TL;DR
Open source AI agents come in four kinds, and the best one depends on your job. Frameworks help engineers build agents, and ready-made agents run out of the box. Coding agents write software, and low-code builders suit business teams. Check the license before you adopt one, because some popular projects limit commercial use. Check that the project is still maintained, because some have been archived. Running an agent also costs money for patching, hosting, tracing and human approval.
Key Takeaways Open source AI agents come in four kinds, which are frameworks, ready-made agents, coding agents and low-code builders, so choose the kind before the project. Read each project’s LICENSE file, because n8n, Dify and the AutoGPT platform carry commercial limits that sit outside the open source definition. Check project health on the repository itself, since Flowise was archived in August 2026 and AutoGen now receives no new features. Open source AI agents cost nothing to license and still carry engineering, inference, hosting, patching and tracing costs. Choose tools that speak MCP, which now sits under the Linux Foundation, so a framework swap does not mean rebuilding every integration. Enterprise use needs identity, governed data access, human approval and tracing around the agent, whichever project you pick. Watch on YouTube
4 Questions Every CTO Should Ask Before Choosing OpenClaw
A short Kanerika guide to the questions a technology leader should answer before adopting OpenClaw, one of the most popular open source agents.
What Flowise Taught Its Users in August 2026 Flowise had more than 55,000 GitHub stars and an Apache 2.0 license. According to their own announcement , the maintainers froze feature work on July 29, 2026 and archived the repository on August 13. They set August 31 as end of life, and teams that had built agents on it had to fork the code or move.
The license did its job, because every fork was legal. It did not keep anyone on call to patch a vulnerability or add the next model provider. Anyone running open source AI agents needs to judge the maintainers and the terms as carefully as the features.
What Counts as an Open Source AI Agent in 2026 An open source AI agent is software that plans, calls tools and finishes tasks for a user. Its license must meet the Open Source Definition , so you can read, run and change the code.
A chatbot answers a question and waits. An agent decides what to do next, calls tools and loops until the job is done. That loop needs memory, permissions and a way to stop.
Our guides to agents versus chatbots and the types of AI agents cover the basics.
A workflow script follows a fixed path that a developer wrote in advance. An agent lets a model choose the tools, routes and next steps. That helps when the path is unknown and costs more when it is not.
Search results mix four different products under that one label, and the first decision is which kind you need.
Agent frameworks are libraries you code against, such as LangGraph or CrewAI, so you build the agent. Ready-made agents are products you install and run, such as goose or Hermes Agent. Coding agents work inside a repository or terminal, such as OpenHands or Cline. Low-code builders give business teams a visual canvas, such as Dify or Langflow. Every star count, version and license in this article was read from the project’s own repository on October 5, 2026. Those numbers move quickly, so check them again before you decide.
The License Check Most Open Source AI Agent Lists Skip The Open Source Definition requires free redistribution. It also bars licenses that restrict anyone from using the program in a specific field of endeavor, including a business.
Several popular agent projects sit outside that line, and their star counts hide it. Here is what each LICENSE file says today.
Permissive Licenses You Can Build On LangGraph, CrewAI, the OpenAI Agents SDK, Microsoft Agent Framework, goose, OpenHands, OpenClaw and Hermes Agent ship under MIT or Apache 2.0. Those licenses allow commercial use, modification and redistribution without asking anyone.
Open Core With Paid Folders Mastra, Langfuse and Flowise license most of their code as open source and reserve separate folders for paid terms. Mastra’s license file puts every ee directory under a different license, and those folders include authentication and the agent builder.
The core is usable on its own. Check whether the feature you need sits in the paid folders before you design around it.
Modified and Source-Available Terms Dify uses a modified Apache 2.0 license. It requires a commercial license to run a multi-tenant service. It also bars removing the logo from its front end. n8n uses its Sustainable Use License , which limits use to internal business purposes or non-commercial use.
The AutoGPT repository splits in two. Its platform folder carries the Polyform Shield license, which bars products that compete with it, while the classic folder stays under MIT.
