Troubleshooting
Find whether a failure comes from transport, authentication, permissions, or the workflow.
The client cannot connect
Confirm the URL points to the MCP endpoint, including /mcp, and that the client supports Streamable HTTP. For a self-hosted service, check /ready and protected-resource metadata. An ordinary browser request to /mcp is not a complete MCP connection test.
Authentication fails
Reconnect through the client's OAuth flow. For a self-hosted deployment, compare the resource URL, issuer, and JWKS settings. Do not substitute an Innflow REST API key or the server signing key for an OAuth access token.
Reads work but execution or deployment fails
Check the operation's required scope and the user's resource permissions. Running a workflow requires an eligible deployed workflow. Deployment and undeployment use management permissions. Inspect the returned error rather than assuming every failure is an authentication issue.
Authoring tools are missing
Authoring is disabled by default. It requires a compatible application and MCP server with the authoring gate enabled. Refresh tool discovery after an operator changes the deployment configuration.
A workflow node cannot find its server
Verify the saved server definition is accessible to the workflow's user and workspace. If the connection has expired, reconnect it. Shared definitions do not share another user's access token.
Agentic mode has no usable tools
Check whether the server returned any tools, whether your tool filter excluded them, and whether tools advertise readOnlyHint: true. Agentic mode excludes write tools by default. Do not enable writes merely to bypass a discovery problem.
A run or write times out
Inspect the execution or operation before retrying. A timeout does not prove that the remote action did not happen. Retain the original idempotency key when retrying the same intended write, and use a new key only for a new intended action.
A tool result is too large
Use summaries, pagination, selected nodes, and targeted execution payloads. Retrieve the smallest result needed to diagnose the issue instead of repeatedly requesting a complete graph or all logs.