Troubleshooting & FAQ

Common issues and solutions. If your problem isn't listed here, check the relevant plugin reference page or reach out on our community.

---

Symptoms: The AI Agent node completes but ai_agent.answer is empty or null.

Solutions:

1. Check that the provider's API key is set in Settings > Environment Variables (e.g. OPENAI_API_KEY for OpenAI, ANTHROPIC_API_KEY for Anthropic)
2. Verify the model name is correct and available on the provider (e.g. gpt-4o, not gpt4o)
3. Ensure the systemPrompt is not empty — some providers return blank responses without a system prompt
4. Check the Run History for error details — look for 401 (invalid key) or 404 (model not found)

---

Symptoms: You added action nodes but they don't appear as tool slots on the AI Agent.

Solutions:

1. Connect the action node to the agent's Tools handle (the dynamic purple handle), not the message input
2. Only action nodes work as tools — triggers and logic nodes (If, Switch, Stop) cannot be tools
3. The agent supports up to 20 tool connections. If you need more, use MCP Server nodes which bundle multiple tools per connection

---

Symptoms: The agent doesn't remember previous messages in the conversation.

Solutions:

1. Ensure a memory plugin (Window Memory or Qdrant Memory) is connected to the agent's Memory handle
2. Set persistent: true on Window Memory (default is true, but check)
3. Verify your sessionKey binding resolves to a consistent value per user — use {{ api_request.headers["x-tensorify-session-id"] }} for Playground compatibility
4. Without a session key, the agent uses agent_memory:{workflow_id} and all users share one buffer

---

Symptoms: The Playground panel displays a prompt to deploy instead of the testing interface.

Solution: Deploy the workflow first from the Deployments tab in the bottom bar. The Playground sends real requests to the deployed endpoint — it cannot test undeployed workflows. Use the canvas Test button for pre-deployment testing.

---

Symptoms: Calling the endpoint with the OpenAI SDK returns a 404 Not Found error.

Solutions:

1. Set the API trigger's protocol to openai-chat (not rest)
2. Ensure POST is in the allowedMethods list
3. Check the path setting — the SDK appends /chat/completions to your base_url
4. Verify the workflow is deployed and the trigger URL is correct

---

Symptoms: Setting stream=True in the SDK returns a non-streaming response or errors.

Solutions:

1. Set streaming: true on the AI Agent node
2. Set protocol: openai-chat on the API Endpoint trigger
3. Ensure POST is an allowed method
4. The workflow must be deployed (streaming doesn't work with canvas Test)

---

Symptoms: The dashboard shows "Runner Offline" or "No Runner" on a workflow card.

Solutions:

1. Start the runner: tensorify runner start
2. Check your API key: tensorify whoami (run tensorify init if not configured)
3. Verify network connectivity — the runner needs outbound access to api.tensorify.io
4. For persistent execution, install as a service: tensorify runner install

---

Symptoms: External services get a 401 when sending webhooks to your endpoint.

Solutions:

1. Check the trigger's auth settings (authType, secret)
2. For api-key auth: the caller must send the key in the configured header (default X-API-Key)
3. For bearer-token auth: the caller must send Authorization: Bearer <token>
4. For webhook triggers with headerSecret: the caller must send the matching header value

---

Symptoms: A Code node fails with NameError or KeyError for a variable you expected to exist.

Solutions:

1. The Code node receives upstream data as input — access fields like input["body"] or input.get("name")
2. Check the edge connection — the Code node only sees data from directly connected upstream nodes
3. Use input.get("field", default_value) for optional fields to avoid KeyError
4. Environment variables are accessed via os.environ["VAR_NAME"], not through input

---

Symptoms: The agent doesn't see any tools from a connected MCP Server node.

Solutions:

1. Verify the MCP server URL is reachable from where the workflow runs
2. Only HTTP transport is supported — SSE transport is not yet implemented
3. Set MCP_API_KEY in Environment Variables if the server requires authentication
4. MCP tools are discovered lazily on the first agent request — check the Run History for connection errors