None of this makes these projects bad choices. A platform team using them internally may be fine, while a software vendor or agency reselling them needs legal review first.
Open Source AI Agent Frameworks Worth Evaluating Frameworks are the largest group of open source AI agent tools, and the shortlist changed a lot since 2025. LangChain is still the largest by stars. Microsoft moved AutoGen into maintenance mode, and Microsoft Agent Framework now carries the enterprise path for its users.
Table 1 compares thirteen frameworks on license, latest release and GitHub stars as of October 5, 2026.
Table 1. Open source agent frameworks compared on October 5, 2026
Framework License Latest release GitHub stars Starts from LangChain MIT langchain-core 1.6.6, Sep 29 147.5k Broad model and tool integrations LangGraph MIT 1.2.13, Oct 5 42.7k Stateful, long-running agents CrewAI MIT 1.15.23, Sep 28 59.4k Role-based agent crews Microsoft Agent Framework MIT python-1.20.0, Oct 2 14.0k .NET, Python and Go teams OpenAI Agents SDK MIT v0.23.1, Oct 2 29.8k Lightweight multi-agent workflows Google ADK Apache 2.0 v2.11.0, Oct 2 21.7k Building, evaluating and deploying agents Pydantic AI MIT v2.54.0, Oct 3 20.4k Typed Python agents smolagents Apache 2.0 v1.26.0, May 29 29.7k Agents that think in code Strands Agents Apache 2.0 python/v1.57.2, Oct 1 8.7k In-process Python and TypeScript agents Agno Apache 2.0 v3.1.1, Oct 2 42.6k Agent runtime and platform Mastra Apache 2.0 plus ee folders 1.74.0, Oct 5 28.6k TypeScript products Haystack Apache 2.0 v3.3.0, Oct 1 26.7k Retrieval and pipelines LlamaIndex MIT v0.14.25, Sep 21 52.4k Document-heavy agents
LangGraph and CrewAI LangGraph describes itself as a low-level orchestration framework for long-running, stateful agents, with durable execution and human-in-the-loop controls built in. CrewAI is built around role-playing agents that work as a crew, which suits teams that think in job titles.
Our LangChain vs LangGraph comparison covers the first decision in depth, and our guide to multi-agent systems explains when a crew beats a single agent.
Microsoft Agent Framework, OpenAI Agents SDK and Google ADK Microsoft Agent Framework is MIT licensed and runs across Python, .NET and Go. The AutoGen README calls it the enterprise-ready successor and links to a migration guide .
The OpenAI Agents SDK is a lightweight framework for multi-agent workflows. Google’s Agent Development Kit is a code-first toolkit for building, evaluating and deploying agents. See our comparison of CrewAI, AutoGen and Microsoft Agent Framework for the Microsoft path.
Retrieval-Led and Typed Frameworks LlamaIndex and Haystack start from documents and retrieval, which fits knowledge agents. Pydantic AI adds typed outputs, smolagents lets agents think in code, and Strands runs inside your own process with no hosted control plane.
Mastra serves TypeScript teams, and Agno gives teams a runtime for agent platforms.
What the Table Shows and What It Hides Release dates show activity. Twelve of the thirteen projects tagged a release in the last six weeks. The exception is smolagents, whose last tag is from May 29, although its repository still received pushes on September 30.
Stars show attention and little else. CrewAI has more stars than LangGraph, yet the two solve different problems, so a higher count does not mean a better fit.
For production criteria such as state, tracing and governance, read our guide to AI agent frameworks and our explainer on agentic RAG .
Ready-Made Open Source AI Agents You Can Run Today Ready-made open source AI agents ship as products you install, connect to a model and give tasks. They attract the most stars and carry the most risk.
OpenClaw OpenClaw is an MIT-licensed assistant that runs on your own computer. It answers in the chat apps you already use, including Slack, Teams, Telegram and WhatsApp. It has more than 390,000 GitHub stars, and its latest release is v2026.9.8.
It also has a heavy security record, which the security section below covers in detail.
Hermes Agent Hermes Agent comes from Nous Research under the MIT license and has more than 250,000 stars. Its README describes a built-in learning loop that turns experience into reusable skills, and it works with any model you point it at.
