What changed in Zihin.
Documented product milestones. Historical entries describe their time; check the documentation for current availability.
October 2026
Conversation control by turn
Chat identifies each turn, supports cancelling the corresponding execution and handles resending the same message without starting duplicate work. Notifications open the right company and context. Documented milestone for the October 14–17 releases.
Source and technical detailsMCP update: data through API or MCP tools
MCP package version 2.2.2 accompanies the removal of direct database connection tools, the db_config type and the db_event trigger. Data integrations use API tools or MCP servers such as the Bridge.
Source and technical details
September 2026
Published Node.js client and connectors
@zihin/agent-client 0.4.0 provides agent invocation, streaming and async triggers. This milestone also records n8n and Activepieces packages and the Paperclip adapter. Installation and marketplace availability vary by platform.
Source and technical details
June 2026
Zihin Chat — app to work with your agents
chat.zihin.ai is live: an installable app (PWA) where each agent is a contact. Organized conversations (unread and pinned), rich replies (tables, receipts and quick-reply buttons), human approvals right in the chat, and proactive notifications that deep-link to the conversation. Per-agent opt-in — you choose which agents your team can reach.
Connectors — updated record
Update of November 2, 2026: see the September milestone for the published n8n, Activepieces and Paperclip packages. Zapier remains outside the published catalog.
OpenRouter — 100+ models with your own key
Connect your OpenRouter account and unlock 100+ models on the platform, with cost reported by the provider itself. Switch models without rewriting the agent — and with no markup from us on usage.
Per-agent spend cap — stops before it blows the budget
Set a dollar limit per agent. When it reaches the cap, the agent stops on its own — even when running on your own key. Cost predictability without having to watch usage.
Action approvals — one member approves what another triggers
New governance rules on top of human-in-the-loop: you define the scope of actions that require approval and let one user approve another's action, right in the chat. Accountability between people, not just between person and agent.
May 2026
Session Control — pause, intervene and block without touching the agent
Five new levers to control attendance without editing the agent: suspend a specific conversation, take over manually replying through the agent's own channel, cancel an in-progress turn (loop or wrong answer), block a contact across all your tenant's agents, and view a contact's aggregated cross-session history. No quota cost when control blocks.
Builder Agent — edit your agents by chatting with AI
Builder Chat is now an AI agent that understands your agents' configuration and proposes validated changes. You describe what you want to adjust — persona, tools, triggers, MCP servers, security policies, skills, workflow, database connections — and Builder creates proposals you approve with one click. Each proposal is validated before becoming a real change, and nothing applies without your confirmation.
Long-term memory — agents remember contacts across sessions
Agent memory is now persistent and cross-session: what it learns about a contact in one conversation is available in the next, even days later. You choose which agents need long-term memory (continuous support, follow-up, customer profile) and which don't (one-shot tasks, public agents without a fixed identity) — per-agent control, without affecting the others.
April 2026
Public site launched in beta
New routes (/sobre, /contato, /seguranca), LGPD consent banner, legal pages (terms, privacy, cookies) and SEO foundations (sitemap, JSON-LD). Ready for organic traffic.
Workroom: live sessions and replay
Every execution event now streams tenant-wide via SSE. Full replay of past sessions, with phases, tool calls and per-call latencies.
Zihin Tunnel in production
First pilot client connected via Tunnel. Outbound WebSocket, on-premise credentials, zero open ports on the client's firewall.
Multi-agent with team memory
An agent can invoke another agent via the invoke_agent tool. Shared memory across agents in the same pipeline. Depth limit controlled by policy.
API changes, endpoint by endpoint, live in the technical changelog: docs.zihin.ai/reference/changelog