Skip to content
dovvnloadingPublic

About

A local-first, graph-based AI workspace. Branch any conversation across a visual canvas instead of one scrolling thread, attach nodes that research the web, run Python, and chart results, then hand multi-step work to a budgeted, fully undoable agent. Works with local (Ollama, llama.cpp) and Cloud hosted models like GPT

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

36 stars

Watchers

1 watching

Forks

Latest commit

 

History

639 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Graphlink

A local-first desktop workspace for branching AI conversations. Ask a question, fork a reply into two lines of inquiry, then bring their answers together. Each branch keeps its own conversation history. Notes, sources, code, charts, and model replies sit next to the question they came from.

Install Graphlink · Connect a model · Build your first graph · API providers · Developer guide

One incident question branches into three replies suggesting a database investigation, a response-path check, and a rollback.

In this example, a deploy pushed p99 latency from 180 ms to 1.4 s. Follow the database branch to see the query plan and the charts that verified the fix. View the full incident canvas.

Getting started

Graphlink opens a native desktop window backed by a local server. For a source install, use Python 3.12 (the version used for the dependency lock), Node.js 24, npm, and Git. Windows uses the WebView2 Runtime. Linux needs GTK, WebKitGTK, and PyGObject. macOS uses Cocoa and WebKit through PyObjC. Native desktop setup by OS.

Clone the repository on any platform.

git clone https://github.com/dovvnloading/Graphlink.git
cd Graphlink

Windows (PowerShell)

py -3.12 -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r requirements.txt
npm --prefix web_ui ci
npm --prefix web_ui run build
.\.venv\Scripts\python.exe graphlink_desktop.py

Linux or macOS

Prepare a Python environment with the native desktop bindings for your OS. The Linux instructions include the GTK packages and the venv option that exposes them to Python. Platform setup.

From the repository root with that environment active, build and launch the app.

python -m pip install -r requirements.txt
npm --prefix web_ui ci
npm --prefix web_ui run build
python graphlink_desktop.py

On first launch, choose Load Sample Workspace in the welcome wizard to explore a small graph without a model or API key. To send your own prompts, connect a model below. To reopen the app, run the last command again. Rebuild the frontend after updating its source.

Connect a model

Where the model runs How to get your first reply
Ollama on your machine Start Ollama and pull a chat model with ollama pull MODEL, replacing MODEL with an Ollama model tag. Open Settings → Ollama (Local) → System Scan. Leave Chat Model on Auto to use a compatible installed model, or pick one by name. Ollama is the default on a fresh install.
Llama.cpp on your machine In Graphlink’s Python environment, install llama-cpp-python with python -m pip install llama-cpp-python. Open Settings → Llama.cpp (Local), choose a GGUF under Chat Model File, then Save Settings. Choose Use This Provider if it is inactive.
Hosted API or a custom endpoint Open Settings → API Endpoint. Pick a provider, supply its credential and any requested Base URL, then choose Load available models or type the exact ID into Default text model. Select Save and activate.

The composer shows the active route. Use its model picker to switch models on that connection. Under Per-task models in API Endpoint settings, you can assign a different model for naming, charts, or web research. To pin the currently selected model to a branch, open the branch root’s chat card menu and choose Pin to Current Model. Switch back to that provider before continuing a cloud-pinned branch.

Build your first graph

  1. On an empty canvas choose Write a message. Ask “Search got slower after a deploy. What should I check first?” Your question and the reply appear as connected cards.
  2. Right-click the reply’s header and choose Branch from Here. Ask “How would I test whether a database query regressed?” Return to the same original reply, branch again, and ask “How would I test the application layer?” The two follow-ups now share an ancestor. Each can continue independently.
  3. Click the header of the first branch’s answer, then Ctrl-click the second (Cmd-click on macOS). Open the command palette with Ctrl+K or Cmd+K, choose Synthesize Branches, and send “Give me one test that distinguishes these explanations.” The resulting card records both source branches.
  4. Use Fit All to see the whole graph. Save keeps it in the Chat Library. To share an editable workspace, choose Export Workspace… in the Library. To share a snapshot, use Export Canvas as PNG in the command palette.

