Product

Put AI to work.
Without rebuilding your stack.

Zihin connects agents to your company's systems and channels through APIs, MCP tools, and integrations. Permissions, steps, and approvals follow your operation's rules.

In operation

From integration to operation.

Published case: agents support purchasing and quoting in agribusiness, connected to the operation's systems.

Executions and tool calls
console.zihin.ai/admin/workroom
Your company

Work Room

Live view of agent runs in your operation.

Live

Active runs

4

Supply Buyer

fd3ed31f-4e8d…

COMPLETED
whatsapp1m 12s
offer #57 created · suppliers notified

Quotation Response

6586385a-b53b…

TOOL CALL
whatsapp18s
ebarn.listar_prazos_pagamento

Lead Qualification

43a043e3-fac5…

THINKING
whatsapp34s
lead wants to sell — pivot to Marketplace

Weekly P&L Close

c9a0f5e1-7b32…

COMPLETED
cron2m 04s
report sent by email

Supply Buyer · cotai-compras

exec fd3ed31f-4e8d-0c7f-cdeb-3932c79c6c3a

Reasoning

8 product codes + branch 0104 received. Validating in ERP.

1.1s

erp_protheus · buscar-produtos ×8

codes: ["1306000069","1306000257", …]

ok · 8/8 validated~1.0s

erp_protheus · buscar-fornecedores ×8

ok · 3–10 suppliers/product~3.2s

Reasoning

Prioritizing by purchase history and average delivery time.

0.9s

ebarn · criar_oferta

offer: #57, items: 8

ok · 658ms$0.087

Process
Purchasing and quoting
Systems
TOTVS Protheus and the eBarn ERP
Connection
MCP Bridge, in production
Published caseGrupo ConceitoeBarn
Connectivity

Legacy, API, or ERP.
Connected to your agent’s tools.

Define the data and operations for each integration, with authentication, contracts and permissions. Databases and ERPs come in through a read-only copy, never through direct agent access.

+ via Replay

Zihin Replay keeps a read-only copy of your ERP inside your network. MCP Bridge publishes business queries over that copy as agent tools.

+ via API

Integrate REST APIs as typed tools, with a contract and validation defined for each operation.

+ via MCP

Connect your own or third-party MCP servers and expose their tools to the agent, with configured authentication and permissions.

Credentials protected with encryption and access controls.

OAuth for services supported by enabled connectors.

How Zihin gets in

your network → outbound 443 → Zihin Cloud

  • Your ERP · Protheus · Sienge · Sankhya. Stays untouched: no writes, no queries competing with your operation.
  • Zihin Replay · read-only copy · in your network. Replicates the agreed tables to a Postgres on your server, with changes reflected in seconds.
  • MCP Bridge · business tools · read-only. Turns the copy into ready queries: critical stock, reorder point, price anomalies, supplier risk.
  • Zihin Tunnel · outbound connection · port 443. A lightweight agent opens the connection from the inside out. No inbound port, no VPN, no public IP.
  • Zihin Cloud · agents under policy. Your agents call the tools through the tunnel, with roles, approvals and audit.
  • Outbound only, never inbound. The only network requirement is outbound 443. No inbound firewall rules.
  • Credentials stay with you. Passwords and internal addresses are resolved in your network and never reach the cloud.
  • Read-only at the source. A read-only capture user in the ERP; the Bridge queries the copy with a read-only role and transactions.
  • Reach under your control. Allowed-host list, request limits and immediate token revocation, with every operation audited.
Ecosystem

Works with what you already use.

Solutions with confirmed integration in the Zihin ecosystem. Others come in through API, MCP or Replay after assessment.

ERPs and data
  • TOTVS Protheus
  • Sienge
  • Sankhya
  • Oracle Database
  • SQL Server
  • PostgreSQL
Channels
  • WhatsApp
  • Slack
  • Microsoft Teams
  • Twilio
  • Resend
Automation and code
  • n8n
  • Activepieces
  • Paperclip
  • TypeScript
  • Python
AI clients · via MCP
  • Claude
  • Cursor
  • GitHub Copilot
  • Windsurf
  • Gemini CLI
  • MCP

Brands and logos belong to their respective owners and indicate compatibility, not partnership or endorsement.

