Graphite

A private, modular node based AI workspace for thinking and creation. Seamlessly connect to Ollama, OpenAI, Gemini, Anthropic, and Groq models in one unified, n…

LLM Mart 7 views 338k listing impressions

A local-first, graph-based AI workspace for branching reasoning, tool use, and multi-provider workflows.


Graphlink replaces the linear chat window with a visual canvas of connected nodes. Instead of forcing every interaction into a single timeline, it lets you branch: conversations, code generation, web research, drafting, and execution each live as their own node, and every branch can follow its own line of inquiry with its own model and context.

It is built with a Python (FastAPI) backend and a Vite/React/TypeScript single-page app, launched as one native desktop window via pywebview — not a browser tab, not Qt. It runs entirely on your machine, and works with local model runtimes (Ollama, llama.cpp) as well as hosted APIs (OpenAI-compatible, Anthropic Claude, Google Gemini).

Graphlink is the second generation of the Graphite project, renamed to avoid collision with unrelated software. The rename is complete: modules, folders, and the UI all use the graphlink name.

Demo

Watch the Graphlink demo

Table of Contents

Features

  • Visual branching canvas — build parallel thought paths, experiments, and delivery tracks in one view instead of one scrolling thread.
  • The Builder — give it a goal and it plans a checklist, then constructs it on your canvas one supervised step at a time, under hard step/token/time budgets (see The Builder).
  • Multiple model backends — run locally with Ollama or direct GGUF loading via llama-cpp-python, or switch to API Endpoint mode for OpenAI-compatible, Anthropic Claude, or Google Gemini. The active mode is switchable in Settings.
  • Per-branch model routing — pin a specific model to a node or a whole branch, on top of the per-task model defaults.
  • Plugin nodes — attach specialist nodes for web research, code execution, drafting, and repository-aware changes (see Plugins), plus a plugin SDK with per-plugin capability grants and optional MCP server integration.
  • The workspace agent — point it at a folder and give it a task; it reads and writes files there, runs shell commands and Python, and asks before anything that changes your machine. Bounded by a per-task turn budget, and it reads an AGENTS.md if the workspace has one.
  • Repository-aware editing — Gitlink loads a GitHub repo into structured context, previews file-level changes, and only writes after explicit approval.
  • Knowledge base and search — ingest documents into a local knowledge store, search it from a node, and search across every workspace at once.
  • Workspaces and library — organize graphs into workspaces with favorites, tags, and archiving; reopen any of them from the Library.
  • Undo that understands agents — full undo/redo over canvas history, including "undo this build", which reverts everything one agent run did as a single action.
  • Charts — generate a chart from any node's content and export it as PNG or SVG.
  • Attachments — stage images, audio, or documents onto a message; the backend classifies and extracts them.
  • Themes and canvas controls — light and dark themes, plus grid, connection routing, node font, and pan-sensitivity controls.
  • Local-first persistence — conversations, notes, navigation pins, and graph layout are stored locally in SQLite, with crash recovery.
  • Diagnostics — a token and cost counter, a command palette, and an exportable diagnostic bundle for troubleshooting.
  • Export — save the whole canvas

From the project's README.

Comments (0)

Sign in to join the conversation.

No comments yet.