Graphicle, on the stand

Graphicle 0.1.0 is a single Rust program that indexes a repository into a code graph and hands a coding agent short, cited excerpts. The project is Apache-2.0, copyright 2026 KarLex AI, Inc. The README says the engine is a Rust port of Graphify, with memory, session capture, and an app that Graphify does not include.

It is local. The README says there is no account and no telemetry, and that nothing leaves the machine unless you ask. The code graph with --code-only, context, the MCP server, and the app work without an API key or a model. A model is required when you want graphicle ask to write an answer, or when a full extract is reading documents and images. Without a model, ask shows the evidence.

Setup is not a read-only glance. graphicle setup indexes the repo, connects the agents it finds, turns on session capture, edits AGENTS.md, and installs Git hooks. It lists the files it changes and backs them up. graphicle disconnect removes what it added. Clients named in the README: Claude Code, Cursor, Codex, VS Code, Claude Desktop, and any MCP client you configure by hand. There is no native Windows build. The README says the Linux install under WSL2 should work and that this has not been tested.

The speed line in the README is one exploratory run, theirs, on a Linux test build of the 0.1.0 release source: on Kubernetes, 31,000 files, a code-only extract finished in 3 minutes 18 seconds, against 26 minutes 38 seconds for Graphify 0.9.88, same graph, one Linux host, 16 CPUs, 8 workers. We did not repeat it.

The fitting. MODEL ran the v0.1.0 Linux binary, checksum verified, on a two-file Python repository.

graphicle setup --yes took 0.2 seconds. It listed every file it touched and reported that no agent client was installed on the machine.

graphicle graph extract . --code-only produced 6 nodes and 10 edges.

graphicle graph affected "price_with_tax()" returned checkout() in shop.py at line 5, handle() in api.py at line 3, and the import in api.py. Those callers were correct.

graphicle ask "how are failed requests retried?" returned both files, with SHA-256-tagged excerpts, inside a 3,000-token budget. It did not write an answer. No model was configured.

We did not test a live MCP session with an agent, and we did not test a large repository.

The verdict. Try it, on a side repository, if an agent is getting lost in a codebase that no longer fits in the window. Skip it for now on native Windows, on a project small enough that the agent already sees every file, or if you do not want a version 0.1.0 tool writing hooks into the repo you ship from.

The full note is in Issue Nº 01.