n8n-nodes-aot-harness
Advanced tools
CHIP + Atom of Thoughts (AoT) Agent Harness — multi-provider n8n community node with a Process Analyst Worker for operational case triage (email, document, automotive service). Pluggable vaults (Obsidian, Supabase) and Human-in-the-Loop gate.
CHIP + Atom of Thoughts (AoT) for n8n — multi-provider agent harness with built-in cost tracking, plus a Process Analyst Worker for operational case triage.
A community node that brings the AoT-Harness (atomic decomposition + QA loop) into n8n. One node, one goal — the harness decomposes the task into atoms, runs them in parallel, QA-scores the result, and returns the polished output plus a per-call cost breakdown.
The Process Analyst Worker turns unstructured operational input into structured next-step decisions for n8n workflows: classification, missing information, urgency, automation potential, risk flags, human review and draft response.
Der Process Analyst Worker übersetzt unstrukturierte operative Eingaben in strukturierte nächste Schritte für n8n-Workflows: Klassifikation, fehlende Informationen, Dringlichkeit, Automatisierungspotenzial, Risiken, Human Review und Antwortentwurf.
Process Analyst — single-call structured triage of one operational case (email, document, support ticket, workshop request, internal note)generic, email_support, document_intake, automotive_servicehuman_review_required is forced true whenever confidence falls below the threshold, missing facts exist, risk flags appear, or input is too short to be reliableSee the Process Analyst Worker section below.
cost.total_usd, tokens, breakdown by provider and by model⚠️ Breaking change for existing v0.2.x users — see Migration below.
In n8n → Settings → Community Nodes → Install:
n8n-nodes-aot-harness
Then add the credential(s) for the provider(s) you want to use: AoT Harness — Anthropic / OpenAI / Google Gemini / Mistral / OpenRouter.
claude-sonnet-4-6).Erstelle eine kurze IDD-Dokumentation für einen Kunden, 42 J., sucht private Haftpflicht…
The node returns:
{
"goal": "...",
"result": "polished final output",
"qa_score": 0.92,
"success": true,
"atoms_done": 4,
"atoms_total": 4,
"provider_used": "anthropic",
"model_used": "claude-sonnet-4-6",
"cost": {
"total_usd": 0.0143,
"total_calls": 6,
"prompt_tokens": 2480,
"completion_tokens": 1120,
"by_provider": { "anthropic": { "cost_usd": 0.0143, "calls": 6, ... } },
"by_model": { "claude-sonnet-4-6": { "cost_usd": 0.0143, ... } }
}
}
Toggle "Enable Mixed-Provider Mode" in the node:
Typical results on a German IDD-documentation task:
| Setup | Cost / run | QA Score | Saving |
|---|---|---|---|
| Single (Claude Sonnet only) | ~$0.014 | 0.92 | — |
| Mixed (Claude Sonnet decompose + Gemini execute) | ~$0.003 | 0.88 | ~78% |
Annualized at 1k runs/month: ~$130/year saved per workflow.
A ready-to-import demo workflow ships in examples/mixed-provider-cost-saver.json — runs both setups against the same goal and reports the delta in a Code node.
A full Versicherungsmakler-use-case lives in examples/schadenmeldung-triage/:
qa_score < 0.75)README.mdThis is the workflow to show an interested broker.
| Provider | Default model | Best for |
|---|---|---|
| Anthropic | claude-sonnet-4-6 | Decomposer, complex reasoning |
| OpenAI | gpt-4o | General-purpose, structured output |
gemini-2.5-flash | Cheap, fast atom executor | |
| Mistral (EU) | mistral-large-latest | GDPR-sensitive workloads, EU residency |
| OpenRouter | anthropic/claude-sonnet-4-6 | One key for 100+ models, model A/B tests |
Per-provider model dropdowns are pre-curated in the node UI.
Goal
│
▼
[Decomposer LLM] AoT decomposition → AtomGraph (1–6 atoms, dependency-aware)
│
▼
[Executor LLM] Atoms run in parallel via Promise.all (where dependencies allow)
│
▼
[Executor LLM] QA-Agent scores 0–1 (retry on failure)
│
▼
{ result, qa_score, success, atoms_used, cost }
In single-provider mode Decomposer = Executor. In mixed mode they're independent — different provider, different model, different credential.
| Mode | Behavior |
|---|---|
| CHIP + AoT | Full pipeline: decompose → solve atoms → QA loop |
| AoT only | Decompose + solve, no QA (faster, less polished) |
| Webhook (Python) | Forwards goal to a running aot-harness Python server. Use this when you want the full Vault/Obsidian pipeline. |
| Process Analyst | Single-call triage of one operational case → structured JSON with classification, entities, urgency, missing info, risk flags, automation potential, draft response, and a Human-Review flag. |
A focused mode for operational case triage. Pick Mode → Process Analyst, paste one inbound case (email, support ticket, document excerpt, workshop request, internal note), and the worker returns a single structured JSON payload that the rest of the n8n workflow can branch on.
| Parameter | Type | Default | Description |
|---|---|---|---|
Operational Input | string (multi-line, required) | — | The case to analyze. |
Process Profile | options | generic | Biases what the analyst pays attention to: generic / email_support / document_intake / automotive_service. |
Output Language | options | de | Language for summary, next_best_action, draft_response, missing_information, risk_flags. Field names + enum values stay in English. |
Human Review Threshold | number 0–1 | 0.7 | Below this confidence the case is flagged for review. Missing info or risk flags also trigger review independent of confidence. |
Include Draft Response | boolean | true | If false, draft_response is null. The worker never sends — drafts are read-only. |
Strict JSON | boolean | true | On parser failure, returns a safe empty analysis with human_review_required = true instead of throwing. |
The active Provider + Model at the top of the node is reused as the analyst LLM (no separate decomposer needed — this is a single call).