01 · Reaches

It doesn't just read.
It acts on your processes.

  • Configurable processes — typed steps and human tasks when enabled

  • Authorized tools — operations defined by each integration and contract

  • Pre-execution checks — review schemas, policies, and connectivity

Process building and validation
console.zihin.ai/admin/agents/984ae206
CCotai — Quotation AssistantLIVEclaude-haiku-4-5
Builder
Workflow "Purchase Request Flow" is preflight-validated on each publication.
Describe what to change…

Behavior (Persona)

purchasing assistant · Grupo Conceito Agrícola · professional tone

Workflow

Purchase Request Flow · 6 steps · strict mode

Typed steps (6)

1COLLECTIdentify need→ consultar_produtos
2TOOL CALLValidate products in ERPerp_protheus.buscar-produtos
3TOOL CALLFind supplierserp_protheus.buscar-fornec…
4VALIDATEValidate quotation→ coletar_dados
5CONFIRMConfirm with user→ criar_oferta
6TOOL CALLCreate offer in eBarnebarn.criar_oferta
Validation
PREFLIGHTACTIVITYALERTS
PREFLIGHT · 1 issue found
Connectivity1 failure

MCP erp-protheus · timeout exceeded

MCP erp-protheus

timeout exceeded · 3 attempts without response

MCP ebarn

ok · 3 tools · 0 failures

Webhook eBarn Chat

2,571 deliveries · no recent errors

Ask Builder to investigate →
Consistencyok

persona ↔ skills ↔ tools aligned

Schemaswarnings

6/6 valid contracts

Telemetry · 24hok

312 runs · p95 1.8s · 0 errors

From request to delivery

A complete task takes more than one step.

Zihin connects the parts: incoming data, process decisions and the delivery that reaches your team.

ZihinFrom request to delivery

Compare purchase orders and highlight the items that need attention.

  1. Receive request
  2. Query sources
  3. Analyze data
  4. Deliver results
Result

Orders that need attention

Orders that need attention
ItemStatusNext step
Material ADue soonReview delivery
Material BPrice changeCheck quotation
Material COn trackMonitor

Deadlines, prices and deliveries brought together in a purchasing analysis.

purchase-analysis.xlsxReady for the team
02 · Governs

Work with defined access and rules.

  • Context-scoped policy (CSP) — inherited, enforced on every call

  • Data protection — access and handling controls defined for each operation

  • Recorded policies — review changes and actions in available logs

  • Track tokens and costs — allowances and limits depend on the plan

Tenant roles and permissions · BYOK when enabled · protected credentials

CotaAI Supplier — Data Security

Protects technical data and supplier credentials

DataAgentPriority: 100

Hide rules

{
  "mask_pattern": "[DADO_PROTEGIDO]",
  "never_expose": [ "request_body_json", "credentials" ],
  "sensitive_fields": [ "token", "api_key", "password" ],
  "restriction_message": "I can’t share internal technical information."
}
Usage and allowance

Visibility to plan your usage.

Review token usage and costs by period. Applicable allowances and limits depend on your plan and account configuration.

Usage and configured limits
console.zihin.ai/admin/agents

Agents

Fleet — cost and spending caps in one place

Spending cap · active
Agents
5
Spend · 7d
$312
At cap
1
AgentCost · 7d% of capStatus
CS
Cotai · Supplies
cotai-suprimentosproduction
$124
62%raise →
$124 / $200
active
CF
Supplier quotation
sub · cotacaoproduction
$38
48%raise →
$38 / $80
active
FN
Invoice follow-up
sub · followup-nfeproduction
$90
100%raise →
$90 / $90
at cap
SE
eBarn support
suporte-ebarnproduction
$51
34%raise →
$51 / $150
active
OB
Customer onboarding
onboarding-clientetest
$9
12%raise →
$9 / $75
active

See this running on your operation — Request a diagnostic →

Zihin Chat · Consumption

Your team operates
with the agents.

