Emergence AI

A one-sentence summary

Emergence AI is a company specializing in enterprise intelligence and AI agents; its core product, CRAFT, enables organizations to carry out data readiness assessments, metadata enhancement, natural language analysis, and constrained multi-agent workflows within their own cloud environments, data centers, or hybrid setups.

Tool Introduction

Emergence AI is the business name used by Merlyn Mind, Inc.; it is aimed at organizations that need to integrate AI into their corporate data and decision-making processes. CRAFT is not a one-click chat tool for individuals, but rather an enterprise-level system composed of solutions, shared platform services, and Kubernetes infrastructure.

The product focuses on running Agents within constraints, policies, and verification mechanisms, and it is used for root cause analysis, anomaly detection, scenario simulation, as well as reliable queries of natural language data. Deployment can take place on infrastructure under the customer’s control, which facilitates the organization and management of data, policies, models, and execution boundaries.

Composition of CRAFT products

The official documentation clearly labels the CRAFT Toolkit as a feature that is still in the planning stage; therefore, it cannot be considered available on a widespread basis at present. Custom data connections are part of the CRAFT data connection framework, and they do not rely on the Toolkit, which has not yet been released.

Three-layer technical architecture

Main functions

Data readiness assessment

CRAFT Assess evaluates data quality, coverage, and compliance with relevant policies, identifying the gaps that an organization needs to address before deploying autonomous agents. It helps transform the vague statement \"the data isn’t ready yet\" into a list of specific issues that can be tracked.

Metadata and data quality improvement

CRAFT Enrich can automatically add metadata, classify data assets, and generate data quality rules and scoring cards. The workflow is orchestrated by Agents and Prefect, but the rules still need to be reviewed by the data owners to prevent the model from misinterpreting the meaning of fields.

Verified Talk-to-Data

Data Insights allows users to ask questions of enterprise databases using natural language; the system generates SQL queries, performs verification, and returns answers, charts, or insights. This solution emphasizes pattern recognition and verification – a fluent response does not necessarily mean that the query is correct.

• Business questions can be raised through dialogue; it is not required that every user be able to write SQL.
• Generate candidate queries by combining the database schema with business rules.
• Validate the structure and results before and after execution.
• Retain multiple rounds of context to continue asking questions or adjusting the scope of analysis.
• Generate interactive charts, dashboards, and shareable results.

Root cause analysis and anomaly detection

Agents can search across multiple enterprise data sources to identify anomalies and trace potential root causes, making them suitable for troubleshooting in operations, manufacturing, and data pipelines. The analysis results are still influenced by data integrity, time range, and business definitions; high-risk decisions require confirmation from domain experts.

Scenario simulation and strategic analysis

The platform can create scenarios based on certain assumptions and compare the possible outcomes, thereby allowing for the testing of plans before taking action. Simulations rely on input data, rules, and model assumptions, and they cannot replace actual market, engineering, or regulatory validation.

Multi-agent orchestration

CRAFT makes use of the A2A protocol to coordinate multiple specialized agents, and it exchanges tasks and status information through JSON-RPC 2.0 as well as server-side event streams. Its stateful pipeline capability supports multi-step workflows and collaborative cancellation, making it suitable for assigning planning, querying, analysis, and verification tasks to different components.

Model provider decoupling

The platform connects to models such as Gemini, GPT, and Claude through LiteLLM, thereby reducing the dependence of solutions on any single vendor. Although these models can be replaced, their output is not always identical; therefore, it is still necessary for teams to evaluate factors such as data terms, costs, context, tool usage, and quality separately.

Asset management

The Asset service provides a unified mechanism for registering data connections, Agents, models, files, and outputs, with separation by organization and project. The Agent Card allows for the verification of protocol version, capabilities, interfaces, and security details before persistence.

