Ten Self-Hostable Open-Source Agent Platforms, Sorted by What a Business Actually Needs

A small-business owner on Reddit described a familiar position. ChatGPT was already part of the daily routine, and the next step was getting AI to handle customer follow-up, sort incoming email, send reminders about outstanding proposals and produce weekly summaries. Research turned up n8n, Zapier, raw APIs, MCP and a swarm of agent builders. The owner had no development background and no wish to learn an entire technical stack for a few automations. The harder worry was upkeep: if building these workflows takes weeks, will they demand constant attention and end up more tangled than the manual process they replaced?

That worry is common among companies shopping for agent tooling, partly because the search term "AI agent" covers wildly different products. There are developer frameworks such as LangChain, CrewAI and Haystack. There are agent-building platforms, automation tools, RAG platforms, and enterprise application platforms that manage business data, pages, permissions and workflows together.

This roundup narrows the field to ten open-source projects that can be self-hosted and already ship some mix of visual app building, workflows, knowledge bases or business-system features. Each gets a short profile covering what it is, which AI capabilities stand out, what it does for a business and which teams it suits.

How the list was chosen

A project qualified if its core code is public and actively maintained, it supports self-hosting, it already offers agent, tool-calling or AI workflow features, and it provides at least one of visual applications, automation, knowledge bases or enterprise business systems. Descriptions rest mainly on each project's website, GitHub repository and official documentation. This is not a benchmark in a uniform test environment and not a complete ranking, so the groupings below are a map and not a scoreboard. Star counts are those shown on GitHub at the time of writing.

The ten in three groups

• AI tied to enterprise data, automation and business processes: a no-code enterprise application platform (profiled first), plus Activepieces, Kestra and Windmill.
• Visual building of agents and AI applications: Coze Studio, Langflow and Dify.
• Enterprise knowledge bases and RAG: MaxKB, AnythingLLM and RAGFlow.

A quick way to choose: a company that needs to manage structured business data, permissions and internal applications should look at that no-code application platform first. One whose main goal is wiring together many SaaS tools and automating chores has a strong candidate in Activepieces. Teams focused on building agents and AI workflows will find Dify, Langflow or Coze Studio good places to begin. Where private document Q&A and complicated knowledge retrieval matter most, MaxKB, AnythingLLM or RAGFlow fit better.

Group one: enterprise applications and business automation

A no-code platform for enterprise applications

The first project (about 23.9K GitHub stars) is an open-source, self-hosted no-code platform for enterprise applications. People mostly use it to build CRMs, ticketing systems, IT management tools, approval flows and other internal software. It supplies the foundations such applications need, namely data models, pages, permissions and workflows, and lets AI take part both in building the application and in adjusting it later. AI employees can also be configured inside the system, combining live business data, tools and processes to analyse, process information and carry out tasks.

Notable AI features:

• AI-assisted no-code building. A user describes a requirement in plain language and the AI creates the tables and relationships, the pages, the workflows and the role permissions. There are two entry points: a no-code portal, where business users keep tuning the application visually, and an AI portal, better for complex interactions and heavily customised pages.
• AI employees acting on the business. The AI takes on a defined role in everyday work, such as analysing data, organising email or extracting information. It can use the system's tools and workflows to query records, fill in forms or start business processes.
• An enterprise knowledge base. AI employees can pull company knowledge automatically or when asked, with the accessible content filtered by the current user's role. Support, sales or internal assistants can then work from the organisation's own material.
• Permissions and execution control. AI employees inherit the data permissions of the current user, and sensitive operations such as changing data can require confirmation first, which stops an agent from reaching beyond what the person behind it may do.

Typical uses include a CRM that tidies up details from customer emails or meeting notes, drafts follow-up suggestions and supports sales with customer, contact and opportunity data; support and ticketing, where the AI examines submitted problems, sets category and priority, drafts replies from the knowledge base and joins ticket workflows; IT management, where it infers request type, priority and handling advice from an employee's own description and works with asset, maintenance and approval data; and other internal systems for procurement, approvals, project management and operations, with AI employees configured per role.

It fits teams with clearly defined needs like CRM, support ticketing, IT management or approvals, who want AI to speed up building and not to invent a whole business system from nothing. Business users can finish most of the configuration with no-code tools and AI, then grow the system through APIs and plugins. It suits internal systems expected to live and evolve for a long time.