  • Web chat and PWA for talking with agents and following their responses

  • Actions configured for approval stay pending until the designated person decides

  • Get activity and pending-task notices with a link to the conversation

Installable app (PWA) · per-user, role-based access — the end user never sees configuration, prompts, or other people's conversations.

Conversation, table and approval
Action approvals

One approves what another triggers.

When a step is configured for approval, the process waits for its designated approver. It can then continue according to the defined workflow.

Approval request and decision
03 · Observes

No black box. From execution to report.

  • Follow recorded sessions, steps, and tool calls

  • Review and manage executions using the controls available for your role

  • History and details are available according to access and retention settings

Usage and results overview
console.zihin.ai/admin/usage

Usage

Monitor resource usage and API calls

24h7d14d30d
Refresh
OverviewLatencyTools
Calls
223

Last 7 days

Tokens
1.4M↗ 6196%

~6.2K per call

Latency
1.8s

Normal

Success rate
100.0%

No errors

Cache hit
51.0%

1.9M cached

Call volume

Over time

223total
May 28May 31Jun 02Jun 04

By provider

Call distribution

Anthropic75%
Google23%
other2%

Calls by model

Top models used

100.0% ok
Claude Haiku167
Gemini Flash51
GPT-4 Mini3
GPT-42

Calls by source

Top sources

webhook180
sse33
builder8
direct2

Zihin Watch

Your AI operation, also in your inbox.

Observability reports to follow agent activity, AI usage and areas that may need attention.

  • Follow activity.Review executions, contacts and trends by period and agent.
  • Understand usage.See AI costs split between your own key and Zihin credits.
  • Spot areas to review.See integration failures and quota usage alongside operational context.

Weekly summaries and monthly email reports for account owners and accounts with email sending enabled.

Email reports
Zihin WatchTo: account owner
Portability

Build the agent once.
Use it anywhere.

The same agent answers in n8n, in your code, in Claude or on WhatsApp. Roles, budget, approvals, audit and session travel with it, whatever the surface.

  • Same rules on every call — roles, per-agent budget, human approvals and audit apply in n8n, in the SDK and over MCP

  • Conversations carry over — the same session continues across surfaces, with history kept on the platform

  • One tested core — streaming, retries and errors behave the same in every published connector

Zapier and Make are not public yet. The Python SDK covers invocation and sessions; context, attachments and triggers are planned for later versions.

Your agent

One definition, every surface

Persona, tools, policies and memory live on the platform. Every call is an agent call, never a raw model call.

  • RBAC
  • Budget
  • Approvals
  • Audit
  • Session

Automation

Packages published on npm; install as a node or piece and invoke the agent from any workflow.

  • n8n
  • Activepieces
  • Paperclip

Code

TypeScript SDK @zihin/agent-client, Python SDK zihin and a REST API with streaming.

  • TypeScript SDK
  • Python SDK
  • REST API

AI clients

Zihin is an MCP server: talk to your agents and manage the platform from the assistant you already use.

  • Claude Desktop
  • Claude Code
  • Cursor
  • VS Code
  • Windsurf
  • Codex
  • Gemini CLI

Channels and triggers

Webhooks, email and schedules put the same agent into conversations and routines.

  • WhatsApp
  • Slack
  • Teams
  • Email
  • Webhook
  • Schedule
Where you command

Command your agents
from your Claude.

  • Inspect and manage agents through MCP tools available to your credential

  • Native Zihin tools via MCP — Zihin is an MCP server for your assistant

  • Tools follow credentials and permissions; confirm sensitive actions in your MCP client

Claude Desktop · Cursor · Claude Code

Zihin tools in an MCP client
Claude · Zihin MCP connected
List the published agents and review the latest session
zlist_published_agents
Result

The latest session is from cotai-fornecedor (offer #53). The agent recovered from an eBarn 500 error on its own — remapped payment methods and terms, then retried: 5 offers recorded. I recommend keeping the retry and strengthening ID validation before submission.

Write a message…Selected model · High

Claude is an AI and can make mistakes. Check its responses.

See this running on your operation — Request a diagnostic →

Your AI, in production.

Tell us what your team needs to solve. We’ll map the systems, process and rules to get started.