Connect text to a note

To count the words in a draft, add a plain note using Add Note in the command palette. Open Text input on that note, choose the Drafter document as the Text source, select Word count, and choose Connect. Run text input writes the count into the note. After you edit the draft, the note marks its input as changed and keeps the previous count. Run it again when you want a fresh count.

You can also use a chat message or another plain note as the source. To copy its text exactly, drag its right-hand text port to the note’s left-hand port and choose Run text input. You can connect one source to several notes. Each destination runs separately. Disconnecting keeps its last result, and saving the graph keeps the connections for the next session. View a connected word-count note.

Build a chart and report from the same table

Choose Add a table on an empty canvas or Add Table in the command palette. To use an existing CSV, open Edit CSV, choose Import CSV file, then Save table. The starter table records 2 hours for Design, 4 for Engineering, and another 3 for Design. Choose Team under Group by and Hours under Sum column. Create chart and Create report produce a bar chart and an editable Drafter document from the same grouped totals. Choose Fit All to see all three cards.

In Edit table, change the first Design entry from 2 to 7 and choose Save table. Each output marks its input as changed and keeps its previous result. Choose Rerun from table on the chart to update Design to 10 while the report still shows 5. Rerun the report when ready. Compare versions lets you inspect the previous result and restore it. On a current result, Inspect source rows shows the original data row numbers and cell values behind each group. For Design, rows 1 and 3 contribute 7 and 3. Saving and reopening the graph preserves these versions and their source connections. View the chart and report after a table edit.

Run connected outputs together

To add a shared intermediate step, choose Create grouped table on the source, then create the chart and report from the grouped table. After editing the original table, open Plugins → Graph Runner, choose Whole graph, and select Preview run. The preview shows the order of the steps and which outputs need updating. The chart and report still show their old totals until you choose Run graph. If a report feeds a word-count note, the runner uses the updated report to calculate the count. Choose Selected nodes to limit the run to a branch and the inputs it needs.

Open Graph Runner again to inspect the saved result. If the app closed during execution, reopen the graph from the Chat Library and choose Resume run. Each completed step and its output were saved together. All transforms run locally. Existing chat replies can supply text to a workflow. See the completed three-step run.

To run a local AI model explicitly, start Ollama and install a chat model. In Plugins → Graph Runner, expand Run a local model into a note. The existing one-shot and saved model-to-note options remain available. For a multi-step workflow, choose a source and model ID, then Create Model node. You can connect an upstream card's named text-output handle to a Model node's text-input handle, or create a second Model from the first one's output. Connect a downstream plain note to the final Model output to copy or count its words. Preview run shows which nodes will run and which are waiting on a real model reply—preview does not call Ollama. Run graph explicitly executes and checkpoints each step in dependency order, without automatically retrying uncertain remote requests. Changed sources mark model outputs stale but preserve earlier replies. This first-class Model slice is local Ollama only; remote/cloud provider profiles, paid approvals, write-capable tools and generalized parallel execution remain follow-on work. Model-node workflow guide · One-shot model guide.

Scientific Workbench

Enable Scientific Core and a domain pack in Settings → Plugins. Open Plugins → Scientific Workbench to create a calculation node. For example, import a numeric CSV, branch a Statistics node from it, and select a measured column to compute its mean and sample standard deviation. Each result retains its inputs, method version, and source fingerprint and can be exported as JSON or CSV. Pack specifications and worked example.

API providers

The directory includes 26 named providers and a Custom endpoint for OpenAI Chat Completions-compatible servers. Search by service or model family, such as Claude, Grok, Kimi, or Qwen. You can browse another provider’s settings while your current connection stays active.

Directory group Providers
Model labs OpenAI, Anthropic Claude, Google Gemini, xAI, Meta, Mistral AI, DeepSeek, Cohere, Moonshot AI, Z.AI, MiniMax, Baidu Qianfan, ByteDance ModelArk, Tencent Hunyuan
Cloud platforms Alibaba Cloud Model Studio, Amazon Bedrock, Azure OpenAI, Google Vertex AI, NVIDIA, Groq, Cerebras, Together AI, Fireworks AI
Model gateways OpenRouter, Hugging Face Inference Providers, Perplexity

If Load available models is unavailable or your deployment is missing from the list, enter its ID directly. Model discovery, tool calling, and media support vary by provider and model. The provider’s Setup guide links to its API documentation.

