n8n-nodes-verica

n8n community node for LLM observability: it sends your AI Agent / LLM
executions to Verica as traces, with inputs, outputs,
tool calls, token usage and cost.

Install

• Self-hosted n8n: Settings → Community Nodes → Install n8n-nodes-verica.
• n8n Cloud: available once the node is verified; meanwhile use the
  importable recipe from Verica's Connect dialog (Traces → Connect → n8n).

Usage

1. Create a Verica API credential: an ingest token (Verica → Settings →
   API tokens, with the ingest scope). The endpoint defaults to the Verica
   cloud.
2. Drop Verica Trace after your AI Agent (or any LLM step), or pick its
   Send a trace action straight from the canvas. The node has one resource,
   Trace, with one operation, Send — an ingest token reaches exactly one
   endpoint. The defaults
   read $json.output, $json.chatInput, $json.intermediateSteps and
   $json.sessionId; set Model so the trace can be priced. It works after
   the AI Agent or "Message a model" out of the box: object outputs are
   flattened to text automatically.
3. Enable Return intermediate steps on the AI Agent so tool calls land in
   the trace (Verica's tool_check grader can then assert on them).
4. Map System Prompt so the trace carries the instructions the agent ran
   with (judges see it in {{ input }}, and "Promote to dataset" can copy it
   into the dataset prompt). n8n does not emit the agent's system message
   downstream: read it from the agent's parameter, e.g.
   {{ $('AI Agent').params.options.systemMessage }} (use your agent node's
   name). Empty = only the user message is sent.

Sessions. Executions sharing a sessionId (the Chat Trigger provides one)
reassemble into a Verica session, one turn per execution. Turns are stored as
deltas matched on byte-identical text, so if you map Input to something
richer than the new message (for example a rebuilt history), resend it exactly
as sent before: any per-turn mutation (like appending "Respond in JSON" to the
last message only) makes every turn degrade to a full copy of the
conversation.

With "Message a model" (OpenAI Responses API), token usage is picked up
from the response automatically: adding the Input Tokens / Output
Tokens options captures the response's usage, including the Reasoning
Tokens (a breakdown of output tokens) and Cached Tokens (a breakdown of
input tokens, priced at the cache rate) when the response reports them. The
response does not echo your prompt: map Input to the node's parameter,
e.g. {{ $('Message a model').params.responses.values[0].content }} (hover
the Prompt field to confirm the parameter path in your n8n version).

Tool calls in the trace need a node that emits them. The AI Agent does
(enable Return intermediate steps); "Message a model" with attached tools
runs its tool loop internally and returns only the final answer, so its
executed calls are not capturable downstream. For tool-using workflows, use
the AI Agent.

With an AI Agent, map Model to the chat-model sub-node's parameter,
e.g. {{ $('OpenAI Chat Model').params.model.value || $('OpenAI Chat Model').params.model }}.
Token usage is NOT capturable in agent workflows: n8n does not propagate the
chat model's tokenUsage to the agent output or to downstream expressions
(open issue n8n#26302); the
trace lands without tokens/cost but stays fully evaluable.

The node is fail-open: an export error never breaks your workflow; the item
passes through with a vericaError annotation instead.

License

MIT