Hermes n8n MCP Plus

Local stdio MCP bridge for managing n8n from Hermes Agent — now with write tools.

This is a fork of CyberSamuraiX/hermes-n8n-mcp, extended with create_workflow, update_workflow, and delete_workflow on top of the original read-only/ops tool set. It gives Hermes full n8n workflow management without exposing n8n over the public internet and without putting API keys in your Hermes config.

What it does

Exposes these MCP tools:

• health — check n8n API reachability and optional Docker container status

• list_workflows — list workflows, optionally filtered by active state

• get_workflow — inspect one workflow with secret-bearing fields redacted

• find_workflows — search workflow metadata

• list_executions — list recent executions

• get_execution — inspect one execution; payload data is off by default

• recent_failures — recent failed/error executions

• export_workflow — fetch redacted workflow JSON for backup/review

• activate_workflow — activate a workflow by ID

• deactivate_workflow — deactivate a workflow by ID

• container_logs — optional Docker logs with line-level redaction

• create_workflow — create a new workflow from a JSON definition. Dry-run by default.

• update_workflow — patch an existing workflow by ID (name, nodes, connections, settings, tags). Previews current state before applying. Dry-run by default.

• delete_workflow — permanently delete a workflow by ID. Previews an export backup before deleting. Dry-run by default.

All three write tools require an explicit confirm=true argument to actually mutate anything. Called with confirm=false (the default), they return a preview of what would happen and make no API call that changes n8n.

Related MCP server: n8n-ops-mcp

Security posture

• Stdio only. No HTTP server. No public port.

• API key is loaded from environment or a local dotenv file.

• .env is gitignored.

• Example config uses REPLACE_ME, never a real key.

• Tool responses redact obvious credential, token, secret, password, and authorization fields.

• Execution payload data is disabled by default in get_execution.

• Workflow activation/deactivation, create, update, and delete are all production mutations. Treat them like loaded weapons.

• Write tools default to a dry-run preview; nothing is created, patched, or deleted unless the caller explicitly passes confirm=true.

• delete_workflow always fetches an export/backup preview of the workflow before a confirmed delete goes through.

Requirements

• Python 3.10+

• Hermes Agent with native MCP enabled

• n8n API key

• n8n reachable from the machine running Hermes, usually http://127.0.0.1:5678

Critical dependency pin: mcp==1.29.0

The official mcp PyPI package released a breaking v2.0.0 on 2026-07-28 that removed mcp.server.fastmcp entirely (renamed to MCPServer, moved module paths, swapped httpx for httpx2, and more). If requirements.txt uses a loose constraint like mcp>=1.29.0, pip install will resolve to 2.0.0 and the server will crash on import with:

ModuleNotFoundError: No module named 'mcp.server.fastmcp'

This repo pins mcp==1.29.0 (the last stable pre-v2 release) as a hard pin, not a floor. Do not loosen this constraint until the codebase is migrated to the v2 MCPServer API. If you ever see the above error, check pip show mcp — if it reports 2.0.0 or newer, run:

pip uninstall -y mcp
pip install "mcp==1.29.0"

Install

git clone https://github.com/csenguttuvan/hermes-n8n-mcp-plus.git
cd hermes-n8n-mcp-plus
python3 -m venv .venv
. .venv/bin/activate
pip install -r requirements.txt

Verify the install landed on the correct SDK version before going further:

pip show mcp
python -c "from mcp.server.fastmcp import FastMCP; print('OK')"

Store your n8n key

Manual version:

install -d -m 700 ~/.config/n8n-mcp-plus
cat > ~/.config/n8n-mcp-plus/env <<'EOF'
N8N_BASE_URL=http://127.0.0.1:5678
N8N_API_KEY=REPLACE_ME
N8N_MCP_TIMEOUT=30
N8N_CONTAINER_NAME=n8n
N8N_MCP_ALLOW_DOCKER_LOGS=true
EOF
chmod 600 ~/.config/n8n-mcp-plus/env

Replace REPLACE_ME locally. Do not commit the real file.