Enterprise identity and fine-grained authorization

CRAFT utilizes OIDC and PKCE to manage multi-tenant identities and supports single sign-on. Permissions are inherited among organizations, projects, and resources through the OpenFGA relationship model, allowing for the distinction between roles such as viewers, developers, operators, and administrators.

Key and data connection

Connection metadata is stored in the asset database, while credentials such as usernames and passwords are saved in Infisical or an external secret management system via the Secrets API. These credentials are not included in the responses from regular APIs; they are inserted only when a connection to the solution is established.

Observability

The platform can collect traces, metrics, and logs using OpenTelemetry, and it can monitor the infrastructure with the Grafana LGTM stack. Langfuse is an optional tool used for tracking model calls, as well as for analyzing prompt versions, performance, and costs; these two functions serve different purposes related to observability.

Automatically generate SDK

CRAFT can generate Python and TypeScript SDKs in accordance with the OpenAPI specifications, thereby enabling solutions to call platform services. These SDKs are intended for use in deployed environments as well as for collaborative development; this does not mean that all enterprise product endpoints are available to the public without any restrictions.

Business-oriented solutions

Data analysis workflow

1. Identify the business issues to be analyzed, the person responsible for the data, and the scope of data that can be used.
2. Create organizations and projects in CRAFT, and configure identity and project permissions.
3. Create read-only accounts for the database, restricting access to only the necessary schemas and tables.
4. Register the connection through the Asset service, and pass the credentials to the Secrets API for storage.
5. Test the network, permissions, and query connections, then have Data Insights read the schema.
6. Pose small-scale questions in natural language to check the SQL code and business definitions generated by the system.
7. Manually verify the answers, charts, time windows, and aggregation methods.
8. Save the validated metric definitions and gradually expand them to more users and data sources.
9. Continuously monitor model invocation, query performance, permissions, and error logs.

Key points for connecting to PostgreSQL

• Prepare the database address that can be accessed via the CRAFT network.
• Create read-only users and grant them only the minimum permissions required for the target schema.
• Register the data connection using a valid identity token and project identifier.
• Have the connection metadata and credentials go into the Asset and Secrets services respectively.
• It is confirmed that the original credentials will not be included in the read response.
• Use a minimal query to test the connection before passing it to Data Insights.

Data governance workflow

1. The inventory process requires evaluating data sources, catalogs, owners, and the scope of oversight.
2. Use CRAFT Assess to identify gaps in quality, coverage, and strategy.
3. Sort the issues based on business impact and Agent usage risks.
4. Run Enrich to supplement asset descriptions, classifications, and candidate quality rules.
5. Let the data owners review the semantics, sensitivity level, and thresholds inferred by the model.
6. Apply rules through workflows to generate quality scorecards.
7. Assign failure rules to the responsible teams and track the resolution.
8. Regular re-evaluations are carried out to prevent drift caused by changes in patterns, data, and policies.

Solution development process

1. Select the reference solution or Starter template that best meets the requirements.
2. Create a FastAPI service and register the solutions and Agent assets as specified.
3. Integration with Keycloak for authentication, OpenFGA for permissions, as well as Secrets and shared storage.
4. Configure the allowed models and budgeting strategies through LiteLLM.
5. Define A2A capabilities, input/output parameters, and validation rules for the Agent.
6. Perform the minimum service tests locally, and then package them as containers and Helm Charts.
7. Deploy to development, testing, staging, and production environments via GitOps.
8. Integrate with OpenTelemetry and optional Langfuse to monitor end-to-end operations.
9. Establish permissions, data, models, and regression tests before making them available to business users.

The official development guidelines set as a goal for the documentation the completion of a minimal Quickstart within 30 minutes, but a full enterprise solution still requires the involvement of teams responsible for infrastructure, identity management, data handling, modeling, security, and operational maintenance. The public website displays only content that falls within the security boundaries of partners, and some of the implementation details are considered confidential documents.

