Summary
Linkr ships an MCP server that lets any agent — Claude Code, OpenCode, Cursor — build Linkr content as files, with no running instance. Every write is validated against the real format, so the tree it produces imports as-is. It is the offline counterpart of the built-in assistant.
Two servers, two targets
MCP is a protocol through which an agent is handed tools. Linkr plans two of them, and the difference lies in what they act on.
Available today
linkr-authoring — files, offline
The agent writes a project tree to disk. No Linkr instance is needed, no data is touched. That is what this page covers.
Coming with the agents
linkr-live — a running instance
The agent acts on the project you have open: add a tab, build a cohort. That is what the Agents page describes.
These are not two versions of one thing. One needs a path on disk, the other a server and an open session: merging them would give a tool nobody could configure.
The name in your configuration
The server still announces itself as linkr. It will take the name linkr-authoring when linkr-live arrives, so that both are distinguishable in an agent using the two of them.
What linkr-authoring can do
An agent handed linkr-authoring can build a complete project: its metadata, its datasets from CSV, its dashboards with tabs and widgets, its analysis scripts. It can also read back an existing tree and change it — rename a widget, adjust a filter, add a tab.
Two guarantees make this usable.
- Everything is validated against the real format. The server holds no format knowledge of its own: it delegates to the same library the application uses. A tree it accepts therefore imports without surprises.
- You must read before writing. A dedicated tool describes what a tree contains, with its true identifiers. This matters because many keys are derived — a tab’s key comes from its name, a widget’s from its position. Editing those files by hand breaks references; going through the tools recomputes them.
Do not edit these files by hand
An agent editing the JSON files directly would produce a tree that looks right but whose references are broken. The tools exist precisely for this: they recompute what needs recomputing and refuse what does not hold together.
What it is for
Three uses stand out.
Preparing demonstration or training content — an example project with its synthetic data, its dashboard and its scripts, built in one pass rather than clicked screen by screen.
Feeding a portal. Linkr’s public content is itself made of trees of this kind: producing them through an agent, with validation, avoids the class of errors you only discover at import time.
Working offline. Since nothing requires a running instance, you can prepare a project from any machine, with no access to the server or to the data.
How to use it
linkr-authoring is added to your agent like any other MCP server — a line of configuration naming the program to launch. The exact instructions live with the package, which is authoritative on that point.
Once registered, the tools show up in your agent and are used in plain language: “build me a demo project on intensive care, with a synthetic dataset and a mortality dashboard”.
Going further
- Agents — the assistant that will act on a running instance instead.
- Entities and sharing — the tree shape these tools produce.
- Import and export — bringing an offline-built tree into an instance.