Activepieces

Activepieces (about 24.0K stars) is an open-source, self-hosted AI automation platform for linking the applications, data and processes a company already uses. It puts agents, automation workflows and data in one environment, which suits sales, support, operations and other work that crosses several tools.

Notable AI features:

• Visual agent creation. Define the agent's task, model and available tools, then place it inside an automation workflow. Existing workflows can also serve as agent tools.
• Building by description. State the goal in natural language and let the AI produce an agent or a workflow, then keep editing visually with conditions, loops and code steps.
• Business tool connections. More than 700 applications are supported, including CRMs, email, collaboration tools and spreadsheets. Through MCP, what gets built on the platform can be exposed to other AI clients that speak MCP.

Typical uses: sales lead handling, where new leads are captured from email or forms, classified by customer type and priority, written into the CRM and given follow-up text; customer support flows that read incoming issues, classify and summarise them, assign tickets and sync data between support tools and the CRM; and daily operations, such as sorting email, extracting details, syncing spreadsheets or databases and generating daily or weekly reports and notifications from several systems.

It is best for small and midsize teams or departments that already run several SaaS tools but have little development capacity. Sales, marketing, support and operations people who constantly move data, send messages and chase follow-ups across systems can start with visual workflows and add APIs or custom integrations as needs grow.

Kestra

Kestra (about 28.0K stars) is an open-source, self-hosted workflow orchestration platform for automation that spans data, AI and infrastructure. Agents run directly as tasks within a workflow, calling tools, running other tasks or triggering existing flows depending on their goals.

Notable AI features:

• Agents inside workflows. An AI copilot can write a new workflow, or revise one, from a plain-language request. Agents can also sit inside existing workflows, where they use models, memory, tools and existing Kestra tasks and flows, then choose the next step from what execution returned.
• Visual and code-based development. Build and adjust through the interface, and extend harder requirements with YAML, Python, Docker and custom scripts.
• Traceable runs. Agents and ordinary tasks share a runtime, so status, logs and failure records can be examined in one place, which eases troubleshooting and maintenance.

Typical uses: data-processing automation that picks next steps from the state of the data and calls existing workflows for cleaning, analysis and aggregation; IT and operations flows that analyse alerts or system events and then run checks, notifications or other follow-ups; and complex business and AI processes that combine models, tools, approvals, data handling and cross-system tasks needing several steps and dynamic decisions.

Kestra suits teams that already have an automation base and involve developers, data engineers or operations specialists. Its orchestration strengths pay off as workflows lengthen and need dependencies, retries, state tracking and multi-step AI decisions. A small team that only wants a few simple automations quickly may find a complete workflow system more than it needs.

Windmill

Windmill (about 17.7K stars) is an open-source, self-hosted platform for development and automation that runs scripts, APIs, workflows and internal apps in one place. It supports visual orchestration alongside code in Python, TypeScript, Go, SQL and other languages, which points it at teams with technical skills.

Notable AI features:

• Agents as workflow steps. An agent runs as one step and combines with conditions, approvals, error handling and other nodes for jobs that need AI judgement.
• Reuse of existing technical assets. Agents can use scripts, databases, APIs and other Windmill workflows already in place, and they can connect to MCP tools, so existing code and internal services are reused directly.
• Permissions, approvals and records. Teams can limit which resources an agent may touch or run, require human sign-off before significant actions, and inspect the tool calls, inputs, outputs, logs and status of every run.

Typical uses: IT and operations automation, where an alert or system state prompts analysis and then scripts run checks, repairs and notifications; data processing, where an agent examines the output of a database query or existing data script and triggers later workflows; and internal AI tools that mix agents, backend scripts and app interfaces into data assistants and operations tools.

Windmill fits companies with developers, data specialists or operations engineers, especially those with a pile of scripts, databases, APIs and internal services they want to reuse for agents. For a team without technical staff that wants a few simple visual workflows, a platform oriented toward development may feel heavy.

Group two: visual agent builders

Coze Studio

Coze Studio (about 21.5K stars) is ByteDance's open-source, all-in-one platform for developing agents. Its focus is creating, debugging and deploying agents visually, combining models, knowledge and tools directly, which suits customer-service assistants, knowledge Q&A, content assistants and similar applications.

