Open Source AI Agent Platform: A Team Selection Guide

By Kortix Blog

An open source AI agent platform should be evaluated by the rights its license grants and the work your team needs to govern—not just whether its repository is public. The Open Source Initiative explicitly distinguishes access to code from open-source licensing. This guide compares two MIT-licensed frameworks with three source-available alternatives, so you can choose without confusing inspectable code, self-hosting, and unrestricted use. Open Source Definition

Documentation reviewed: September 16, 2026. This is a documentation-based selection guide, not a hands-on benchmark or security certification.

Short answer: choose by workload and operational ownership

Start with the layer you want to own. For this comparison, a platform means a product that brings together an operating interface and services around agents; a framework means building blocks your developers compose into an application. LangGraph describes itself as low-level orchestration infrastructure, while CrewAI provides Python abstractions for Crews and Flows. Neither framework's license should be treated as the license or feature entitlement for its associated commercial platform. LangGraph overview · CrewAI repository

Our selection recommendation is to shortlist Kortix for repository-backed team work, LangGraph for custom stateful orchestration, CrewAI for role-based multi-agent workflows, Dify for visual LLM applications and retrieval pipelines, and NocoBase for AI inside business applications. These are workload recommendations based on documented designs, not rankings of quality, speed, or security. Kortix · LangGraph · CrewAI · Dify · NocoBase

A personal assistant is a different selection problem from a shared team system. Use this guide to evaluate organizational operation; for the narrower positioning discussion, see personal AI agents versus a company OS. That discussion is not evidence for the third-party capabilities compared here.

Open source, source available, and self-hosted are different

Open source concerns license rights, including redistribution and use without discrimination against fields of endeavor. Source available describes inspectable code whose license can impose restrictions outside that definition. Self-hosted describes who runs software, not which rights the license grants. A public repository or a Docker installation does not settle all three questions. Open Source Definition · Kortix self-hosting

The current LangGraph and CrewAI framework licenses are MIT. Kortix's current repository instead carries Elastic License 2.0, including a restriction on providing a hosted or managed service that exposes a substantial set of its features, and restrictions on circumventing license-key functionality. Although Kortix's product copy says “open source,” this guide classifies that repository as source available. LangGraph license · CrewAI license · Kortix license

Dify's license is modified Apache 2.0, requiring commercial permission for the specified multi-tenant use and restricting changes to frontend branding. NocoBase's current agreement incorporates Apache 2.0 with overriding supplementary terms, including restrictions on public no-code, low-code, and AI-platform SaaS/PaaS offerings. NocoBase distinguishes Community from Standard, Professional, and Enterprise commercial editions. Treat both current distributions as source available here, not plain Apache-licensed software; review the exact intended use with your legal team. Dify license · NocoBase agreement

Compare the shortlist

The matrix is split into two tables for readability. Unknown means the reviewed documentation does not establish that capability or entitlement for the specified scope; it does not mean the product lacks it. Model choice means a documented integration path, not equivalent behavior across models. Evidence was reviewed on September 16, 2026; the linked licenses and documentation can change.

Workload, license, deployment, and model choice

Option / best-fit workload recommendation	License and edition boundary	Deployment responsibility	Model choice	Evidence / reviewed
Kortix: shared agents producing repository-backed work	Elastic 2.0, source available; some governance controls require an enterprise key	Self-hosted control plane via Docker Compose; sandbox provider still required; managed cloud also offered	Bring provider credentials to the self-hosted gateway	Repository, license, deployment · 2026-09-16
LangGraph: custom stateful orchestration	MIT framework; do not extend that license to LangSmith services	Standalone framework runs in your application; LangSmith Deployment is a separate option	LangChain provider integrations can be used; LangGraph does not require LangChain	Overview, license, models · 2026-09-16
CrewAI: role-based Crews and event-driven Flows	MIT framework; AMP is a commercial control plane	Run the Python framework yourself or evaluate AMP deployment separately	Multiple provider SDK integrations; other providers through LiteLLM	Repository, license, models · 2026-09-16
Dify: visual LLM apps and retrieval pipelines	Modified Apache 2.0, source available; Community, Cloud, and enterprise offerings are distinct	Community Docker Compose is operator-run; Cloud is hosted	Provider and self-hosted model integrations documented	Repository, license · 2026-09-16
NocoBase: AI-enabled business applications	Apache-based agreement with supplementary restrictions; Community and commercial editions	Docker Compose application and database deployment under your operation	Local OpenAI-compatible inference documented for the Lina localization scenario	Repository, license, deployment, local model example · 2026-09-16

