Coda

Connectome data analysis: a browser-based node editor for querying and analyzing connectomic data.

[!TIP] Beta. The core functionality is there, including support for (almost) all major connectome datasets. We're still adding features, improving the UI and fixing bugs, but from here on we aim to keep saved workflows opening in later builds. Bug reports and feature requests are very welcome!

Highlights

• node-based editor for building analysis/exploration pipelines
• works with neuPrint, CAVE, CATMAID and local data
• download/export results as CSV, SWC, images, Neuroglancer URLs, etc.
• convert workflows to Python or R code to run locally
• AI assistant for generating workflows from natural language queries
• MCP server so Claude, ChatGPT & co. can build workflows for you
• share workflows via links

Quickstart

Just go to https://coda.science/.

New to Coda? Watch the two-minute video Your first Coda workflow, which carries on into the rest of the tutorial playlist.

Other useful links:

• "Feature Overview"
• "Field Guide"
• "Node Guide"
• "MCP Server": drive Coda from your own AI client

Check out ? ▶ Guides for in-app tutorials, and the Coda Zoo for workflows shared by other users.

• add credentials to access neuPrint/CAVE/CATMAID datasets via Connections icon in the top toolbar
• New to create a new graph: from scratch, with the Workflow Wizard, from the Coda Zoo or from a preconfigured dataset
• click and drag to pan, scroll to zoom
• + button (lower right) to add nodes or Tab to search for nodes
• Space for the command palette
• "Run" button or ⇧R to run all stale nodes
• I to show/hide the inspector

Why not just use neuPrint, Codex or CATMAID directly?

Codex and neuPrint are great for initial exploration of the datasets but if you want to do more than just look at a few neurons, you quickly hit a wall and have to start writing code. CATMAID (the oldest of these tools) actually has a lot of analysis tools, but hasn't made the transition to modern (segmentation + meshes) connectomics.

Coda is designed to fill the gap between exploration-only and full-on coding. Also: because we don't play favourites, you can combine data from multiple sources in one workflow! And if you do want to write code, Coda can help you get started by generating Python or R code from your workflow.

Why "Coda"?

The neat acronym aside, a coda symbol in musical notation is a marker used for navigating through a complex piece - similar to what Coda does for connectomics data. Furthermore, the plural of coda (Italian for "tail") is code which is strangely fitting.

Development

pnpm install
pnpm dev          # http://localhost:5173

pnpm test         # vitest
pnpm typecheck
pnpm lint
pnpm build        # static bundle in dist/

Requires Node ≥ 20 and pnpm. On a fresh machine: brew install node && npm i -g pnpm. (Node 25+ no longer bundles corepack, so install pnpm directly.)

Layout

src/
├─ core/          graph engine — headless, no React (lint-enforced)
│  ├─ types.ts        socket type system + column schemas
│  ├─ node.ts         the NodeDefinition contract
│  ├─ graph.ts        document model, topo sort, serialisation
│  ├─ inference.ts    edit-time type/schema propagation + link validation
│  └─ scheduler.ts    DAG executor, provenance-keyed cache, hybrid eval
├─ data/          DataSource interface, one folder per backend, the mock connectome
├─ nodes/         built-in nodes (query / table / analysis / output …)
├─ packs/         node packs (Connectome, Cortex, ZapBench, …), switched on under Plugins
├─ help/          the in-app help documents (a pack keeps its own under packs/<id>/help)
├─ export/        Python notebook and R Markdown export
├─ store/         zustand document state, undo, persistence
├─ ui/            React Flow editor, param widgets, viewers
│  └─ panels/     command palette (+ fuzzy matcher), inspector, toolbar
└─ wizard/        the Workflow Wizard: graphs built from four answers, the starters, the demos

src/core and src/data must stay headless: a lint rule blocks imports of React, zustand, the store and the UI from those directories. The point is that the engine stays unit-testable without a DOM, and reusable later by a non-React consumer (a CLI runner, or a Python-side executor consuming the same graph JSON).

Known limitations / bugs

• Muting blocks downstream rather than passing input through Blender-style.
• No virtualisation. One DOM node per graph node, and table viewers cap rendered rows. Fine for tens of nodes; a 1000-node graph would need work. Explore Dataset pages rather than scrolling a long list for the same reason.
• A non-default neuPrint deployment needs CORS or the dev server. It is tried directly first; failing that, pnpm dev and pnpm preview proxy /np/<deployment>/… (https to public hosts only), and a static build has nothing serving that path.
• Not all nodes are currently able to emit Python or R code. The codegen is a work in progress.

Analytics

The published site counts page views with GoatCounter, and the dashboard is public — that link is the whole of what is collected, viewable by anybody.

No cookies, no localStorage, no tracker id, and no stored IP or full User-Agent: the beacon reports a page name, a referrer, and the coarse browser/OS/country/screen-width facts GoatCounter aggregates. There is deliberately no event tracking — nothing observes what you build on the canvas. To opt out entirely, run localStorage.setItem('skipgc', 't') in the console on any Coda page.

The tag is injected at build time and only when CODA_ANALYTICS is set, which happens in this repository's deploy workflow and nowhere else — so a fork you build and host reports nothing to anybody. See docs/analytics.md.

Contributors

• Jordan Matelsky (Kording lab/Mindspan) contributed the DotMotif node, which searches a network for a motif pattern locally in the browser.
• Casey Schneider-Mizell (Allen Institute for Brain Science) provided the inspiration for the Cortex Gallery node.

Licence

MIT