route Y protected by auth?", "what are the external dependencies of file Z?". Uses GraphQLite (embedded SQLite) for persistent graph storage — no external database needed.

I want you to install codesteward-mcp for me. Execute all the steps below autonomously.

OBJECTIVE

Install the codesteward-mcp MCP server and add it to the user's MCP configuration so that the AI agent can query the structural code graph of any repository.

DONE WHEN

  • uvx is available (or codesteward-mcp is installed via uv pip install)
  • The codesteward-graph MCP server entry appears in the active MCP server list
  • The agent can call graph_status successfully

TODO

  • Check that uv is installed
  • Ask the user for the path to the repository they want to analyse
  • Add the MCP server entry to the appropriate config file
  • Verify the server starts by calling graph_status

Steps

1. Check prerequisites

Ensure uv is installed:

uv --version

If not installed, direct the user to https://docs.astral.sh/uv/getting-started/installation/ or run:

curl -LsSf https://astral.sh/uv/install.sh | sh

2. Ask the user one question

Ask the user:

  1. Repository path — the absolute path on their machine to the repository they want to analyse (e.g. /Users/alice/projects/my-app). This becomes DEFAULT_REPO_PATH.

3. Add to MCP config

Add this entry to the MCP settings file. GraphQLite is the default backend — an embedded SQLite graph database that persists to ~/.codesteward/graph.db. No external database server needed.

For Cline (cline_mcp_settings.json):

{
  "mcpServers": {
    "codesteward-graph": {
      "command": "uvx",
      "args": [
        "--from", "codesteward-mcp[graph-all,graphqlite]",
        "codesteward-mcp", "--transport", "stdio"
      ],
      "env": {
        "DEFAULT_REPO_PATH": "<absolute-path-to-repo>",
        "DEFAULT_REPO_ID": "<short-repo-name>"
      }
    }
  }
}

For Claude Code (~/.claude/settings.json):

{
  "mcpServers": {
    "codesteward-graph": {
      "command": "uvx",
      "args": [
        "--from", "codesteward-mcp[graph-all,graphqlite]",
        "codesteward-mcp", "--transport", "stdio"
      ],
      "env": {
        "DEFAULT_REPO_PATH": "<absolute-path-to-repo>",
        "DEFAULT_REPO_ID": "<short-repo-name>"
      }
    }
  }
}

For Cursor (~/.cursor/mcp.json):

{
  "mcpServers": {
    "codesteward-graph": {
      "command": "uvx",
      "args": [
        "--from", "codesteward-mcp[graph-all,graphqlite]",
        "codesteward-mcp", "--transport", "stdio"
      ],
      "env": {
        "DEFAULT_REPO_PATH": "<absolute-path-to-repo>",
        "DEFAULT_REPO_ID": "<short-repo-name>"
      }
    }
  }
}

Replace <absolute-path-to-repo> with the path from step 2 and <short-repo-name> with a short identifier (e.g. my-app).

graph-all installs tree-sitter grammars for all 14 supported languages (Python, TypeScript, JavaScript, Java, Go, Rust, C#, Kotlin, Scala, PHP, C, C++, COBOL). Use graph instead to install only the four core languages (Python, TypeScript, JavaScript, Java).

4. Verify

After saving the config, call graph_status. You should see a YAML response with backend_connected and last_build fields. If last_build is null, call graph_rebuild to parse the repository — this may take 10–60 seconds depending on repo size.

Available tools

Tool What it does
graph_rebuild Parse the repository and build the structural graph
codebase_graph_query Query the graph — lexical, referential, semantic, dependency, or raw cypher
graph_augment Add agent-inferred relationships (confidence < 1.0) to the graph
graph_status Check graph state: node/edge counts, last build time, backend connectivity

EXECUTE NOW

Start with step 1. Work through each step autonomously, pausing only to ask the user the question in step 2.