Show HN: Graphicle, a Rust port of Graphify with memory for coding agents

What it claims

Graphicle is a single Rust binary that builds a code graph and a project memory over a repository, then serves cited excerpts to coding agents over MCP. The code-graph engine is a port of Graphify that tracks Graphify's releases. The MCP server exposes Graphify's tool names (query_graph, get_neighbors, shortest_path) alongside memory tools, so an agent already wired to Graphify keeps working. The retrieval tools call no model; the agent's own model reads the excerpts.

Evidence

• Kubernetes (31k files), code-only extract: 3 min 18 s for the 0.1.0 engine versus 26 min 38 s for Graphify 0.9.88. Both built the same graph. This was one exploratory run each, on one Linux host with 16 CPUs and 8 workers.
• Django, code-only extract: 7.5 s for a pre-release build versus 72.8 s for Graphify 0.9.77. These are medians of five alternating runs on a shared Linux host with 8 workers.
• Memory: compared against seven memory systems. The comparison was pre-registered, and one open model served as reader and judge for every system. A pre-release build "trails none" on three development sets. That means a tie at the top is possible, and these are development sets, not held-out sets. Per-system scores are on a linked page, not in the README.

How to use it

brew install graphicle-ai/tap/graphicle
cd your-repo
graphicle setup

Without Homebrew:

curl -fsSL https://github.com/graphicle-ai/graphicle/releases/latest/download/install.sh | sh

graphicle setup makes several changes beyond indexing:

• connects every MCP client it finds
• enables session capture
• redirects Claude Code's searches to the code graph first
• appends a section to AGENTS.md
• installs Git hooks for commit, merge and checkout

Each changed file is backed up once as <file>.graphicle-bak, and graphicle disconnect reverts everything. graphicle setup --yes runs non-interactively.

Other useful commands:

• graphicle ask "…" returns cited excerpts within a token budget (default shown: 3,000 tokens).
• graphicle config set model anthropic (or openai, ollama) adds written answers to ask. Any OpenAI-compatible server also works.
• graphicle graph extract . --code-only then graphicle graph affected "retry()" lists dependents of a symbol.
• graphicle open launches the browser app, where proposed memories are approved.
• graphicle context --workspace searches across the repositories in a workspace.
• graphicle doctor runs diagnostics.

Storage is one SQLite store:

• macOS: ~/Library/Application Support/Graphicle
• Linux: ~/.local/share/graphicle
• graph extract also writes to .graphicle/ in the project.

The app, HTTP and MCP endpoints bind to 127.0.0.1. There are SDKs for Python and TypeScript, and a team-server recipe in deploy/. Deleting a memory or session also removes everything derived from it and issues a hash-chained deletion receipt.

Limits

• Network use: "no API key" holds only for code-only work. A full graphicle graph extract without --code-only sends documents and images to whichever LLM it finds an API key for. Use --code-only if that matters to you.
• Platforms: there is no native Windows build. WSL2 is expected to work but is untested.
• Size limits: defaults are 100,000 text files and 1 GiB of text per project; both can be raised. Very large files are only partly analysed or kept as plain searchable text.
• Staleness: a memory goes stale when the function it references changes. Edits elsewhere in the same file "usually" leave it current; the exact rule is documented separately.
• Session capture: covers Claude Code and Codex only. It drops <private>…</private> spans and redacts credentials.
• Maturity: both benchmarks used pre-release or test builds. The Kubernetes figure comes from a single run.
• Licence: Apache-2.0. Parts derived from Graphify remain partly MIT. A commercial team edition is in private preview. Contributions require a DCO sign-off; there is no CLA.