n8n-nodes-hitl-stack-agent

An n8n community node that sends an AI agent's answer to
Cerebro for human-in-the-loop review — without pausing your workflow.

When this node runs, it sends the agent's answer (and, optionally, the upstream
question and context) to Cerebro, which turns it into a review case for a
human to approve or correct. The workflow continues immediately — the
execution is never parked — so the review happens out-of-band rather than
blocking the run.

Installation · Credentials ·
Operations · Node parameters ·
Output · Example workflow ·
Compatibility · Resources

Installation

Follow the
community node installation guide
in the n8n docs.

In the n8n UI (self-hosted):

1. Go to Settings → Community Nodes and click Install.
2. Type the exact package name n8n-nodes-hitl-stack-agent into npm Package Name.
   There is no search box: n8n installs by name. (The Browse button only opens
   npm's website in a new tab.)
3. Tick I understand the risks of installing unverified code from a public
   source and click Install.

Only the instance owner or an admin can install community nodes, and community
nodes must be enabled on your instance.

Manually (npm):

npm install n8n-nodes-hitl-stack-agent

After installation the HITLStackAgent node appears in the nodes panel.

Prerequisites

• A running n8n instance (self-hosted, or n8n Cloud once the node is verified).
• A Cerebro account with an API gateway Base URL, an API key, and at least
  one agent configured to route answers to human review. Refer to your Cerebro
  documentation for how to set up an agent and enable reviews.

Credentials

This node uses a single credential type: Cerebro API (cerebroApi).

The credential test verifies that the Base URL and API key are valid. The key
is only used to authenticate with your Cerebro gateway; it is never forwarded to
agents or included in the node output.

Operations

The node has a single behaviour: send the current item to Cerebro for review
(non-blocking). It does not pause or resume the execution.

• It must receive exactly one input item per execution. To review a batch,
  put a Loop Over Items node upstream.
• The agent's answer is read from the item's output field; the question shown
  to the reviewer is taken from the upstream context (see parameters below).

Node parameters

Output

The node passes the input item through and adds a review object describing what
was sent:

{
  "output": "…the agent answer…",
  "review": {
    "mode": "cerebro",
    "status": "pending",
    "agentId": "ag_...",
    "traceId": "…correlation id…",
    "delivered": true,
    "acceptedSpans": 1
  }
}

• status is "pending" only when Cerebro accepted the answer for review;
  otherwise it is "failed" — so a delivery failure never reads downstream as
  "awaiting human input".
• delivered / acceptedSpans reflect Cerebro's response. Delivery is
  best-effort: a transport error is reported on the item (and status: failed)
  rather than throwing, so a telemetry hiccup doesn't fail your workflow item.
• traceId lets you correlate the n8n run with the Cerebro review case.

How it works

flowchart LR
  A[Upstream AI Agent] -->|answer| B(HITLStackAgent)
  B -->|answer + question| C[Cerebro Services]
  B -->|item + review marker| D[Next node]
  C -.->|opens a case| E[Human reviewer]

The node collects the agent's answer (and, optionally, the upstream question and
context) and hands it to Cerebro, which handles reviewing internally and
opens a case for a human. How Cerebro processes and routes that data is its own
concern — this node treats it as a black box and does not wait for the review.

Example workflow

Chat Trigger → AI Agent (+ chat model) → HITLStackAgent → (next steps)

1. The AI Agent produces an answer on output.
2. HITLStackAgent (with an Agent selected and Include Upstream Context on)
   sends the question + answer to Cerebro and continues.
3. A reviewer sees the case in Cerebro; your workflow proceeds in parallel.

Compatibility

• Requires n8n 1.x or newer (n8nNodesApiVersion: 1).
• Node.js >=20.
• No run-time dependencies — the node uses only the n8n-workflow peer API and
  Node's built-in crypto.

Limitations

• Handles one item per execution (use Loop Over Items for batches).
• The agent's answer must be on the item's output field.
• The question shown to the reviewer is read from an upstream item (within
  Context Depth) that has a chatInput, input, query or question field,
  as a Chat Trigger provides. If the question exists only in the AI Agent's own
  prompt setting, the reviewer sees the answer without the question.
• Model name and token usage are not captured in an AI Agent + chat-model
  sub-node topology, because n8n does not expose sub-node output to downstream
  nodes.

Resources

• n8n community nodes documentation
• Creating nodes

License

MIT