Our Hermes Agent guide explains the memory design and the enterprise gaps.
goose and Browser Use goose is an Apache 2.0 agent for desktop, command line and API, written in Rust, that connects to more than 70 extensions through MCP. It belongs to the Agentic AI Foundation at the Linux Foundation, which gives it a neutral home.
browser-use lets an agent drive a real browser, and it has about 117,000 stars under the MIT license. Pick the agent whose permissions you can actually control. An assistant with access to your inbox and files carries a different risk from a library in a container.
Kanerika Service
Agentic AI Services
Kanerika designs, builds and governs production agents on open source and managed stacks, with evaluation, access controls and audit logging in place.
Explore Agentic AI Services Open Source Coding Agents Coding agents are open source AI agents that work inside a repository or terminal, and search results treat them as a separate category. OpenHands is MIT licensed and has more than 90,000 stars. It describes its Agent Canvas as a self-hosted control center for coding agents, including Claude Code, Codex and Gemini.
OpenCode has more than 210,000 stars under MIT, and Cline has about 70,000 under Apache 2.0. SWE-agent is MIT licensed, has about 20,000 stars and turns a GitHub issue into a proposed fix.
OpenAI’s Codex CLI and Google’s Gemini CLI are Apache 2.0 licensed too, with more than 127,000 and 107,000 stars.
In each case the agent code is open source, and the model behind it is a separate decision with its own terms and cost. For a buyer’s view of these tools, see our guide to AI coding agents .
Low-Code Builders for Open Source AI Agents Low-code builders put open source AI agents on a visual canvas for business teams. Dify has about 158,000 stars, n8n has about 207,000 and Langflow has about 156,000.
Langflow is MIT licensed. Dify and n8n deserve a careful read, because their licenses limit multi-tenant hosting and commercial resale.
Flowise belongs on this list only as a warning. The project was archived on August 13, 2026, and its Apache 2.0 code remains available to fork. Our roundup of AI agent builder platforms covers the managed alternatives.
Ready-Made, Coding and Low-Code Agents Compared Table 2 puts the three product-style groups of open source AI agents side by side, with the same live data as Table 1. Several carry more stars than LangChain, the largest framework at about 147,000, and their licenses differ more than the frameworks do.
Table 2. Ready-made, coding and low-code agents compared on October 5, 2026
Project Kind License GitHub stars Latest release OpenClaw Ready-made agent MIT 391.4k v2026.9.8, Oct 3 Hermes Agent Ready-made agent MIT 251.4k v2026.9.24, Sep 24 goose Ready-made agent Apache 2.0 55.0k v1.53.0, Oct 2 browser-use Browser agent MIT 117.2k 0.13.10, Sep 4 OpenCode Coding agent MIT 211.9k v1.18.34, Sep 30 Codex CLI Coding agent Apache 2.0 127.9k rust-v0.160.1, Oct 5 Gemini CLI Coding agent Apache 2.0 107.2k v0.62.0, Sep 29 OpenHands Coding agent MIT 90.0k v1.24.0, Sep 25 Cline Coding agent Apache 2.0 69.9k desktop-v0.0.43, Oct 2 SWE-agent Coding agent MIT 20.5k v1.1.0, May 22, 2025 n8n Low-code builder Sustainable Use License 206.7k [email protected] , Oct 5Dify Low-code builder Modified Apache 2.0 157.9k 1.17.1, Sep 10 Langflow Low-code builder MIT 155.5k v1.12.4, Sep 29
SWE-agent is the outlier on release date, with its last tag in May 2025, although the repository still received a push on September 28, 2026. Read the repository itself as well as the tag.
Where Open Source AI Agents Work Best Today Production use clusters around tasks where the work is repetitive, the data is reachable and a person can review the result.