This fork's tools also read N8N_API_KEY / N8N_API_URL directly from the environment, not only from a dotenv file, so you can alternatively inject them straight from ~/.hermes/config.yaml using ${N8N_API_KEY} interpolation — see below.

Hermes config

Add this to ~/.hermes/config.yaml under mcp_servers. If the original n8n bridge is already registered, add this as a second, separate entry (n8n_plus) rather than replacing it — this keeps a safe read-only fallback available:

mcp_servers:
  n8n:
    command: /Users/admin/.hermes/mcp-installs/n8n/.venv/bin/python
    args:
      - /Users/admin/.hermes/mcp-installs/n8n/server.py
    enabled: true
    env:
      N8N_API_KEY: "${N8N_API_KEY}"
      N8N_API_URL: "http://localhost:5678/api/v1"

  n8n_plus:
    command: /Users/admin/projects/hermes-n8n-mcp-plus/.venv/bin/python
    args:
      - /Users/admin/projects/hermes-n8n-mcp-plus/server.py
    enabled: true
    env:
      N8N_API_KEY: "${N8N_API_KEY}"
      N8N_API_URL: "http://localhost:5678/api/v1"

Indentation matters. Both n8n: and n8n_plus: must sit at the same indent level, directly under mcp_servers:, with no other top-level key breaking the block in between. Validate the file parses correctly before reloading:

python3 -c "import yaml; d = yaml.safe_load(open('/Users/admin/.hermes/config.yaml')); print(list(d.get('mcp_servers', {}).keys()))"

Then reload MCP in Hermes:

/reload-mcp

Or from shell:

hermes mcp test n8n_plus

Tools register with the server-name prefix, e.g. mcp__n8n_plus__create_workflow, mcp__n8n_plus__health, distinct from the original bridge's mcp__n8n__* tools if running both side by side.

Smoke test outside Hermes

. .venv/bin/activate
python -m py_compile server.py
python -c "import server; print('imported OK')"
hermes mcp test n8n_plus

If import server hangs or throws ModuleNotFoundError: No module named 'mcp.server.fastmcp', re-check the mcp==1.29.0 pin above — this is almost always a dependency version problem, not a code problem.

Using the write tools

All three write tools follow the same dry-run-by-default pattern. Example flow for create_workflow:

Call mcp__n8n_plus__create_workflow with workflow={"name": "test", "nodes": [], "connections": {}} and confirm=false.

Returns a preview, no mutation:

{
  "ok": false,
  "error": "Dry run only. Set confirm=true to create the workflow.",
  "workflow_preview": { "name": "test", "nodes": [], "connections": {} }
}

Once the preview looks right, re-run with confirm=true to actually create it. The same pattern applies to update_workflow (previews current state and proposed patch) and delete_workflow (previews an export backup before deleting).

Recommended test order for any new environment: create_workflow then list_workflows to confirm it landed, then update_workflow, then delete_workflow — each with a throwaway workflow, verified against the n8n UI at each step.

Docker logs

container_logs shells out to Docker. If the user running Hermes cannot access Docker, set:

N8N_MCP_ALLOW_DOCKER_LOGS=false

The rest of the API tools will still work.

Notes for production use

• Keep n8n bound to loopback behind your reverse proxy.

• Do not expose this MCP bridge over Caddy, nginx, or Docker ports.

• Rotate n8n API keys if they ever hit chat logs, terminals, CI output, screenshots, or issue trackers.

• Back up workflows before mutating them. update_workflow and delete_workflow both preview state before confirming, but always check the preview yourself before passing confirm=true.

• Never loosen the mcp==1.29.0 pin in requirements.txt without first testing against the MCPServer v2 API.

Roadmap

• Migrate from FastMCP (v1.x) to MCPServer (v2.x) once the v2 API stabilizes and this fork's tool set is verified compatible.

• Consider adding run_workflow as a fourth write tool for manual execution triggers.

License

MIT. See LICENSE.