Notable AI features:

• Visual agents and workflows. Set models, prompts and tools in the interface, or draw multi-step workflows on a canvas, adding branching and data handling as needed.
• RAG and knowledge bases. Connect company documents and knowledge data so agents answer from specified content.
• Plugins and external tools. Agents call outside services through plugins, bringing search, lookups and data processing into a run.
• One place for development, debugging and deployment, which cuts the switching between tools.

Typical uses: AI customer support answering product, service and after-sales questions from knowledge bases while plugins fetch extra business information; enterprise knowledge assistants offering employee Q&A, document lookup and information organising from internal material; and content and marketing helpers that combine search, generation and workflows for research, organising material and first drafts.

It suits teams just starting with agents who want customer-service, Q&A or content assistants fast without dedicated AI developers. Product, operations or business teams can test ideas visually, then add knowledge bases, plugins and more complex workflows later.

Langflow

Langflow (about 153.9K stars) is an open-source visual platform for building AI applications, chiefly agents and AI workflows. Modular components join models, tools, data and other agents while Python extensibility remains, which fits scenarios that call for constant experimenting and iteration.

Notable AI features:

• Drag-and-drop building and multi-agent composition. Configure agents on a canvas and connect models, prompts, tools and components into complete flows. Components and flows can act as agent tools, and several agents can split work such as retrieval, analysis and organising.
• MCP support. Connect to external MCP services, or expose a Langflow workflow to other MCP-aware applications.
• Models and Python extension. Models and data sources can be mixed freely, and custom Python components cover gaps by connecting enterprise APIs and processing logic.

Typical uses: knowledge and document applications that combine internal documents, retrieval, models and agents for Q&A, document analysis and extraction; multi-agent workflows in which different agents handle retrieval, analysis and organising within one process; and prototypes of support assistants, research assistants and content generators that later connect to existing products through APIs.

It suits teams with some technical involvement that experiment often with models and agent structures. Product or data staff can build visually while developers add depth through Python, custom components and APIs. It is strong for prototyping and continuous experiments. For teams with no technical staff that simply want fixed business workflows live fast, later maintenance may lean heavily on developers.

Dify

Dify (about 154.0K stars) is an open-source platform for developing AI applications such as agents, knowledge assistants and more involved AI workflows. It spans the lifecycle from creating and debugging an application to publishing it and managing it at runtime, which suits teams that plan to keep producing AI applications on one platform.

Notable AI features:

• Agent and workflow building. Configure an agent's model, instructions, tools and knowledge bases visually, or sketch multi-step workflows on the canvas, mixing branching, retrieval, tool calls and points where a person supplies input.
• RAG and knowledge processing. Turn PDFs, slide decks and similar files into knowledge bases, covering cleaning, chunking, indexing and tests of retrieval quality.
• Models and plugin extension. Connect several model providers and add tools, data sources, triggers and agent abilities through plugins.
• Publishing and runtime management. Release finished applications as web apps or APIs while watching logs, feedback, latency and usage.

Typical uses: customer support and knowledge assistants that mix enterprise knowledge bases with outside tools to handle customer questions or staff lookups; business information processing that chains extraction, classification, analysis, generation and tool calls; and quick launches where finished agents or workflows are published as applications or APIs and integrated with existing products and systems.

Dify fits teams that have moved past experimenting with one agent and intend to keep building support tools, knowledge assistants and content-processing utilities. Business and product people can do most configuration visually, with developers adding plugins, APIs and deeper integrations when required. For small and midsize businesses or product teams managing several AI projects, one platform can be easier than a patchwork of separate tools.

Group three: knowledge bases and RAG

MaxKB

MaxKB (about 22.6K stars) is an open-source, self-hosted enterprise agent platform that works well as a starting point built around knowledge bases and knowledge Q&A. Besides basic RAG it offers visual workflows and tool calling, which can be stretched into support, office assistants and similar scenarios.

Notable AI features:

• Knowledge bases and RAG. Upload documents or crawl online content, then chunk, vectorise and retrieve for internal materials and specialist Q&A.
• Visual workflows and tool calling. Combine conversation, retrieval, conditions, forms, document extraction and other nodes on a canvas, and connect MCP services so models can call outside tools.
• Many models, local deployment. Public models such as OpenAI, Claude, Gemini, DeepSeek and Qwen are supported, as are local ones through Ollama, vLLM and similar.
• Integration with existing systems. Finished agents embed into websites and internal apps through public links, embed code or APIs.

