Connecting clients
Wire Claude Code, Claude Desktop, or any MCP client to your workspace.
Both surfaces are standard streamable-HTTP MCP servers; anything that speaks MCP connects. Mint a token with the scopes you actually want the agent to hold — the token is the entire authorization story.
Connect
claude mcp add gorunner --transport http \
https://<your-workspace-url>/mcp/build \
--header "Authorization: Bearer grt_…"
# and/or your workflows-as-tools:
claude mcp add gorunner-tools --transport http \
https://<your-workspace-url>/mcp/tools \
--header "Authorization: Bearer grt_…"Choosing scopes per use case
| Use case | Scopes to mint |
|---|---|
| Agent may call your automations, nothing else | workflow:run |
| Agent may also read results | workflow:run, run:read |
| Agent builds and tests, human publishes | workflow:read, workflow:write, workflow:run, run:read, connection:read |
| Full authoring including publish | the above + workflow:publish |
Sizing advice
Connect /mcp/tools broadly — it's small, safe vocabulary. Connect /mcp/build where you actually want authoring; its 12 tools plus your other servers all draw from the same client tool budget.
First prompts to try
- "Using gorunner, build a workflow that receives a webhook, validates the payload with a code step, and emails me when the total is over 500. Validate it, run it in test, show me the run, then stop before publishing."
- "List my production runs from today and diagnose the most recent failure."
- With
/mcp/toolsconnected: just ask for the outcome — "look up order 9137's total" — and watch it pick your workflow.