Skip to content

cline.md

Latest commit

 

History

History
86 lines (65 loc) · 2.94 KB

cline.md

File metadata and controls

86 lines (65 loc) · 2.94 KB
Raw

Cline Integration

Cline is the largest open-source AI coding agent in VS Code (61K+ stars, 5M+ installs as of May 2026). It's powerful, but its full-file rewrite model is token-heavy: editing 10 lines in a 500-line file sends all 500 lines back. That cost compounds fast on real codebases — and the Cline community has been asking about token strategy in discussion #1539 for a long time.

engram fixes that without you changing how you use Cline. The MCP server intercepts file reads at the agent boundary and replaces them with structural summaries. Cline keeps doing what Cline does. The token bill drops.

Prerequisites

npm install -g engramx
cd ~/your-project
engram init .

That's it on engram's side. No extra configuration.

Add engram to Cline as an MCP server

Cline reads MCP server configuration from your VS Code settings. Open the Cline panel, click the gear icon, then "MCP Servers" → "Edit MCP Settings". Add:

{
  "mcpServers": {
    "engram": {
      "command": "engram-serve",
      "args": ["/absolute/path/to/your-project"]
    }
  }
}

Replace the path with your actual project root. Cline will now have access to the same six MCP tools Claude Code does:

  • query_graph — natural-language graph queries
  • god_nodes — most-connected entities in the codebase
  • graph_stats — high-level codebase summary
  • shortest_path — find connection between two concepts
  • benchmark — measure token reduction on this repo
  • list_mistakes — past failure modes engram has seen here

What changes in your sessions

Cline's agent will start using query_graph before reading large files. Where it would have read all 500 lines of auth/middleware.ts, it'll first ask the graph "how does auth work in this project," get a 200-token structural answer, and only fall back to the full file if the structural view isn't enough.

In practice, on a real codebase the average Cline session sends roughly 80% fewer tokens for context retrieval. Your raw API bill drops the same amount. Edit-and-fix cycles get faster because the model sees less noise.

Tracking it

engram cost -p ~/your-project

After 24 hours of normal Cline usage, the table fills in. After a week, you can run engram cost --digest to write a Markdown report you can paste into a chat or pin to a Slack channel.

A note on file-write operations

engram's interception is read-only. Cline's full-file write behavior is unchanged — engram doesn't try to slim down what Cline writes back, only what it reads in. Output-side optimization is a separate concern (see projects like Kilocode RTK for that angle).

Combine with: Claude Code, Cursor, Continue, Aider

If you alternate between Cline and another agent on the same repo, engram's graph is shared. The same .engram/graph.db powers all your tools — index once, save tokens everywhere.

See the integration index for setup in each.