Skip to main content

Integrations Overview

Connect external tools and platforms to Zihin by invoking agents. Every integration — a no-code connector, a custom backend, or an embedded product — calls a hosted Zihin agent, not a raw LLM.

The principle: external input = agent invocation​

When something outside Zihin needs intelligence, it invokes an agent. The agent runs the full loop on Zihin's side — tools, MCP, memory, multi-step reasoning — under your tenant's RBAC, quota, HITL and audit. You send a message and get a governed result; you never wire up models, prompts, or tool execution yourself.

External platform ──▶ Zihin agent (react loop: tools · MCP · memory · HITL)
(n8n, Zapier, under tenant RBAC · quota · audit
your backend…) ◀── result (content, usage, session, sources)
Agents vs. the LLM API

This pattern is about agents (/api/v2/agents/:id). If you only need a stateless model call or media processing, see the LLM API instead. New integrations should prefer agents.

The shared core: @zihin/agent-client​

Every connector is a thin shell over one core package — @zihin/agent-client. It hides the transport (SSE streaming, async triggers) behind a small, governed API and is the reference client for any integration.

import { createClient } from '@zihin/agent-client';

const zihin = createClient({ apiKey: process.env.ZIHIN_API_KEY }); // zhn_live_*
const res = await zihin.invokeAgent({ agentId, message: 'Summarize this order' });

Choose your path​

PathUse it whenStart here
No-code connectorYou automate in a workflow tooln8n · Activepieces · Zapier
Agent orchestratorYou run agent companies in PaperclipPaperclip
SDK / your backendYou build a custom app or service (Node, serverless)@zihin/agent-client
AI IDE / MCPYou connect Claude Desktop, Cursor, Codex…MCP Server

Platform support​

PlatformStatusHow it connects
n8nAvailableZihin Agent node (community node) over the core
ActivepiecesAvailable (npm)Zihin piece (Invoke Agent action) over the core
ZapierPrivate (invite only)Zapier app (Invoke Agent) over the core — how to join
PaperclipAvailable (npm)zihin-paperclip-adapter external adapter over the core
Custom backendAvailable@zihin/agent-client (Node ≥ 20.16)
MakePlannedDeclarative — needs the synchronous JSON endpoint
AI IDEs (MCP)AvailableMCP Server
Current versions
  • @zihin/agent-client 0.9.1 · n8n-nodes-zihin 1.0.3 · @zihin/piece-zihin 0.6.1 · zihin-paperclip-adapter 0.7.1 — published on npm.
  • Zapier app 0.2.0 — private integration, by invite (Zapier).

These are package registry versions. Update installed connectors in their host platform separately. Activepieces and Paperclip currently depend on the @zihin/agent-client 0.8.x line; their dependency range does not select the latest 0.9.x client.

Marketplace approval and installed versions are independent of npm publication.

Behavior change in 0.7

Starting with client 0.7.0 and the corresponding connector releases, a turn the model provider refused (REFUSED) or that hit the context limit (CONTEXT_LIMIT) is an error in every connector; before, the backend's neutral sentence was returned as the answer. SESSION_BUSY (another turn still running in the session) is now retryable. Each page has a What changed section.

Why the same core everywhere

A single, tested client means every platform inherits the same behavior: governed invocation, streaming, session continuity, retry/timeout safety, SSRF protection and sanitized errors.

What you need​

  1. An agent. Create one in the console or via the Agents API. You invoke it by its agentId.
  2. An API key. zhn_live_* (or zhn_test_*) from Settings → API Keys. The tenant is embedded in the key. Direct invocation works with any role; provisioning async triggers needs an admin/editor key. See Authentication.

Security model​

  • Governed by the agent — every call runs under the agent's RBAC, quota and audit; HITL pauses are surfaced back to the caller (rolling out per platform).
  • Key handling — the key is sent only to Zihin hosts. Connectors keep an explicit Allow custom base URL opt-in (off by default) so the domain whitelist stays on unless you deliberately target a private deployment.
  • Sanitized errors — failures never leak the key or stack; they carry a requestId for support.

Next: pick a connector above, or read the Agent Client SDK reference.