Shared knowledge, permissions, isolation, and review

Option	Shared knowledge	Permissions	Isolated execution	Human review	Evidence / reviewed
Kortix	Shared skills and company memory in a project repository	Resource permissions documented; self-host SAML, SCIM, custom roles, groups, and audit read require an enterprise key	Per-session sandbox; containers by default, microVMs with Platinum; provider-dependent	Change requests for repository work; configurable action approval gates	Repository and security section, self-host edition details · 2026-09-16
LangGraph	Short- and long-term memory primitives; team access design remains application work	Agent Server auth handlers are documented separately; self-hosted server has no default authentication	Per-session execution sandbox: unknown from reviewed framework/server docs	Interrupts pause execution for approval or edits; require checkpointing and application wiring	Overview, server access control, interrupts · 2026-09-16
CrewAI	Agent- and crew-level knowledge sources	AMP documents feature and entity RBAC; equivalent organization controls in the MIT framework: unknown	Current agent docs deprecate built-in code-execution settings; replacement isolation guarantee: unknown	Task configuration exposes human_input; do not assume a universal pre-action gate	Knowledge, AMP RBAC, agents, tasks · 2026-09-16
Dify	RAG pipeline; Cloud knowledge access configured separately from workspace role	Cloud documents Owner, Admin, Editor, Normal; Community parity: unknown in reviewed permission docs	Cloud Code node documents Python/JavaScript sandboxes; whole-agent or Community-equivalent isolation: unknown	Cloud Human Input node supports review branches; self-host config also references the node; email links allow any holder to respond	Repository, Cloud members, Code node, Human Input, self-host configuration · 2026-09-16
NocoBase	AI Knowledge Base plugin provides RAG; edition entitlement unknown in reviewed plugin page	Built-in data queries follow current-user permissions; custom workflow tools use independent permission logic	Per-session execution sandbox: unknown in reviewed deployment/AI docs	Lina translation example requires separate publishing after review; universal action approval: unknown	Knowledge plugin, permissions, deployment, Lina example · 2026-09-16

Which option fits your team scenario?

Shared company knowledge and controlled business tools

Kortix is an AI management system that keeps agents, shared skills, memory, and connectors around a versioned project repository. Shortlist Kortix when reviewable repository work is your desired unit of collaboration. Do not choose Kortix on the assumption that self-hosting means a disconnected installation: its self-hosting page requires a sandbox provider and says air-gapped topologies are scoped separately. The same page makes enterprise-key requirements explicit for several governance features. Kortix repository · Self-hosting details

Dify is an LLM application development platform combining visual workflows, retrieval, and model integrations. Shortlist Dify instead when the main deliverable is a workflow-driven application over documents, rather than a repository change. Before choosing Dify for restricted data, establish the exact edition's knowledge visibility and reviewer identity rules: Cloud workspace roles and possession of an emailed review link answer different access questions. Dify repository · Members · Human Input

Custom developer orchestration

LangGraph is a low-level framework for stateful workflows; CrewAI is a Python framework built around collaborating Crews and event-driven Flows. Prefer a LangGraph evaluation when explicit state and custom control flow are central, or a CrewAI evaluation when role-based collaboration matches your design. In either case, budget separately for the surrounding application and deployment controls rather than assuming an MIT library includes its vendor's commercial management plane. LangGraph overview · CrewAI repository

Do not choose Kortix merely to obtain orchestration primitives if your team already owns the application interface and wants to embed a framework. Conversely, do not choose a framework merely because its license is permissive if nobody will own the surrounding system. These are architectural recommendations; for the separate managed-runtime decision, read the OpenAI Agents API for teams.

