Connecting an AI assistant
Rinkt exposes a Model Context Protocol server, so an AI assistant can build, run and debug your workflows directly rather than you describing them to it.
Connecting
claude mcp add --transport sse rinkt-workflows --scope user \
https://<your-rinkt-host>/api/mcp/sse
The first connection opens an OAuth flow in the browser against your Rinkt account. For a
headless client, an API key prefixed lov_-style is sent in an Authorization: Bearer
header instead.
The same endpoint works for Claude Desktop, Cursor and any other MCP client that supports SSE transport.
What it can do
The assistant works in three loops.
Build and edit
It reads the DSL and expression references, browses the activity catalogue, then writes or
edits the workflow — a single activity, a block wrapped in TryCatch, or a whole new
workflow. Every save snapshots the previous state, so a change can be rolled back after
previewing the diff.
Run and debug
It starts a run, then reads the activity-level logs, the first failing activity, the values an activity actually ran with, and the page as it was when the run broke. It can step through with breakpoints, or replay a run from the input that produced it.
Drive a robot
With an online robot it can open a page, read its accessibility tree, verify selectors before acting, then click, type and wait — and on the desktop side, resolve an xPath or match an element by text or image. Sessions last two hours.
Tool categories
| Category | What it covers |
|---|---|
| Workflow | List, search, read, create, update, delete; version history and rollback |
| Editing | Edit one activity, insert steps, wrap steps in control flow, move, delete |
| Execution | Run, wait, stop, replay, signal; logs, errors, screenshots, recordings, breakpoints |
| Discovery | Activity catalogue and schemas, DSL and expression references, templates, datasets |
| Web inspection | Robot sessions, page analysis, selector validation, click, type, wait, viewport |
| Desktop | xPath resolution, OCR and image matching, click, type, screenshots |
| Project | Project info, environments, resources, reports and widgets |
Worth knowing before you rely on it
A session cannot be introspected. Once a robot session exists, there is no way to ask it what page it is on. The assistant has to track that itself.
One step at a time. A robot session executes a single step at a time; a second step issued while one is running fails with a lock error. Resolving a desktop xPath is the exception — it reads the accessibility tree over a separate channel and takes no lock.
Desktop targeting and desktop tooling disagree. Targeting is by xPath, but most of the desktop tools match text or template images, so they cannot tell you anything about a selector. Use the xPath check for that.
Reading order
An assistant that reads the DSL reference once at the start of a session, and the
expression reference before writing any {{ ... }}, produces workflows that need far less
correcting. Both are available as tools, and as pages here:
Writing workflows and Expressions.