n8n

Tracira has a native n8n node. Add monitoring to any workflow with a single node, no HTTP Request configuration required: connect once with your token, then branch on the verdict with an IF node.

Verdict response

{
  "ok": true,
  "id": "b1c2d3e4-...",
  "status": "pass",
  "verdict": {
    "pass": true,
    "rules": [{ "name": "Tone Validator", "triggered": false }]
  }
}

Reacting to decisions: the Tracira Trigger node

The package also ships a Tracira Trigger node ("Watch decisions") that starts a workflow the moment an output gets a verdict or a human decision in Tracira. It fires on approved, rejected and edited by default, and you can opt into taken over (a reviewer handled it themselves, so a waiting workflow can stop expecting an approval that will never come), taught (a reviewer coached the AI after an output already went out), flagged, passed and error. Activating the workflow registers the trigger with Tracira automatically; there is nothing to set up in the dashboard.

When a reviewer edits an output, output already carries their corrected version, so a workflow that maps output delivers what the human approved rather than the version they replaced. correctedOutput and aiOutput are there when you need to tell the two apart.

The typical human-in-the-loop pattern uses two workflows:

Workflow A: [ AI drafts ] → [ Tracira: Send an output ]           (submit for review)
Workflow B: [ Tracira Trigger ] → [ IF decision = approved ] → [ Send the reply ]

Decision events include the AI output and its metadata, so the approval workflow can deliver the approved reply with no extra lookup.

Holding an action for approval

When your AI decides to do something with side effects (issue a refund, delete a record), set After Check to Wait for a Human to Approve. That reveals Action Name and Action Summary, both required, plus optional Action Parameters (JSON), Callback URL and Callback Events.

The two halves do different jobs. The action describes what is being reviewed: the summary is what reviewers read verbatim to approve or reject, so write it as a clear sentence like Refund EUR 49.00 to Alice Martin (order #8841). The choice itself decides whether a person is asked at all, and it holds the output in the review queue whatever your rules conclude.

Rules still run either way, and they remain the right tool when only some actions need a person: a data-field rule on action.params.amount flags refunds over a threshold and lets smaller ones through. When a rule flags the output, its explanation is what the reviewer reads.

Set a Callback URL so the workflow resumes automatically once a person decides, or leave it blank and start the second workflow with the Tracira Trigger instead. You can also submit an action with no AI Output at all, when there is no message to review and only a step to approve.

Logging a conversation

For a chatbot workflow, send one output per exchange - do not paste the full message history into every output:

• Input Text: only the newest user message (the one your AI just replied to)
• AI Output: the reply to it
• Session ID: your conversation or thread ID (e.g. {{ $json.threadId }}) - the same value on every turn

Tracira stitches all outputs sharing a Session ID into a single conversation thread, so reviewers read the whole exchange in order with nothing repeated.

Hosted instructions (self-improving prompts)

Tracira can host the instructions (system prompt) your AI runs with, so the prompt lives next to the people reviewing its output instead of hardcoded in your AI node. Reviewers see, edit, and restore every version on the Instructions page in the dashboard; your workflow always drafts with the latest.

1. Add Tracira → Get Instructions at the start of the workflow. Use the same Project Name and Task Name as your Send an output node, and put your prompt into Starter Instructions - it is saved as version 1 on the very first run. Map the returned Instructions into your AI node's system message.
2. In Send an output, map the returned Version into Instructions Version so every output links back to the exact instructions that produced it.
3. To make the prompt self-improving, react to changed decisions with the Tracira Trigger node: have an AI step rewrite the current instructions to follow the reviewer's comment, then save the result with Tracira → Update Instructions. The next run fetches the improved version automatically.

Attaching files

For small files, use the Input Attachments or Output Attachments fields on Send an output. Each attachment has four sources: Upload File (map a binary field, keep under ~3 MB), From URL (a publicly accessible HTTPS link), Tracira Upload (a key from the Upload a File operation), or Already in Tracira (a file stored on an earlier output).

For large documents, add the Tracira → Upload a File operation before Check an output. It uploads the binary straight to Tracira storage (up to 32 MB) and returns a key; map that key into an attachment with source Tracira Upload.

[ AI node ] → [ Tracira: Upload a File ] → [ Tracira: Send an output ]

Getting a file back for a redo

Every event on the Tracira Trigger carries an attachments array listing the files stored on that output, each with a filename, a key, and an absolute url.

That is what makes a redo possible for document work. A workflow that parsed a PDF has finished long before the reviewer reads it, so when they send it back with a comment the workflow no longer holds the file. Add the Tracira → Download a File operation, map the attachment's url (or key) into File, and it lands in a binary field your AI node can read alongside the reviewer's comment.

[ Tracira Trigger ] → [ Tracira: Download a File ] → [ AI node ] → [ Tracira: Send an output ]

On that final Send an output, set Revision Of to the original output ID so the versions form a chain, and add an attachment with source Already in Tracira mapping the same key. The document stays on the new version without being uploaded again.

Manual setup without the node

Prefer to wire it up yourself, or running an n8n version where the community node is unavailable? Everything above is a thin wrapper over the public API: send a POST to https://tracira.com/api/logs from an HTTP Request node with an Authorization: Bearer YOUR_TOKEN header. See the HTTP guide for the payload reference.