ZGI Opens Its Source: Models, Knowledge, Skills, and Workflows in One Agent Runtime

Building an AI agent is easier than it was two years ago. Connect a language model, write a prompt, attach a few documents, and a conversational demo can be ready quickly.

The difficult part comes afterward.

Sales wants the agent to query CRM. Operations wants automatic reports. Engineering wants access to internal APIs. Management starts asking about permissions, logs, token costs, and data security. Once AI enters the business, “build an agent” becomes an infrastructure problem.

This is one reason ZGI has opened its source code.

ZGI is an Agent Runtime for enterprises. It brings models, knowledge bases, data, Skills, workflows, and runtime governance into one environment, helping teams move from a working demo toward an agent system that can operate in real business.

1. Why enterprise agents eventually need a shared foundation

Many enterprise AI initiatives begin in a similar way.

Support builds a knowledge bot. Engineering connects to the Claude API. Operations creates a DeepSeek workflow. Another department uses a private model.

At first, each project works.

Six months later, the problems emerge: numerous model configurations, API keys scattered across projects, duplicate knowledge bases, and workflows understood only by their original creators. When an agent fails, it is hard to tell whether the model, data, or tools caused it.

More AI capabilities can create new silos.

ZGI starts by bringing these scattered resources into one workspace. Models, agents, knowledge bases, Skills, workflows, API keys, and execution records can be managed together over time.

Enterprises need to turn AI capabilities into reusable organizational assets.

2. ZGI goes beyond an agent builder

There are already many agent builders that let teams connect a model, arrange a few nodes, and produce an application.

ZGI emphasizes Agent Runtime because executing real tasks requires a complete operating process.

A sales lead agent, for example, might read an email, query CRM, assess intent against past interactions, and update the system. High-value leads may need a salesperson's attention. Special pricing or contracts may require automatic execution to pause for confirmation.

This involves models, company data, tool calls, workflows, and permissions together.

In ZGI, Model Gateway connects GPT, Claude, DeepSeek, Qwen, Gemini, Ollama, and internal models. Knowledge bases and databases supply business context. Skills perform actions such as creating files, querying data, and invoking tools. Workflows organize the sequence around business rules.

The goal is to move from AI knowing what should happen to AI carrying the task through.

Conceptual illustration: ZGI brings models, knowledge, Skills, workflows, and governance into one environment.

3. Models can change without rebuilding company capabilities

Models evolve quickly. A task that suits Claude today may suit GPT later. Batch processing may favor DeepSeek on cost, while sensitive work requires an internal private model.

When an application is tightly coupled to one model, switching providers can force changes to the business logic.

ZGI manages models centrally to separate model capabilities from business processes as far as possible. Agents and workflows can remain stable while the model underneath changes according to business needs.

This may seem minor in a demo. It becomes important in production.

The assets an enterprise wants to retain are its business logic, knowledge, Skills, and workflows, rather than dependence on one model API.

4. Skills turn capabilities into reusable assets

Skills are an important part of ZGI. They package specific capabilities so multiple agents can invoke them repeatedly.

A company might create a daily business report Skill that queries a database, calculates results, produces charts, and outputs a standard report. Another Skill might retrieve customer details from CRM and internal systems. Existing business tools can also be exposed as capabilities agents can call.

These functions previously tended to live in scripts, prompts, or someone's computer. Packaged as Skills, they can be reused over time.

Developers can extend them for their own business rather than waiting for a platform vendor to add a feature.

Models offer general capabilities. Skills are closer to a company's own capabilities.

5. Workflows bring agents into actual business processes

Real enterprise processes rarely consist of one question and one answer.

Quality checks may require reading materials, identifying problems, retrieving rules, assessing risk, generating results, and updating a business system.

Support may need to identify an issue, retrieve knowledge and order data, answer routine questions, and send higher-risk cases to people.

Initial recruitment screening likewise involves reading resumes, extracting information, comparing it with role requirements, preparing assessments, and identifying matters that require an interviewer's judgment.

These processes need stable workflows.

ZGI combines model calls, knowledge retrieval, conditions, loops, HTTP requests, databases, code execution, and tool calls in one flow.

AI can then participate in business operations rather than remaining a standalone chat page.

6. More capable agents make governance more important

Teams initially focus on results: Are the answers accurate? Can the task be completed?

After launch, other questions follow. Why did a task fail? Which model ran? How many tokens were consumed? Which node took longest? Who ran the agent? What data did it access?

At ten calls a day, these questions may seem manageable. At hundreds or thousands, logs, costs, permissions, and error tracing become essential.

ZGI therefore brings execution logs, token consumption, model use, node status, API keys, and permissions into runtime governance alongside agents and workflows.

Agents can become more proactive without companies losing sight of their actions.

7. Why source access and self-hosting matter

Deployment location may seem less sensitive when AI only writes copy. It becomes more consequential when agents access internal knowledge, customer data, databases, and business systems.

Enterprises need to understand the system and decide how data flows.

ZGI makes its source available and supports self-hosting. Teams can deploy the relevant services on their own infrastructure and connect their own models, knowledge bases, databases, and internal services.

This matters for enterprises that require private deployment, internal network operation, or data isolation.

ZGI currently uses the ZGI Community License. Personal, research, educational, and internal organizational use can be free. Hosted multitenant offerings, white-label business models, and other covered commercial uses require the appropriate commercial authorization under the license.

Rather than claiming unrestricted free use, we want to describe the value precisely: teams can inspect the source, deploy the system themselves, and retain control of their AI capabilities.

8. Which teams should try ZGI?

A general AI chat tool may be enough for occasional questions.

ZGI becomes more relevant when an enterprise uses multiple models, connects internal knowledge and databases, gives agents tools and workflows, and needs to manage permissions, tokens, logs, and private deployment.

Support, sales, operations, engineering, HR, and internal knowledge teams can all begin with a small use case.

Build a daily report workflow, create a knowledge assistant, or delegate a routine data retrieval and organization task.

ZGI currently offers a way to start exploring for free. Run one real task first, then decide whether to expand.

A foundation enterprises can understand and control

Discussions of open-source AI often focus on whether model weights are available.

As agents enter enterprises, access to the runtime matters too. Companies need an environment they can understand, deploy, extend, and govern.

Models determine whether AI can reason. Knowledge and data determine whether it understands the business. Skills determine whether it can act. Workflows determine how those capabilities cooperate. Runtime determines whether everything can keep operating.

This is the connection ZGI wants to make with opening its source: giving enterprises a runtime foundation they can hold in their own hands, rather than simply another AI demo.

GitHub: https://github.com/zgiai/zgi

Website: https://www.zgi.cn/

Documentation: https://docs.zgi.ai/

If enterprise AI is going to become part of real business, the layer that runs it should be sufficiently open as well.