n8n-mcp-lite

n8n-mcp-lite

MCP redesigned for real-world workflow scale.

If your n8n workflow has more than ~40 nodes, standard MCP becomes structurally unusable. n8n-mcp-lite remains viable.

---

Benchmark

Measured on identical workflows. Token counts via the OpenAI tokenizer.

Above ~50 nodes, standard MCP stops being context-viable. n8n-mcp-lite remains usable.

Standard MCP implementations serialize the full workflow graph on every context turn. In large workflows, this produces token counts that exceed practical context windows entirely — making reliable AI assistance structurally impossible above a certain workflow size.

n8n-mcp-lite addresses this at the architecture level, not through compression or summarization hacks.

---

Who This Is For

n8n users with growing workflows

• Want AI to help manage, debug, and extend workflows safely
• Don't want to understand n8n JSON internals
• Need confidence that AI won't break production

Developers building production automation

• Care about deterministic serialization
• Need selective context for large workflows
• Want rollback and structural guarantees before any mutation
• Build workflows with 30, 50, 100+ nodes

---

Quick Start

1. Complete installation for your client (Cursor / Claude Desktop).
2. Ask your AI assistant: "List my n8n workflows."
3. Ask: "Scan workflow [name]." — returns a lightweight overview without loading full JSON.
4. Ask: "Focus on [node name]." — loads only the nodes you need.
5. Make changes. Every mutation runs security checks and creates an auto-snapshot before touching n8n.

No n8n internals knowledge required. The server handles format translation, validation, and context management automatically.

---

5-Minute Validation Test

After installation:

1. Ask: "List my workflows."
2. Pick a workflow with 30+ nodes. Ask: "Scan it."
3. Note the token estimate in the response.
4. Compare it to the raw JSON size of that workflow.
5. Ask: "Focus on [one node name]." — see how much context drops further.

If you have a workflow above 50 nodes, the difference will be immediately visible.

---

Real-World Scenarios

Large workflows (50+ nodes)

Standard MCP sends the entire workflow graph on every turn. At scale, this exceeds context windows entirely — responses degrade or fail. n8n-mcp-lite sends only what's needed, keeping interaction viable regardless of workflow size.

Debugging broken expressions

Ghost Payload attaches an inputHint to each focused node showing the exact $json field names arriving from upstream — pulled from real execution data. No more guessing what fields exist.

Safe AI refactoring

Every mutation runs a 7-layer security preflight before touching the n8n API. Hardcoded credentials, broken expressions, SQL injection patterns, and structural errors are caught and blocked. The API is never called on a failing check.

Recovering from a bad change

Every mutation auto-snapshots the workflow before executing. rollback_workflow restores any prior state exactly. Up to 20 snapshots per workflow, automatically pruned.

---

What This Is

n8n-mcp-lite is an MCP server that connects AI clients (Claude, Cursor, or any MCP-compatible client) to an n8n instance. It exposes 26 tools across six categories — reading, writing, activation, execution, versioning, and a node knowledge base — all operating on a compact serialization format purpose-built for AI interaction.

The server is designed to be:

• Token-disciplined — minimal context at every tool boundary
• Architecturally safe — every write operation runs security preflight and creates an auto-snapshot
• Deterministic — LiteNode format is stable and predictable across calls
• Model-agnostic — works with any MCP-compatible AI client

---

Installation

Prerequisites

• Node.js ≥ 18.0.0
• An n8n instance with API access enabled
• An n8n API key (Settings → API → Create API Key)

Clone the repository and build:

git clone https://github.com/LunkiBR/n8n-mcp-lite.git
cd n8n-mcp-lite
npm install
npm run build

---

Cursor

Create or edit ~/.cursor/mcp.json (global) or .cursor/mcp.json (per-project):

{
  "mcpServers": {
    "n8n-mcp-lite": {
      "command": "node",
      "args": ["/absolute/path/to/n8n-mcp-lite/dist/index.js"],
      "env": {
        "N8N_HOST": "https://your-n8n.example.com",
        "N8N_API_KEY": "your-api-key"
      }
    }
  }
}

Restart Cursor after saving. The MCP server will appear in the active tools panel.

---

Claude Desktop

Edit the Claude Desktop configuration file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "n8n-mcp-lite": {
      "command": "node",
      "args": ["/absolute/path/to/n8n-mcp-lite/dist/index.js"],
      "env": {
        "N8N_HOST": "https://your-n8n.example.com",
        "N8N_API_KEY": "your-api-key"
      }
    }
  }
}

Restart Claude Desktop. The server registers automatically on startup.

Note: Do not use relative paths in args. Claude Desktop sets the working directory to the system root on some platforms. Always use absolute paths.

---

Generic MCP Clients

Any client that supports the Model Context Protocol stdio transport:

{
  "command": "node",
  "args": ["/absolute/path/to/n8n-mcp-lite/dist/index.js"],
  "env": {
    "N8N_HOST": "https://your-n8n.example.com",
    "N8N_API_KEY": "your-api-key"
  }
}

The server communicates over stdin/stdout using the standard MCP protocol. No HTTP server is started.