Deployment method

CRAFT can be deployed in the customer’s own cloud, data center, or hybrid environment, and it requires to run on a Kubernetes cluster that complies with CNCF standards. The architecture strives to avoid reliance on any single cloud, making use of Helm, ArgoCD, Terraform, as well as alternative object storage and key management solutions.

Dependencies of the underlying platform

• Kubernetes and Helm are responsible for container orchestration and deployment.
• ArgoCD is used for GitOps delivery in development, testing, demonstration, and production.
• Keycloak handles identity and single sign-on.
• OpenFGA manages the permissions regarding the relationships between organizations, projects, and resources.
• PostgreSQL, Redis, and object storage platforms are used to store platform status and events.
• Infisical or external secret management is used to store connection credentials.
• The OpenTelemetry and Grafana stack provide tracking, metrics, and logging.

Price and procurement methods

As of August 22, 2026, Emergence AI’s official website does not disclose the fixed price of the CRAFT package, nor the pricing per seat or the limits of the free version. The product is offered through corporate partnerships and customized quotes; potential customers must submit information regarding their business needs, data requirements, deployment specifics, and security concerns in order to receive a proposal.

Preparations before requesting a quote from a company

1. List the specific business problems; do not start with vague, general Agent requirements.
2. Statistics source, database type, asset size, and update frequency.
3. Determine the boundaries for customer cloud, on-premises, hybrid, or isolated deployments.
4. Specify the permitted models, data regions, sensitivity levels, and compliance requirements.
5. Estimate the scale of users, queries, concurrency, Pipelines, and model calls.
6. List the identities, keys, directories, and observability systems that must be connected.
7. Define accuracy rates, SQL security, data quality, and business acceptance metrics.
8. The quote is requested to list separately licensing, implementation, infrastructure, models, and support.

Which users are it suitable for

• Enterprise Data Officer: Assess whether the data is suitable for use with Agents and natural language analysis.
• Data governance team: Automatically adds metadata, classifies assets, and creates quality rules.
• Business and analysts: Use natural language to query data and create visualizations.
• Data Engineering Team: Identifies abnormalities in data pipelines and tracks potential root causes.
• Platform Engineer: Deploys and maintains enterprise Agent platforms in a Kubernetes environment.
• Security and Compliance Team: Controls data access through identity, permissions, keys, and audit mechanisms.
• Semiconductors and manufacturing organizations: Developing isolated analysis solutions for domain-specific data.
• Joint design partners: Develop customized Agent applications on CRAFT shared services.

Typical use cases

• Before deploying the enterprise agent, evaluate data integrity, ownership, and policy gaps.
• Automatically generate descriptions, classifications, and candidate quality rules for data catalogs.
• Allow sales staff to query sales, operations, or manufacturing databases using natural language.
• Convert query results into charts, dashboards, and shareable insights.
• Orchestrate processes such as data retrieval, analysis, verification, and reporting across multiple agents.
• Detect anomalies in the data pipeline and trace them back to the source table, rules, or upstream systems.
• Compare the outcomes of scenarios under different assumptions before taking strategic actions.
• Deploy industry-specific Agent applications in a controlled internal network or hybrid environment.

Product advantages

• It covers data readiness, governance, analysis, and Agent workflows, rather than just providing a chat interface.
• It can be deployed in customer clouds, data centers, or hybrid infrastructures, thereby reducing the pressure associated with transferring data outside.
• Based on Kubernetes and common open-source components, it avoids reliance on single cloud-specific services.
• Unified identity, permissions, and credentials are managed through Keycloak, OpenFGA, and Secrets API.
• It supports multiple model providers, allowing teams to choose models based on security, performance, and cost.
• A2A and stateful pipelines are suitable for breaking down complex, multi-step tasks.
• Natural language to SQL conversion, combined with pattern recognition and validation mechanisms, facilitates the creation of auditable processes.
• OpenTelemetry and Langfuse cover infrastructure monitoring and model invocation monitoring, respectively.

