Skip to content

llms-install.md

Latest commit

 

History

History
73 lines (55 loc) · 2.27 KB

llms-install.md

File metadata and controls

73 lines (55 loc) · 2.27 KB
Raw

Installing engram (guide for AI agents like Cline)

engram is a local code-graph context layer for AI coding agents. It indexes a repository into a local SQLite knowledge graph and exposes it over MCP (stdio), so the agent can ask about code structure, call sites, and git-mined mistakes instead of re-reading whole files. Apache-2.0, local-only, zero cloud, no API key, no account.

1. Install the CLI

npm install -g engramx

(Or skip the global install and run it on demand with npx — see the MCP config below, which uses npx.)

2. Index the project

cd /absolute/path/to/the/project
engram init .

This builds .engram/graph.db from the repo (AST extraction + git history). It runs entirely locally and takes a few seconds on a typical repo.

3. Register engram as an MCP server

Add this to the MCP settings (in Cline: MCP Servers → Configure, i.e. cline_mcp_settings.json):

{
  "mcpServers": {
    "engram": {
      "command": "npx",
      "args": ["-y", "-p", "engramx", "engram-serve", "/absolute/path/to/the/project"]
    }
  }
}

Replace /absolute/path/to/the/project with the absolute path to the repo you indexed in step 2.

4. Verify

engram doctor -p /absolute/path/to/the/project

It should report that the graph is present and the AST provider is reachable.

MCP tools you get

  • query_graph — natural-language search of the code graph
  • god_nodes — the most-connected entities (core abstractions)
  • graph_stats — node / edge counts and confidence breakdown
  • shortest_path — trace the connection between two symbols
  • benchmark — the structural context-token reduction on this repo (a reduction in tokens entering context, not a dollar/bill saving — your real cost is dominated by prompt caching)

Notes

  • Requirements: Node ≥ 20. For the grep call-site feature, ripgrep (rg) on PATH (optional — engram falls back gracefully without it).
  • Keeping the graph fresh: re-run engram init . after large changes, or engram reindex <file> for one file, or engram install-hook to update it automatically as files change.
  • Privacy: nothing leaves the machine — engram never makes a network call and stores everything in local SQLite under .engram/.