carnaby/patchbay

Patch any data into a live dashboard. Your AI agent does the wiring.
Public APIs and your own CSVs, processed by declarative rules, rendered as real-time dashboards,
and built either by hand or by an AI agent over the Model Context Protocol.
### Try it live → patchbay.me.uk
No sign-up needed to build a dashboard. To let your agent build one, add https://patchbay.me.uk/mcp as a
connector (how).
Who it is for: anyone who wants something between heavy BI and monitoring tools (Grafana, Metabase) and static
homelab start pages (Homepage, Dashy). Patchbay is a declarative data pipeline with the speed of a notebook, and an AI
agent's work ends as a live, shareable dashboard instead of a chat answer. The agent checks its own work by looking
at a server-rendered snapshot of the dashboard.
> In a recording studio, a patchbay is the panel where you wire any source into any device with short cables.
> This is that panel for data: sources on one side, widgets on the other, and processing rules in between.
## Why it exists
Most dashboard tools assume a human clicks every chart into place. Patchbay is AI-first: its primary
interface is an MCP server that lets an agent such as Claude create dashboards, inspect your data, propose
how to process it, apply the changes, and then look at a rendered image of the result to fix what it got
wrong. Everything the agent can do, a person can do in the editor as well, with undo/redo.
It is free, needs no API keys, and runs as a single process you can host anywhere.
## Highlights
🤖 MCP-native. 15 tools, 2 prompts and a reference resource. OAuth 2.1 (PKCE, dynamic registration, Client ID
Metadata Documents) for real clients, personal access tokens for scripts and development.
👀 Agents can see. rendersnapshot returns a PNG of the dashboard, rendered server-side with the same chart
code the browser uses, so the agent can check its own work.
⚡ Real time. Datasets refresh on the server while someone is watching and are pushed over Server-Sent Events,
only when the data actually changed. "Current value" sources can accumulate into live time series.
🧪 Declarative processing. Filter, derive, group/aggregate, time buckets, rolling windows and more, with a small
sandboxed expression language (no eval, ever). The same pipeline runs for the browser, the agent and the renderer,
in worker threads when the data is large.
↩️ Undo everything. Undo/redo in the editor, a version for every save, one-click restore. When an agent changes
a dashboard you have open, the change appears live, highlighted, with an Undo button.
🏠 Wall screens and start pages. Maps, clocks, news feeds and a search box next to the charts. Presentation mode
(press F, or /d/ ?present=1) runs unattended on a TV; set a dashboard as your browser's home page and ask
your agent to add whatever you find.
More features
🧷 Concurrent editing. Pipeline steps have stable ids, so an agent and a person can edit different steps of the
same pipeline at the same time; unsaved edits are replayed on top of an agent's newer version.
💡 Suggestions without an LLM. A rules engine profiles any dataset (time, measures, dimensions, coordinates…)
and proposes fitting widgets with a one-line rationale. Agents get the same suggestions as ready-to-apply operations.