Usage restrictions and precautions

• The product is designed for enterprise deployment and requires Kubernetes, identity management, data handling, and operational capabilities.
• The official website does not provide standard prices; the costs associated with pilot implementations and full-scale deployment must be evaluated separately.
• The CRAFT Toolkit is still a planned module and cannot be purchased as a feature that is already available.
• Natural language generation of SQL can still misinterpret metrics, times, relationships between tables, or business definitions.
• Multiple agents increase model calls, latency, failure points, and the complexity of debugging.
• The fact that models can be replaced does not mean that different models are entirely equivalent in terms of quality and compliance.
• Automatically generated data quality rules may result in false positives, false negatives, or unreasonable thresholds.
• Self-hosting provides greater control, but it also entails increased responsibilities regarding patches, backups, monitoring, and capacity.
• The availability of public development documentation does not mean that all internal services and code are accessible to the public.
• The licenses of various repositories on the official GitHub vary, so it cannot be said that they are all completely open source.
• DPA restricts sensitive data that has not been designated as such; organizations must first carry out data classification.

Security and Governance

Pre-launch governance inspection

1. Determine the owner, purpose, sensitivity level, and permitted users for each data source.
2. Create read-only accounts for analyzing databases, and restrict the schemas, tables, and networks.
3. Separate human users, service accounts, and Agent permissions to avoid sharing accounts with high privileges.
4. Set timeouts, row counts, costs, and restrictions for SQL operations.
5. Define definitions and test cases for manual approval of business metrics.
6. Check for personally and commercially sensitive data in prompts, SQL queries, results, and logs.
7. Set approval, rollback, and auditing for Agent actions that have side effects.
8. Regularly verify authorized inheritance, key rotation, backup restoration, and model changes.

Privacy, DPA, and Compliance

The website’s privacy policy applies primarily to visitors of the official site, whereas the processing of data in corporate services is governed by the customer agreements and the DPA issued in 2025. The DPA specifies that Emergence, as the processor, may receive and store API calls, prompts, model outputs, as well as technical and usage data, and it is required to process and delete this data in accordance with the customer’s instructions.

The Trust Center lists GDPR, SOC 2 Type 1, SOC 2 Type 2, and SOC 3, and provides information on security features such as multi-factor authentication, single sign-on, and encryption for both data at rest and during transmission. Specific reports and sensitive information generally require access requests through the Trust Center; purchasers should verify that the audit scope, dates, and the services purchased are covered by it.

GitHub and open-source projects

Emergence AI maintains an official GitHub organization, but the repositories there address different issues and come with varying licenses. According to the CRAFT documentation, the core platform components are designed to be open-source and can be hosted independently; however, as of the date of verification, no single public repository containing the complete source code for all CRAFT enterprise products has been found.

The differences between Agent-E and CRAFT

Agent-E was an early web automation agent project made available to the public; it used a browser to handle forms, searches, and various website operations. CRAFT is a modern platform for enterprise data and intelligence. Both projects are maintained by the same official organization, but their product purposes, methods of deployment, and licensing terms cannot be considered identical.

APIs and developer capabilities

Basic information

Recommendation score

It receives a rating of 4.4 out of 5 points. CRAFT integrates data readiness, natural language analysis, governance, multi-Agent capabilities, permission controls, key management, and observability into a Kubernetes deployment that can be controlled by the customer; it is suitable for large organizations with high requirements regarding data boundaries and reliability.

The main obstacles are the lack of standard public pricing, the high complexity associated with deployment and management, and the need to clearly distinguish between planned functions and those that are currently available. It is more suitable for companies that have teams dedicated to platform development, data governance, and security; it is not a lightweight product designed for individual users to try out agents quickly.

Frequently Asked Questions

What is Emergence AI?

