Graphicle – A local, Rust‑based code‑graph & memory engine for AI‑powered coding assistants

What it is

Graphicle is a single‑binary, open‑source tool written in Rust that builds a code graph of a repository (files, functions, classes, calls, imports) and stores project‑wide documentation, decisions and past coding‑assistant sessions. The data live in one SQLite file on the developer’s machine, so there is no external database, Docker container, or cloud service.

Why it matters for AI‑assisted development

• Cited context on‑demand – When a coding agent (Claude Code, Cursor, Codex, VS Code, etc.) asks a question, Graphicle returns the exact file‑and‑line excerpts that answer it, staying inside a token budget you set. The model that powers the agent reads those excerpts, so the answer can be verified.
• Local‑first & private – No API keys, accounts or telemetry are required. All indexing, searching and memory lives on the developer’s computer; only optional model calls (e.g., to Anthropic, OpenAI, Ollama) go out.
• Speed & memory efficiency – Benchmarks in the repo show the 0.1.1 pre‑release engine runs 3 ×–22 × faster and uses 0.3 ×–0.97 × less peak memory than the comparable Python tool Graphify on 13 real‑world codebases.
• One‑stop installation – A single brew install (or script) gives you the code‑graph builder, a memory server, an MCP (Multi‑Client Protocol) server, a web UI and a CLI. No Python, Docker, or separate database is needed.

Core features (as described in the README)

Feature	What it does
Code graph	Parses source with tree‑sitter, records files, symbols, calls, imports; provides CLI (graphicle graph …) and an interactive HTML view.
Project memory	Indexes Markdown, PDF, Office, HTML docs and past Claude Code / Codex sessions; memories are linked to the symbols they describe and become stale when the code changes.
Cited context	graphicle ask / graphicle context return excerpts with file‑name and line‑range citations, respecting a token budget, without invoking a model.
MCP server	Exposes the graph and memory tools over the MCP protocol so any compatible client (Claude Code, Cursor, VS Code, etc.) can call them.
Session capture	Records each completed turn of a coding assistant; reviewed memories are stored and later automatically supplied to new sessions.
Git integration	Installs Git hooks that sync changed files before each answer and backs up every file it modifies (*.graphicle-bak).
Cross‑repo workspace	graphicle context --workspace can span multiple repositories in a workspace.
Web UI	graphicle open launches a browser view showing the graph, source files and memories, where you can approve memories.
Privacy controls	Respects .gitignore, skips secrets, redacts private spans, and only makes a daily release‑check request (opt‑out via GRAPHICLE_NO_UPDATE_NOTIFIER).

Who would use it?

• Developers who rely on LLM‑based coding assistants and want deterministic, verifiable citations for the assistant’s suggestions.
• Teams that need a shared, local knowledge base (docs, decisions, prior assistant sessions) tied to the exact version of the code they are working on.
• Tool builders who want a fast, Rust‑native code‑graph backend that can be accessed via MCP from any editor or IDE.

How to get started (quick‑start from the README)

1. Install – brew install graphicle-ai/tap/graphicle (or the one‑line curl script).
2. Index a repo – cd your-repo && graphicle setup. This builds the graph, captures the current session, adds a short entry to AGENTS.md, and installs Git hooks.
3. Ask a question – graphicle ask "where is tax added to prices?" – you’ll see cited excerpts; configure a model (graphicle config set model anthropic) to get a full answer.
4. Explore – graphicle open (browser UI), graphicle graph extract . --code-only, graphicle graph affected "retry()".
5. Undo – graphicle disconnect removes hooks and restores backed‑up files.

Limitations & current status (as of the README)

• The Windows build is a preview and not code‑signed yet.
• Benchmarks are based on pre‑release builds (0.1.0/0.1.1); the engine is still evolving.
• Token‑budget handling only limits the amount of cited context; the quality of the final answer still depends on the downstream LLM you configure.
• Long‑running sessions and large memory accumulation are not yet fully tested.

---

All information above is taken directly from the repository’s README; no external assumptions have been added.