Business applications and data permissions

NocoBase is an AI and no-code business-system platform with data models, workflows, and a configurable interface. Shortlist NocoBase when agents need to act within a business application, but examine each tool's permission path: the official AI permissions guide says built-in queries follow user permissions while custom workflow tools can follow independent business logic. A role assigned to an AI employee is not proof that every custom tool preserves the same data boundary. NocoBase repository · AI permissions guide

A bounded platform evaluation

The following is a proposed buying exercise, not testing we performed. Choose two candidates from the relevant layer and request the same evidence from each.

Define one representative task

Write a task such as “prepare a customer briefing from approved records and propose a CRM update.” Specify the allowed records, intended output, prohibited actions, and person authorized to approve the update. Reject a demonstration that substitutes unrestricted sample data for the permission problem you need to solve.

Request proof of permission boundaries and isolation

Ask for an authorized and an unauthorized user to attempt the same retrieval and action. Request the exact edition, permission configuration, execution provider, and network boundary. Do not count a deployment container as proof of per-session isolation; require documentation for the actual execution design and a demonstration of the denied path.

Inspect completed-work evidence and review gates

Require a reviewable output, the records used to produce it, and a clear approve/reject step before the chosen external action. Test rejection, timeout, and a reviewer without authorization. Keep repository merge approval separate from approval to send a message or modify a business record; ask the vendor to show where each gate is enforced.

Document operating burden

Assign an owner for upgrades, backups, credentials, model access, and incident response before making the selection. List which services leave your network and which controls require a paid entitlement. After selecting a platform, use the reliable AI agent workflows guide for implementation controls rather than expanding this comparison into a reliability tutorial.

Frequently asked questions

What is the difference between an agent platform and a framework?

In this guide, a platform packages services and an operating interface around agents, while a framework supplies building blocks for your own application. LangGraph's low-level orchestration and CrewAI's Crews and Flows are framework examples. Evaluate the associated commercial deployment products separately rather than assuming their capabilities ship under the framework license. LangGraph · CrewAI

Does self-hosting mean everything runs locally?

No. Kortix's self-hosting documentation, for example, places the control plane on your infrastructure but requires a sandbox provider and separately scopes air-gapped topologies. For a local-only requirement, ask for a complete dependency map covering model inference, embeddings, execution, connectors, and updates—not merely the location of the dashboard. Kortix self-hosting

Are community and paid governance controls the same?

Do not assume parity. CrewAI documents organization RBAC in AMP, its commercial control plane, while Kortix's self-hosting page lists an enterprise key for SAML, SCIM, custom roles, groups, and audit read. Request the entitlement for each required control in the edition you will actually deploy. CrewAI AMP RBAC · Kortix self-hosting

Can a team change model providers later?

Several candidates document provider choice: LangChain integrations usable with LangGraph expose a common model interface, and CrewAI documents multiple native provider SDKs plus LiteLLM integrations. Treat those paths as integration options, not proof of identical outputs. Our recommendation is to repeat your acceptance task when changing models and check embeddings separately. LangChain models · CrewAI models · CrewAI knowledge

Does human review cover every action an agent takes?

The documented review mechanism matters. LangGraph interrupts pause where developers place them; Dify's Human Input node routes a workflow according to a response; Kortix documents change requests for repository work and configurable action gates. Ask for the exact gate before your sensitive action rather than treating any review feature as universal protection. LangGraph interrupts · Dify Human Input · Kortix repository

Next step: inspect the license and deployment you will use

Choose a workload first, then verify license rights, edition entitlements, and deployment dependencies before approving a pilot. For Kortix, start with the current license and official self-hosting documentation. Bring the same questions to every shortlisted vendor, and leave undocumented controls unresolved rather than scoring them as either present or absent.