Pulse usage

You use Pulse by pointing an MCP-aware AI assistant at it and asking questions in natural language.

Connect your assistant

Point the assistant at your deployment's /mcp endpoint, or at the hosted playground to try it without any setup. The playground is public and therefore read-only: you can ask anything, but it never changes the engine.

Host	How to connect
claude.ai	Settings → Connectors → Add custom connector, URL https://pulse-playground.miragon.ai/mcp (needs a public HTTPS endpoint)
Claude Code	claude mcp add --transport http pulse https://pulse-playground.miragon.ai/mcp
Claude Desktop	stdio only: bridge the HTTP endpoint with mcp-remote (below)
Any other MCP host	add it as a streamable-HTTP server

Claude Desktop validates claude_desktop_config.json against the stdio shape, so a bare "url" entry is dropped silently. Open Settings → Developer → Edit Config, add the bridge, and restart the app:

{
  "mcpServers": {
    "pulse": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://pulse-playground.miragon.ai/mcp",
        "--transport",
        "http-only"
      ]
    }
  }
}

For a server on your own machine, use http://127.0.0.1:8400/mcp there: 127.0.0.1 rather than localhost, which Node may resolve to IPv6 while the server listens on IPv4.

Once connected, you'll see Camunda and analytics tools available in the assistant. Try asking "show me the open incidents". Claude calls camunda7_show_incidents_dashboard and renders the result in the incidents widget.

What you can ask

Topic	Example questions
Processes	"Show me all process definitions." · "Start a new instance of loan-approval."
Tasks	"What tasks are assigned to me?" · "Complete task 12345 with approved=true."
Incidents	"Triage the open incidents." · "Retry all failed jobs from this morning."
Analytics	"Where do loan-approval instances spend the most time?" · "Find failed instances from the last 24h."
Migrations	"Plan a migration from v1 to v2 of loan-approval."

Most data-heavy answers come back as an interactive widget: process lists, task inboxes, incident dashboards, instance details, and analytics dashboards all render inline. Inline a widget stays a compact card: a list shows its first rows and how many there are, and Show all opens it fullscreen where your chat app offers that.

What the assistant may change depends on the deployment's toolset. A read-only deployment (the playground, or any server without login unless its operator widened it) answers every question but offers no write tools and no action buttons. Starting instances, completing tasks, retries and resolving incidents need operations (the default with login); deletions, modifications, migrations and signals need admin, which an operator has to enable explicitly. Whether a write the assistant starts asks you first is up to your chat app's tool-approval settings; the server marks every destructive tool (destructiveHint) so the app can ask. A widget button that cancels, suspends or resolves asks first and names what it acts on (instance, business key, engine); after any action the view reloads what it changed: completing a task shows the next one. The cockpit's Refresh re-reads the open view, e.g. after the assistant changed something.

Your settings

Ask "open my settings" (or click Settings in the cockpit) for the full page: the language first, then one section per active module, so an assistant with only some modules connected shows only those sections.

Setting	Section	Effect
Language	Language	UI language (Automatic: the chat app's); a chosen one also sets the language tool summaries come back in
Engine availability	Engines and dashboards	Which engines appear in pickers, and the default the cockpit opens on
Pinned dashboards	Engines and dashboards	Which saved dashboards come first in pickers (needs login)
Look-back period	Analytics	Applied whenever you ask an analytics question without naming a window
Comparison bucket	Analytics	How many instances a window needs before a comparison is trusted

You can also open a single section ("open my engine settings" or "show my analytics settings") or change a value in passing, without any page: "switch the UI to German", "default my analytics to 30 days". Only the setting you name changes; the rest keeps its value.

Settings are saved per user account, so they need a deployment with login: signed in, they follow you across conversations. Without a login, and on a read-only deployment, the page shows the defaults (widgets follow the chat app's language) but hides Save. Light or dark is never a setting: the widgets always follow the chat app's theme, and your system's without one. Saved dashboards exist only with login on a deployment that allows writes; otherwise the dashboard picker says "Saved dashboards are unavailable".

Tips

• Ask follow-ups. Once a widget is on screen, you can drill down by clicking, or ask for "the same thing but for the last 7 days."
• Combine tools. "Find failed instances from yesterday, then retry the transient ones" works as a single sentence (the retry needs operations).
• Failed jobs and failed external tasks clear by a retry, not by marking them resolved: the incident views offer Retry for them and Mark as resolved only for custom incidents. Batch actions (batch retries, migrations) only queue work: ask for the batch status afterwards. Completing a task that was delegated in Tasklist hands it back to its owner, and the task stays open.
• The two health checks judge different things and say how: the engine overview rates one engine by its open incidents (by default critical at 50 open or 25 in one cluster), the analytics health by Prometheus alert rules (critical only while a critical alert fires; unknown when an engine sends no metrics). "Has this activity failed before?" is answered from incident history, resolved incidents included.
• A process view counts every deployed version of its key: old versions keep running after a redeploy; only the diagram is one version, and it says which. A number the engine could not confirm shows as "—" (or "≥N" when only part of it was read), never as 0, and an engine that does not answer is an error, not an empty view.
• If a tool isn't doing what you expect, ask the assistant to show you the raw arguments. It'll print them, and you can spot mismatches quickly.