Customer support triage suits a framework such as LangGraph, with a knowledge base behind it and a human handoff when confidence is low. Our guide to AI agents for customer support covers the design. Document and contract review suits retrieval-led tools such as LlamaIndex or Haystack, plus an approval step before anything is sent. Data extraction and web automation suit browser-use, although an API route is safer whenever an API exists. Software development suits OpenHands, OpenCode or Cline, with every change reviewed before it merges. Research assistants suit a ready-made agent such as goose, which can reach sources through MCP. Internal workflow automation suits Dify or Langflow when business teams own the logic. See our guide to agentic automation for the wider pattern. Notice what the list leaves out. Anything that moves money, deletes data or messages customers needs a human approval step, whichever agent runs it.
Open Source AI Agent Projects That Stopped Moving Older lists, including the earlier version of this article, still recommend open source AI agents that have slowed or ended. Table 3 shows what the repositories say today.
Table 3. Agent projects that slowed, split or stopped
Project Status on Oct 5, 2026 Evidence Look at instead Flowise Wound down Maintainers froze development Jul 29, archived the repository Aug 13, end of life Aug 31, 2026 Dify or Langflow AutoGen Maintenance mode README says no new features, last release python-v0.7.5 on Sep 30, 2025 Microsoft Agent Framework AgentGPT Archived Repository archived, last push Apr 29, 2025 A ready-made agent such as goose OpenAgents (xlang-ai) Dormant Last push Nov 18, 2024 LangGraph or CrewAI SuperAGI Dormant Last push Jan 22, 2025, last release v0.0.14 in Jan 2024 CrewAI MetaGPT Slowed Last release v0.8.1 on Apr 22, 2024, last push Jan 21, 2026 CrewAI or CAMEL Aider Slowed Last release v0.86.0 on Aug 9, 2025, last push May 22, 2026 OpenHands, OpenCode or Cline AutoGPT Active, split license Platform under Polyform Shield, classic under MIT, platform release v0.8.2 on Sep 30 Read the license section above
The earlier version of this article named AutoGPT, AgentGPT, OpenAgents and CAMEL. Two of those have ended or stalled and one now has a split license. CAMEL is the one still moving, with a push on the day we checked.
It also praised OpenDevin, which is now OpenHands and has more than 90,000 stars.
Stars hide all of this. AgentGPT still shows 36,295 stars although it is archived, and AutoGen shows 61,265 although it receives no new features. That is why stars are a poor health signal.
Five Checks Before You Commit Run these checks on the repository itself. A README or a ranking page can hide the facts that matter.
Look at the last tagged release, and treat a long silence as a warning. Check whether the repository is archived and whether the README or discussions announce an end of life. Read the LICENSE file, because the badge on the README can hide an ee folder or a competing-use clause. Open the security advisories and see whether issues are published, fixed and dated. Find out who governs the project, a foundation or a named set of maintainers. The first, second and fourth checks take a minute with the GitHub CLI. Replace OWNER and REPO with the project you are testing.
gh api repos/OWNER/REPO --jq '{archived, pushed_at, license: .license.spdx_id, stars: .stargazers_count}'
gh api repos/OWNER/REPO/releases/latest --jq '.tag_name + " " + .published_at'
gh api "repos/OWNER/REPO/security-advisories?state=published&per_page=100" --paginate --jq '.[].severity' | sort | uniq -cThe license field reads NOASSERTION when a project uses a custom license, so open the LICENSE file itself. A project that passes all five checks is not guaranteed to last. It is far less likely to leave you with a fork to maintain.
MCP and A2A Protocols That Keep an Agent Portable Three open standards decide whether your tools survive a framework change. The Model Context Protocol connects agents to tools and data, and Anthropic donated it to the Linux Foundation on December 9, 2025. Neutral governance lowers the risk that one vendor can change the rules.
The same foundation took in goose and AGENTS.md, a convention that gives coding agents project instructions, according to the Linux Foundation announcement .
Google created Agent2Agent, which is now a Linux Foundation project under Apache 2.0. The A2A site lists a steering committee that includes AWS, Cisco, Google, IBM Research, Microsoft, Salesforce, SAP and ServiceNow. Version 1.0.1 shipped on May 28, 2026.