Typical uses: AI customer support that links product knowledge bases, FAQs and business material to an agent answering customer questions, and professional knowledge applications that use industry material for search, organising and analysis in finance, education, research and other specialised fields.

It is particularly suitable for teams with a large body of enterprise documentation who want knowledge-focused agents relatively quickly. The visual workflows and knowledge features keep the entry barrier fairly low, and local deployment with local models suits organisations that care about internal knowledge use, data control and third-party integration.

AnythingLLM

AnythingLLM (about 65.4K stars) is an open-source, local-first AI application mainly for building assistants over private documents and internal knowledge. A desktop edition and a self-hosted edition both come with a fairly complete graphical interface, so a team can start without assembling a complex development environment first.

Notable AI features:

• Private documents and RAG. PDF, Word, text and other files go into separate workspaces, and the AI answers from the chosen content while keeping citations.
• Agent flows. Chain scraping, API requests, model prompts, file reads and writes and further steps into flows that can be rerun on demand, all through a visual editor.
• Multiple users and permissions. The self-hosted edition manages several users and permissions, so workspaces and access to knowledge can be separated by user or group.

Typical uses: internal knowledge and document assistants answering questions, summarising and extracting data from company policies, product material, training documents, contracts and reports; and research and document organisation that combines internal documents, web information and tool calls to collect, arrange and analyse information.

It is better for individuals, small teams or companies piloting a private assistant within one department. Without dedicated AI developers, a team can get going quickly on document Q&A and information organising. When the main aim is letting AI use internal knowledge securely and not building elaborate business workflows, it is the lighter option.

RAGFlow

RAGFlow (about 89.7K stars) is an open-source RAG and agent platform aimed at complex document processing and enterprise knowledge retrieval. Compared with simpler Q&A tools, it suits settings with difficult document types, large volumes of knowledge and stricter needs for retrieval accuracy and source traceability.

Notable AI features:

• Understanding complex documents. It handles PDF, Word, PPT, spreadsheets, images and long documents, pulling out information usable for retrieval and analysis.
• High-accuracy RAG. Vector retrieval, keyword search, reranking and metadata filtering work together to raise accuracy and traceability.
• Agent workflows. A visual canvas combines retrieval, models, conditions, loops, classification and other steps, and more complex patterns such as multi-agent flows, planning and reflection are supported.
• Tools and data sources. Agents call outside tools and support MCP, and enterprise knowledge can sync from Confluence, S3, Notion, Google Drive and other sources.

Typical uses: complex enterprise knowledge bases that unify retrieval over large sets of policies, technical manuals, reports and specialist material; legal and compliance analysis that retrieves supporting evidence from regulations, contracts, cases and internal documents and then writes analysis with sources; and research and decision support that combine internal data with external material so agents retrieve, analyse and draft reports.

It fits teams with large knowledge volumes, intricate documents and some technical resources. In legal work, manufacturing, technical support, research and similar fields it shows its strength when material includes many long documents, tables, images or specialist content, and when retrieval accuracy and source evidence matter a lot.

Matching a platform to a team

• No developers, a handful of SaaS tools, tasks that repeat daily: Activepieces, or the no-code application platform when the work centres on structured records and approvals.
• Developers or data engineers, long multi-step jobs, retries and dependencies: Kestra, or Windmill when many scripts and internal services already exist to be reused.
• A product or data team that experiments with models and agent shapes every week: Langflow for fast iteration, Dify for a steadier path from prototype to published app, Coze Studio for a quick first assistant.
• A pile of documents and a need for answers with sources: MaxKB or AnythingLLM for a fast start, RAGFlow when documents are long, messy and accuracy matters.

Whichever row applies, the sensible first test is a single narrow workflow run end to end, with attention to who will maintain it in six months.

What to take from the list

Agent products increasingly overlap. Visual workflows, knowledge bases, MCP and tool calling now appear on more and more platforms, so feature checklists separate the options less than they used to. The questions that matter in practice are whether a platform fits smoothly into work that already exists and how much extra operating and maintenance burden it brings. A business with structured data and permissions to manage, a team stitching together SaaS tools, a group prototyping agents, and an organisation drowning in documents will each land on a different platform from the same ten, and each can reasonably start small before committing.