Skip to main content

n8n

n8n supports MCP servers via its AI Agent → Tools node.

Three paths: hosted (recommended), direct HTTP calls, or a local server next to n8n.

Path A — Hosted endpoint (recommended)​

Prerequisites​

  • n8n 1.46+ (self-hosted) or recent n8n Cloud
  • A ColabHive API key

Config​

  1. In n8n, add an AI Agent node (or Chat Trigger workflow).
  2. Under Tools, click Add Tool → MCP Server.
  3. Configure:
    • Endpoint: https://mcp.colabhive.com/mcp
    • Transport: SSE (or HTTP for sync request/response)
    • Authentication: Header
      • Header name: X-API-Key
      • Header value: hive_xxx (or use an n8n credential expression)
  4. Save. n8n auto-discovers tools.

Try it​

Search the web for "ColabHive 2026 launch" using web-search-public, then summarize the top 3 results with qwen-2.5-7b-instruct-public.

The agent routes both calls through mcp.colabhive.com.

Path B — Direct HTTP calls to Actions API​

If you don't need the AI Agent node and just want sequential HTTP calls, use the HTTP Request node against api.colabhive.com/api/builder/v1/actions/{slug}:invoke directly. Cover detail in Plugin Integration → n8n.

Use Path B when:

  • Workflow is deterministic (you know which model to call)
  • AI Agent's tool budget is a constraint
  • You want to skip MCP entirely

Path C — Local server next to n8n (advanced)​

There is no published container image for colabhive-mcp. Run the PyPI package, with its HTTP extra, on a machine n8n can reach:

pipx install 'colabhive-mcp[http]'
colabhive-mcp serve --host 127.0.0.1 --port 8765

Then point n8n at http://127.0.0.1:8765/mcp. If n8n runs in a container, 127.0.0.1 inside it is the container itself: bind the server to an address the container can reach, and read Security before binding anything other than 127.0.0.1.

The local server runs in the same multi-tenant mode, so the key must be supplied per request — set it in n8n's MCP credential.

Recipes​

Workflow: doc → embedding → search index​

[Webhook trigger]
  ↓ (doc text)
[AI Agent (MCP: colabhive)]
  → embeddings-public
  ↓
[Postgres node: insert into vectors table]

Workflow: email → classify → route​

[IMAP trigger]
  ↓
[AI Agent (MCP: colabhive)]
  → uses your trained 'support-ticket-classifier'
  → uses a callable multilingual LLM discovered from the live manifest if non-English
  ↓ (label)
[Switch] → Slack / Zendesk / archive

Here MCP shines: the agent picks the right tool based on content.

Limits​

  • n8n Cloud free tier has no outbound to private IPs — Path A and C both work fine since mcp.colabhive.com is public.
  • AI Agent node tool budget per execution is ~30 calls. For heavy fan-out, prefer Path B.

Security​

  • API key lives in n8n credentials. n8n encrypts these at rest with N8N_ENCRYPTION_KEY. Set it.
  • For multi-tenant n8n (multiple users), each user must use their own ColabHive API key in their own credential — don't share.
  • See Security for the full doc.

Troubleshooting​

"Failed to connect to MCP server"​

curl https://mcp.colabhive.com/health
# OK?

Then check the credential in n8n — wrong header name (X-API-Key, not X-API_Key) is a common gotcha.

Tools listed but calls return "Unauthorized"​

The key is wrong or revoked. Update the n8n credential.

Slow tool calls​

For generative / batch models, increase the AI Agent timeout in n8n. The MCP server polls up to 5 minutes by default.

See also​