Emergence AI is a company that develops enterprise intelligence and AI Agent systems; its core product, CRAFT, is used for data preparation, governance, natural language processing, and controlled multi-Agent workflows.

What is CRAFT?

CRAFT is an intelligent platform that can be deployed within enterprise infrastructure; it consists of business solutions, shared services such as those related to identities and assets, as well as a Kubernetes infrastructure.

What is the difference between CRAFT Assess and Enrich?

Assess is responsible for identifying issues related to data quality, coverage, and policy gaps, while Enrich is tasked with adding metadata, categorizing assets, and creating quality rules. The former focuses on diagnosis, whereas the latter is aimed at improvement.

Is the CRAFT Toolkit available yet?

No. The current official documentation explicitly designates the Toolkit as a module that is still in the planning stage; therefore, verification certificates and automatic formalization tools cannot be described as being readily available.

Is it possible to query a database using natural language?

Yes. Data Insights offers natural language to SQL conversion, multi-round analysis, and visualization, but the query logic, business metrics, and results still need to be verified.

Which databases are supported?

The public introductory documentation provides detailed information on connecting to PostgreSQL, and its architecture allows for the addition of additional data connections as needed. Other specific databases should be evaluated one by one during project assessment.

Does CRAFT support private deployment?

It can be deployed in the customer’s cloud, data center, or hybrid environment, and runs on Kubernetes clusters that comply with CNCF standards. The specific requirements regarding offline operations, images, and support boundaries must be specified in the contract.

Can I choose different large models?

It is possible to access providers such as Gemini, GPT, and Claude through LiteLLM. Enterprises need to configure keys, data policies, budgets, and quality tests separately.

Is Emergence AI free?

No free public version or fixed trial quota of CRAFT was found; the product is primarily acquired through business partnerships and customized quotes.

How much is CRAFT?

The official website does not specify a fixed price. The cost can be influenced by factors such as modules, data sources, users, the amount of model usage, the deployment environment, as well as the scope of implementation and support; it is necessary to contact sales for more information.

Are APIs and SDKs available?

It provides platform APIs, and Python and TypeScript SDKs can be generated in accordance with the OpenAPI specifications. Actual access requires a deployed environment, an authentication token, and project permissions.

Is Emergence AI open-source?

It cannot be simply answered as being entirely open source. Projects such as Agent-E come with public licenses, and CRAFT utilizes open-source components that can be self-hosted; however, the complete enterprise platform still includes commercial products and controlled content.

What is Agent-E?

Agent-E is a web automation agent project under the MIT license, made available by Emergence AI; it allows browser tasks to be executed using natural language. It is not the CRAFT enterprise platform itself.

Can Emergence World be used for commercial purposes?

By default, it is not allowed. This research repository uses the CC BY-NC 4.0 license; it permits use for non-commercial research and educational purposes provided that attribution is given, while commercial use requires separate permission.

Will the data leave the corporate environment?

CRAFT can be deployed on infrastructure under the customer’s control, but external models, telemetry, support services, and sub-processors may still generate data streams. Enterprises should verify each element based on the actual architecture and the DPA.

Is it suitable for small teams?

If a small team already possesses Kubernetes capabilities as well as expertise in data and security, and if its requirements are complex enough, it can consider conducting a pilot project; generally, individuals or small teams are better suited for using analysis tools that require fewer setup requirements.

Summary

CRAFT from Emergence AI is designed for organizations that wish to enable agents to use corporate data in a reliable manner; its main advantage lies in the fact that it integrates data evaluation, metadata enhancement, natural language analysis, multi-agent orchestration, and governance controls within a single deployable framework.

When making a selection, it is necessary to start with the specific business challenges and data boundaries, and to verify the accuracy of SQL queries, permission isolation, the flow of model data, operational costs, and the acceptance criteria. It is also important to clearly distinguish between the Assess and Enrich tools that are already available, the Toolkit project that is still in the planning stage, and independent open-source research projects such as Agent-E.