Operational Input:
Sehr geehrte Damen und Herren,
mein Golf macht beim Bremsen seit ein paar Tagen ein schleifendes Geräusch.
Können Sie mir nächste Woche einen Termin anbieten?
Viele Grüße
Max Mustermann
Profile: automotive_service · Language: de · Threshold: 0.7
Output:
{
"mode": "process_analyst",
"profile": "automotive_service",
"analysis": {
"process_type": "terminanfrage",
"business_area": "service",
"summary": "Kunde fragt nach einem Werkstatttermin für nächste Woche wegen Bremsgeräuschen am VW Golf.",
"detected_entities": {
"customer_name": "Max Mustermann",
"company": null,
"email": null,
"phone": null,
"vehicle": "VW Golf",
"license_plate": null,
"requested_date": "nächste Woche",
"topic": "Bremsgeräusch / Werkstatttermin"
},
"urgency": "medium",
"confidence": 0.82,
"missing_information": [
"Telefonnummer",
"Kennzeichen",
"konkreter Wunschtermin"
],
"next_best_action": "Rückfrage an Kunden mit Bitte um Kennzeichen, Telefonnummer und zwei mögliche Zeitfenster für nächste Woche vorbereiten.",
"automation_potential": "high",
"human_review_required": true,
"risk_flags": [
"Terminwunsch unklar",
"Fahrzeugdaten unvollständig"
],
"recommended_workflow_tags": [
"service_intake",
"appointment_request",
"human_review"
],
"draft_response": "Hallo Herr Mustermann, vielen Dank für Ihre Anfrage. Damit wir Ihnen einen passenden Werkstatttermin anbieten können, senden Sie uns bitte noch Ihr Kennzeichen, eine Telefonnummer und zwei mögliche Zeitfenster für nächste Woche.",
"_profile": "automotive_service",
"_trigger_reasons": ["3 missing fact(s)", "2 risk flag(s)"]
},
"provider_used": "anthropic",
"model_used": "claude-sonnet-4-6",
"cost": { "total_usd": 0.0021, "total_calls": 1, "prompt_tokens": 480, "completion_tokens": 220 }
}
The downstream Switch node branches on analysis.human_review_required → either send the draft via Gmail or route to a Makler/Service-Annahme inbox.
human_review_required is deterministic — set in code (not just by the model) whenever any of these is true:
confidence < humanReviewThresholdmissing_information.length > 0risk_flags.length > 0_trigger_reasons lists exactly which of the above fired — useful for explainability + debugging.recommended_workflow_tags always contains human_review when the gate fires, so a Switch node can match on a single tag.| Profile | Pays attention to |
|---|---|
generic | Process type, business area, missing info, next step. Branch-neutral. |
email_support | Customer concern, urgency, missing info, draft reply, escalation, support category. |
document_intake | Document type, mandatory fields, missing fields, plausibility, extraction hints. |
automotive_service | Vehicle, license plate, requested date, service topic, callback need, missing vehicle info, next workshop step. |
v0.3.0 is a breaking change because the node now requires a provider-specific credential.
After updating:
Anthropic (matches the v0.2.x default behavior).ANTHROPIC_API_KEY.claude-sonnet-4-6).The legacy aotHarnessApi credential type is still registered so existing credential entries don't disappear from your n8n credentials list — but the new node doesn't read it. Delete it after migration if you like.
| Credential | Required env var | Notes |
|---|---|---|
| AoT Harness — Anthropic | ANTHROPIC_API_KEY | Default provider |
| AoT Harness — OpenAI | OPENAI_API_KEY | Optional OpenAI-Organization header |
| AoT Harness — Google Gemini | GEMINI_API_KEY | Get one at aistudio.google.com |
| AoT Harness — Mistral (EU) | MISTRAL_API_KEY | EU-hosted, GDPR-compliant |
| AoT Harness — OpenRouter | OPENROUTER_API_KEY | Optional HTTP-Referer / X-Title |
You only need credentials for providers you actually use.
MIT
FAQs
CHIP + Atom of Thoughts (AoT) Agent Harness — multi-provider n8n community node with a Process Analyst Worker for operational case triage (email, document, automotive service). Pluggable vaults (Obsidian, Supabase) and Human-in-the-Loop gate.
We found that n8n-nodes-aot-harness demonstrated a healthy version release cadence and project activity because the last version was released less than a year ago. It has 1 open source maintainer collaborating on the project.
Did you know?
Socket for GitHub automatically highlights issues in each pull request and monitors the health of all your open source dependencies. Discover the contents of your packages and block harmful activity before you install or update your dependencies.
Research
/Security News
A malicious NuGet package impersonating Sicoob exfiltrated client IDs, PFX passwords, and banking certificates through Sentry telemetry.
Security News
Feross Aboukhadijeh joins TBPN to discuss Socket's $60M Series C, 500%+ ARR growth, AI's impact on open source, and the rise in supply chain attacks.
Security News
OSV withdrew 157 OSV malware reports after automated false positives incorrectly flagged trusted npm and PyPI packages, sending bad records into tools that rely on OSV data.