Using Promptfoo in n8n Workflows

This guide shows how to run Promptfoo evaluations from an n8n workflow so you can:

• schedule nightly or ad‑hoc LLM tests,
• gate downstream steps (Slack/Teams alerts, merge approvals, etc.) on pass‑rates, and
• publish rich results links generated by Promptfoo.

Prerequisites

Shipping a custom Docker image (recommended)

The easiest way is to bake Promptfoo into your n8n image so every workflow run already has the CLI:

# Dockerfile
FROM n8nio/n8n:latest          # or a fixed tag
USER root                      # gain perms to install packages
RUN npm install -g promptfoo   # installs CLI system‑wide
USER node                      # drop back to non‑root

Update docker‑compose.yml:

services:
  n8n:
    build: .
    env_file: .env # where your OPENAI_API_KEY lives
    volumes:
      - ./data:/data # prompts & configs live here

If you prefer not to rebuild the image you can install Promptfoo on the fly inside the Execute Command node, but that adds 10‑15 s to every execution.

Basic “Run & Alert” workflow

Below is the minimal pattern most teams start with:

Execute Command node configuration

promptfoo eval \
 -c /data/promptfooconfig.yaml \
 --prompts "/data/prompts/**/*.json" \
 --output /tmp/pf-results.json \
 --share --fail-on-error
cat /tmp/pf-results.json

Set the working directory to /data (mount it with Docker volume) and set it to execute once (one run per trigger).

The node writes a machine‑readable results file and prints it to stdout, so the next node can simply JSON.parse($json["stdout"]).

:::info The Execute Command node that we rely on is only available in self‑hosted n8n. n8n Cloud does not expose it yet. :::

Sample “Parse & alert” snippet (Code node, TypeScript)

// Input: raw JSON string from previous node
const output = JSON.parse(items[0].json.stdout as string);

const { successes, failures } = output.results.stats;
items[0].json.passRate = successes / (successes + failures);
items[0].json.failures = failures;
items[0].json.shareUrl = output.shareableUrl;

return items;

An IF node can then route execution:

• failures = 0 → take green path (maybe just archive the results).
• failures > 0 → post to Slack or comment on the pull request.

Evaluating n8n AI Agent prompts and outputs

If your goal is to test the prompt inside an n8n AI Agent / OpenAI node (not just run Promptfoo from a workflow), treat the n8n node like any other app contract:

1. Put the agent prompt in a file,
2. Map incoming n8n fields to tests.vars, and
3. Assert on the exact JSON or tool-call shape that downstream n8n nodes expect.

This works well when you want to regression-test an agent before wiring it into a larger workflow.

Validate JSON that downstream n8n nodes consume

If your agent is supposed to emit structured data for a Set, Code, Switch, or HTTP Request node, validate the payload directly.

prompts:
  - file://./prompts/n8n-support-router.txt

providers:
  - openai:gpt-5-mini

tests:
  - vars:
      customer_message: 'Customer wants to cancel order #4815 and asks for a refund'
    assert:
      - type: contains-json
        value:
          type: object
          required: [route, priority, reply]
          properties:
            route:
              type: string
              enum: [billing, support, sales]
            priority:
              type: string
              enum: [low, medium, high]
            reply:
              type: string

Use contains-json when the model may wrap JSON in prose or a markdown code block. If your node must return only JSON, use is-json instead.

Validate tool calls for agent workflows

If your n8n setup uses an OpenAI-compatible agent that should call tools before continuing, validate that Promptfoo sees a real tool call and that it matches your schema.

prompts:
  - file://./prompts/n8n-calendar-agent.txt

providers:
  - id: openai:gpt-5-mini
    config:
      tools: file://./tools/calendar-tools.yaml

tests:
  - vars:
      user_request: "Move tomorrow's standup to 3pm and notify the team"
    assert:
      - type: finish-reason
        value: tool_calls
      - type: is-valid-openai-tools-call

That pattern is especially useful when your n8n workflow branches on whether the LLM produced a tool invocation versus a final answer.

Useful building blocks

• /docs/configuration/tools for defining tool schemas
• /docs/guides/evaluate-json for JSON and schema assertions
• examples/openai-tools-call for a concrete OpenAI tool-calling config
• examples/eval-tool-use for finish-reason and tool-use checks across providers

Advanced patterns

Run different configs in parallel

Make the first Execute Command node loop over an array of model IDs or config files and push each run as a separate item. Downstream nodes will automatically fan‑out and handle each result independently.

Version‑controlled prompts

Mount your prompts directory and config file into the container at /data. When you commit new prompts to Git, your CI/CD system can call the n8n REST API or a Webhook trigger to re‑evaluate immediately.

Auto‑fail the whole workflow

If you run n8n headless via n8n start --tunnel, you can call this workflow from CI pipelines (GitHub Actions, GitLab, …) with the CLI n8n execute command and then check the HTTP response code; returning exit 1 from the Execute Command node will propagate the failure.

Security & best practices

• Keep API keys secret – store them in the n8n credential store or inject as environment variables from Docker secrets, not hard‑coded in workflows.
• Resource usage – Promptfoo supports caching via PROMPTFOO_CACHE_PATH; mount that directory to persist across runs.
• Timeouts – wrap promptfoo eval with timeout --signal=SIGKILL 15m … (Linux) if you need hard execution limits.
• Logging – route the stderr field of Execute Command to a dedicated log channel so you don’t miss stack traces.

Troubleshooting

Next steps

1. Combine Promptfoo with the n8n AI Transform node to chain evaluations into multi‑step RAG pipelines.
2. Use n8n Insights (self‑hosted EE) to monitor historical pass‑rates and surface regressions.
3. Check out the other CI integrations (GitHub Actions, CircleCI, etc) for inspiration.

Happy automating!