Cloud deployment details and provider directory screenshot
  • Azure OpenAI takes your resource URL and deployment name.
  • Google Vertex AI takes your project and location endpoint and a Google Cloud OAuth access token. Replace the token when it expires.
  • Amazon Bedrock takes a regional endpoint and a Bedrock API key. For Runtime, use https://bedrock-runtime.<region>.amazonaws.com/openai/v1 and enter a model or inference-profile ID. For Mantle, use https://bedrock-mantle.<region>.api.aws/v1 for catalog discovery. Some models use the /openai/v1 path on the same host.
  • Custom endpoint accepts a local server without an API key on loopback. For remote services, enter a key for that URL.

Graphlink’s provider picker filtered to cloud platforms in API Endpoint settings.

Tools on the canvas

When you want to… Use
Research a claim or draft a document Web Research gathers sources on a branch. Artifact / Drafter opens a Markdown document you can revise alongside the conversation. You can also search documents you’ve ingested into the local knowledge base.
Work with a repository Review Lens reviews a GitHub pull request and links findings to lines in the diff. Gitlink previews proposed file changes and lets you apply selected files to a local checkout.
Run code or preview a page Virtual Environment Runner runs Python with per-node dependencies and your user account’s permissions. HTML Renderer opens a separate source and preview surface.
Generate media Fal Media Studio browses Fal’s live catalog from a canvas node. Configure Settings → Fal.ai, review the price, and confirm each paid generation. The node tracks the job and shows its results. Studio guide · Fal API and agent workflows
Give one branch its own instructions Add a System Prompt from Plugins, or use a Conversation Node to keep a linear exchange inside one card.

The Builder and the workspace agent

Open Builder to give the graph a goal. Choose Plan the build, edit the checklist on its plan card, then choose Start build. Co-pilot asks before edits to the canvas. Autopilot can make ordinary canvas edits on its own within your step, token, and time budgets. Python execution and web fetches still require approval. A paused or interrupted build can resume from its plan card. Stop ends the run. View the Builder and a recipe.

Open Agent for tasks that need files. It starts in a private scratch folder, or you can choose an existing project with Choose folder…. The agent can read files, propose edits, and request approval to run shell commands or Python. Its activity appears on the canvas.

Local data and privacy

Graphs and workspaces live in ~/.graphlink/chats.db. The knowledge base has its own database at ~/.graphlink/knowledge/knowledge.db, and stored images live under ~/.graphlink/assets/. Local models can answer offline once installed. Hosted inference sends conversation context to the selected provider. Web research, GitHub integrations, and Fal media jobs also contact their respective services.

Saved keys and GitHub tokens live in ~/.graphlink/session.dat. Graphlink uses account-bound Windows DPAPI when available. On macOS, Linux, or when DPAPI fails, saved credentials are plaintext. For providers that support them, Settings shows the environment variables you can use in place of a saved key. Asset storage and cleanup · Security policy.

Development

The desktop launcher starts FastAPI on loopback and opens the native webview. The Python backend owns the graph and persists edits. The React and TypeScript frontend renders the canvas and sends actions over a WebSocket. Shared types are generated from Python contracts. Backend · Frontend · Wire contracts · Plugin examples.

For automation, run python -m backend.headless --port 8765 and connect the local MCP stdio adapter. Inspection is read-only by default. Canvas edits require explicit opt-in and a separate write token. The desktop and headless launchers share the same single-instance lock for each data directory. Headless and MCP setup.

Contributor setup and checks · Codebase map · MIT license

About

A local-first, graph-based AI workspace. Branch any conversation across a visual canvas instead of one scrolling thread, attach nodes that research the web, run Python, and chart results, then hand multi-step work to a budgeted, fully undoable agent. Works with local (Ollama, llama.cpp) and Cloud hosted models like GPT

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

36 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages