ZBrain

Flows assemble from 159 components that read and write enterprise systems, among them Salesforce, ServiceNow, HubSpot, Zendesk, Jira Cloud, SharePoint, Microsoft Dynamics, Snowflake, Databricks, Postgres, Stripe and Twilio, plus HTTP, SOAP and code steps for anything else.

Agents in a crew take default tools for Google search, knowledge base search, HTTP calls and deep research, custom tools written in code with npm packages pulled in at a pinned version, and both ZBrain and external MCP servers whose endpoints the agent calls at run time. A Salesforce lead, a ServiceNow record or a Zendesk ticket can therefore be read, updated and passed on inside one run.

Sourcedocs.zbrain.ai/zbrain-documentation/zbrain-builder/zbrain-ai-agents/agent-crew/define-crew-structure.mdread 2026-10-18

Agent Crew arranges agents on a visual canvas under a supervisor that receives the input, delegates parts of the task to child agents, evaluates what they return and decides the next step, on LangGraph, Google ADK, Microsoft Semantic Kernel or Microsoft Agent Framework. Crews import and export as JSON, and agents from Azure AI Foundry or AWS Bedrock can be added as children. Single agents run a Flow, a visual blueprint of triggers, components, routers, loops, delays, queues and conditions, with data passed between steps, and a running Flow can be stopped mid execution.

Sourcedocs.zbrain.ai/zbrain-documentation/zbrain-builder/zbrain-ai-agents/agent-crew.mdread 2026-10-18

Knowledge bases ingest PDF, Word, PowerPoint, Excel, CSV, JSON and text files, images and video, from uploads or connected tools, with choices of embedding model, chunking strategy, retrieval technique and vector store, and a DeepDoc parser for complex documents.

Retrieval runs as vector, full text or hybrid search, or over a knowledge graph whose depth of connections the builder sets, and each source shows its chunks, metadata and graph of entities.

Automated reasoning extracts rules and variables from policy documents so agents apply them, information schemas turn unstructured files into structured fields, and a retrieval testing screen checks what a query returns.

Sourcedocs.zbrain.ai/zbrain-documentation/zbrain-builder/knowledge-base.mdread 2026-10-18

A human in the loop toggle on any agent in a crew pauses the run after that agent's task and shows its intermediate result, so a reviewer can approve it, edit the input or rerun the step before work passes downstream. In Flows, a Wait for Approval step holds execution until someone approves or rejects, and approval links let a decision come from outside the platform without stopping the run. Guardrails on apps and agents check responses for policy violations, jailbreak attempts and irrelevant content, and guardrail agents in the store screen output before it reaches people.

Sourcedocs.zbrain.ai/zbrain-documentation/zbrain-builder/flows/flow-components/approval.mdread 2026-10-18

ISO 27001:2022 and SOC 2 Type II certifications cover the platform, and data is encrypted with AES 256 in transit and at rest, across inputs, outputs and model calls.

Role based access control separates builders from operators and sets who can view, edit or create knowledge bases, apps, agents, prompts and Flows, while a multi tenant design isolates each customer's data.

Enterprise plans add custom SSO through enterprise identity providers and an infosec review, and the platform runs SAST, DAST and dependency scanning, network ACLs, daily backups and IAM controlled storage, with API keys masked when models are configured.

Sourcedocs.zbrain.ai/zbrain-documentation/zbrain-builder/security-features.mdread 2026-10-18

The Runs view lists every Flow execution with a unique ID, status, duration, token usage and cost, filters by time window, opens each run's step by step details and retries failed runs. Crews have their own monitoring screen showing each agent's task status and outputs, and agent dashboards track tasks and performance. Monitor captures every input and output of deployed agents and apps, logs successes and failures in an event dashboard and sends notifications, so operations teams see when an agent starts returning poor answers.

Sourcedocs.zbrain.ai/zbrain-documentation/zbrain-builder/flows/flow-runs-and-logs.mdread 2026-10-18

State is held within runs and sessions. Apps recall earlier turns of the same session through follow up conversation, with a context token limit that sets how much history is kept, and crews share context through scratchpads that each agent can now create and edit at its own level, which keeps one agent from overwriting another's notes. LangGraph crews carry state across the steps of a run, and a Clear Agent Data toggle controls what an agent keeps between tasks. No memory of users or past cases persists across sessions on its own.

Sourcedocs.zbrain.ai/zbrain-release-notes/zbrain-builder-release-notes/version-2.4.7-or-apr-7-2026.mdread 2026-10-18

Builder is cloud and model agnostic and deploys on premises, in a customer's private cloud, or on AWS, Azure or Google Cloud. On the hosted service, customers choose the data storage region that meets their residency rules, can delete or manage stored data at any time, and keep their data isolated from other tenants. Enterprise plans include deployment assistance from ZBrain's team for private installs.

Sourcezbrain.ai/zbrain-builderread 2026-10-18

The agent store holds hundreds of prebuilt agents filed by function and task, with sales, customer service and finance the largest groups and procurement, legal, operations and marketing behind them, covering jobs such as lead enrichment, RFP responses, ticket assignment and escalation, rebate auditing, regulatory monitoring and document comparison.

Prebuilt agents deploy from Builder and then take a team's own input sources, knowledge and model settings. A prompt library supplies templates to agents and apps, and packaged solutions cover customer support, sales copilot work, due diligence research and candidate profiling.

Sourcezbrain.ai/agentsread 2026-10-18

Each agent and crew watches a queue fed by input sources that fire on events: a new Gmail message, a file in Amazon S3 or Google Drive, a new Google Sheets row or Google Forms response, a Jira issue, a Notion database item, a Salesforce record created or updated, a ServiceNow record, a Zendesk ticket, a Discord message or new Zoho data.

Flows also start from a catch webhook protected by basic or header authentication, a synchronous webhook that returns a response, or human input. Agents can be shared by public or private link with an operator dashboard, and apps built on them publish as chat interfaces backed by the same agents.

Sourcedocs.zbrain.ai/zbrain-documentation/zbrain-builder/zbrain-ai-agents/agent-crew/set-up-input-sources.mdread 2026-10-18

Administrators register language and embedding models from OpenAI, Azure OpenAI, Google Gemini, Groq, Bedrock Claude or a custom endpoint, set any of them as the default, and supply their own API keys. Inside a crew, each agent can run on its own model, which overrides the crew default and lets teams match cost and quality to each role. Builder is model agnostic, with routing and switching between models such as GPT 5.4, Claude 4.6 and Gemini 3, and using a customer's own key avoids ZBrain's 20% markup on model usage.

Sourcedocs.zbrain.ai/zbrain-documentation/zbrain-builder/settings.mdread 2026-10-18

The REST API lists agent templates, lists, updates, deploys and tests a tenant's agents, and manages apps, knowledge bases, automated reasoning and prompts, so agent operations can run from a team's own code. Crews attach ZBrain MCP servers and external MCP servers side by side, the A2A protocol supports coordination with outside agents, and agents built in Azure AI Foundry or AWS Bedrock import into a crew. Flows import from JSON, any Flow can be called through its webhook URL, and custom code tools extend agents with npm dependencies.

Sourcedocs.zbrain.ai/zbrain-documentation/zbrain-builder/api-tutorials/agents.mdread 2026-10-18

Monitor schedules evaluations of deployed agents and apps at set intervals against expected outputs, using deterministic checks such as health, exact match, F1, Levenshtein similarity and ROUGE L, and LLM judged metrics such as response relevancy and faithfulness, then alerts the team on failures.

Each Flow step can be tested on its own before publishing, crews test as one coordinated unit, Flows and prompts keep version histories that can be restored, and the Prompt Manager tests prompts with usage analytics. Retrieval testing and the automated reasoning playground show how knowledge and rules respond to sample queries.

Sourcedocs.zbrain.ai/zbrain-documentation/zbrain-builder/monitor.mdread 2026-10-18

Flows can pull content from the web while they run. The Utilities component scrapes a web page as one of its data operations, and the Apify component runs Apify's scraping actors on that third party service and brings the resulting datasets into the Flow. Agents add Google and Bing search, a deep research tool and HTTP requests.

All of these fetch and extract content. None of the 159 Flow components or the default agent tools opens a browser that an agent clicks through, fills in forms on or signs in to, and no component controls a desktop, so a portal with no API or connector stays outside what an agent can operate.

Sourcedocs.zbrain.ai/llms.txtread 2026-10-18