---

Architecture

Why This Exists

AI-assisted automation breaks at scale because of context inflation. Standard MCP was not designed for workflows that grow past a few dozen nodes. n8n-mcp-lite is a serialization layer designed for AI reasoning — not a wrapper around the n8n API.

The Core Problem

Standard MCP implementations pass the raw n8n workflow JSON to the model context on every call. A 78-node workflow produces over 600,000 tokens of raw JSON — well beyond any practical context window, and containing mostly positional metadata, default values, and structural noise irrelevant to the task at hand.

This is not a tokenizer efficiency problem. It is a serialization design problem.

The n8n-mcp-lite Approach

n8n API (raw JSON)
       │
       ▼
┌─────────────────┐
│  Simplify Layer │  Strip defaults, normalize types, compress connections
└────────┬────────┘
         │  LiteNode format
         ▼
┌─────────────────┐
│  Graph Analysis │  Build adjacency graph, detect segments and branches
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│  Focus Engine   │  Compute boundaries, classify upstream/downstream,
│                 │  reduce non-target nodes to one-line dormant summaries
└────────┬────────┘
         │  Focused context payload
         ▼
       Model

The result: the model receives exactly the nodes relevant to the current task, with full parameter detail for focused nodes and compact summaries for everything else.

Subsystems

Serialization Layer (src/transform/simplify.ts) Converts raw n8n JSON to LiteNode format. Strips all-default values, omits empty fields, normalizes node type prefixes (n8n-nodes-base.httpRequest → httpRequest), and compresses the connection graph. The format is stable across n8n versions.

Graph Analysis (src/transform/graph.ts) Builds a directed adjacency graph from the serialized workflow. Detects independent branches, segments by router output index, identifies entry and exit points, and classifies node relationships for focus boundary computation.

Focus Engine (src/transform/focus.ts) Given a set of target nodes, computes the boundary of the focus area, classifies every other node as upstream, downstream, or parallel, and reduces non-focused nodes to single-line dormant summaries. Supports node selection, branch selection, range selection, and iterative expansion via expand_focus.

Ghost Payload — when an executionId is provided to focus_workflow, the engine reads that execution's input data and attaches an inputHint to each focused node showing the exact $json field names arriving from upstream. Eliminates expression guessing in debugging workflows.

Security Preflight (src/security/preflight.ts) Every mutating tool (create_workflow, update_workflow, update_nodes) runs a multi-layer validation pass before touching the n8n API:

If any error-level finding is detected, the mutation is blocked and the response contains a structured errors array with exact field locations and fix instructions. The n8n API is never called.

Version Store (src/versioning/version-store.ts) Before every mutation, the current workflow state is serialized to a JSON snapshot in .versioning/{workflowId}/. Up to 20 snapshots are retained per workflow; oldest are pruned automatically. rollback_workflow restores any snapshot exactly.

Approval Gate (src/approval/approval-store.ts) Optional team safety layer, enabled via N8N_REQUIRE_APPROVAL=true. When enabled, mutating tools return a pending response with an approve_token instead of executing immediately. The caller must re-submit with the token to proceed. All attempted mutations are recorded to an append-only audit log at .versioning/audit.log, regardless of approval mode status.

Node Knowledge Base (src/knowledge/) Static embedded database of n8n node schemas, workflow pattern templates, webhook payload schemas, expression recipes, and documented node quirks. Queried via get_node, search_patterns, get_payload_schema, search_expressions, and get_n8n_knowledge.

---

Feature Comparison

---

Tool Reference

Reading

Writing

Activation & Execution

Versioning

Node Knowledge Base

Approval (Team Mode)

---

Configuration

All configuration is via environment variables passed through the MCP client config.

---

Design Principles

Context is a resource. Waste is architectural failure. The model should receive the minimum context necessary to complete the task correctly. Sending more is not safer — it is noisier.

Determinism is not optional. LiteNode format is predictable. The same workflow produces the same serialized output on every call. The model can reason about structure without compensating for variance.

Safety before execution — without exception. No mutation reaches the n8n API without passing preflight. No preflight-passing mutation executes without a prior snapshot. Rollback is always available.

The model is a client, not a collaborator in protocol design. The server makes no assumptions about which model is on the other side of the MCP boundary. The protocol and the format work identically across all compliant clients.

Minimal surface, maximum capability. 26 tools cover the complete workflow lifecycle — create, read, update, delete, activate, debug, rollback, and learn — without redundancy.

---

Feedback & Contributing

This project is actively optimizing for real-world production cases, not toy examples. Issues, bug reports, feature requests, and real-world results are all welcome:

• Bug? → Open a bug report
• Feature idea? → Open a feature request
• Token savings or workflow story? → Share your feedback

If you're running workflows above 50 nodes, open a feedback issue with your node count, scan_workflow token estimate, and use case. Edge cases from real workflows directly shape what gets built next.

Code contributions are welcome. When adding or modifying tools, maintain the LiteNode format contract and ensure all mutating paths run through the preflight pipeline.

The structured issue templates live in .github/ISSUE_TEMPLATE/.

License

MIT