The MCP specification repository lists a release dated July 28, 2026, so the standard is still moving. On October 5, 2026 the READMEs of CrewAI, Google ADK, Microsoft Agent Framework and Agno mention A2A. The READMEs of CrewAI, the OpenAI Agents SDK, Google ADK, Pydantic AI, Strands, Agno, Mastra and Haystack mention MCP.
A missing mention does not prove missing support, so read each project’s documentation for transports, authorization and version compatibility. Protocol support also does not replace security review, because every MCP server you add is code that your agent can run.
In practice, expose your internal tools as MCP servers and treat the agent framework as replaceable. That protects the work you put into your tools. Our MCP vs A2A guide explains when each protocol applies, and MCP and context-aware agents walks through the pattern.
On-Demand Webinar
Model Context Protocol (MCP): The Key to Building Context-Aware AI Agents
A recorded Kanerika session on how MCP connects agents to tools and data, and how to design agents around it.
Watch the Webinar → How to Choose an Open Source AI Agent by Scenario Start from the job and the team, then check the license and the health of each open source AI agent. Table 4 maps common scenarios to a first pick.
Table 4. A first pick for common scenarios
Scenario Start with Why Watch for Long-running workflow with approvals LangGraph Durable execution and human-in-the-loop controls are built in Needs engineers who can run it Prototype of role-based agents CrewAI The crew and role model maps to job titles Add an evaluation set before production Microsoft estate on Azure Microsoft Agent Framework Successor to AutoGen with .NET, Python and Go support Smaller community than LangGraph or CrewAI TypeScript product team Mastra TypeScript-native framework Authentication sits in the paid ee folders Document-heavy knowledge agent LlamaIndex or Haystack Retrieval is the starting point Measure retrieval quality separately Assistant in chat apps OpenClaw or Hermes Agent Runs on your own hardware and talks through chat apps Broad access needs strict permissions and fast patching Developer productivity OpenHands, OpenCode or Cline Permissive licenses and large communities Review every change before it merges Visual builder for business teams Langflow or Dify Canvas with agent features, MIT for Langflow Dify limits multi-tenant hosting Tools that must outlive any framework MCP servers plus goose Open protocol under the Linux Foundation Audit every MCP server you add
Treat a first pick as a starting point that your own tests must confirm. Run the same small task through two candidates on your own data before you commit.
Score each candidate on task success, cost per task, latency, the number of human corrections and the effort to wire in your tools. Two weeks of that evidence beats any ranking page, including this one.
Teams also need to know whether their data and processes are ready for any agent. Our agentic AI enterprise adoption guide and the readiness checklist below help with that.
Checklist
Agentic AI Readiness Checklist
A practical checklist for judging whether your data, controls and teams are ready to run AI agents in production.
Get the Checklist → What Open Source AI Agents Really Cost to Run A zero license fee moves the cost of open source AI agents into five other places, and the bill grows with usage.
Engineering time comes first, because someone must build, test and own the agent. Model inference comes next, whether through a paid API or through your own GPUs running Ollama or vLLM, which are MIT and Apache 2.0 licensed. Our guide to open source LLMs covers the model side.
Agents that must read screenshots, scanned documents or video need a model that accepts those inputs, which our guide to multimodal models compares.
Inference is also where retries and runaway loops hurt, because every failed run and oversized context window is billed like a good one. Human review adds a further cost, since someone has to handle approvals and exceptions.
Hosting is small for light agents, since the Hermes Agent README says it can run on a $5 VPS. That is the project’s own claim, and a team deployment needs more.
Patching is the hidden line. GitHub’s published advisory feeds on October 5, 2026 list 722 advisories for OpenClaw, 14 of them critical, and 210 for n8n, 22 of them critical. LangGraph lists 10, Dify lists 21 and CrewAI lists none.
Advisory counts reflect both exposure and disclosure habits, so read them as patch workload rather than a quality score. Tracing and evaluation add the last cost, because an agent without them cannot be debugged or approved. Our AI agent development cost guide breaks down the build side.
A simple way to compare options is cost per completed task. Add inference, hosting, tracing, review time and a share of engineering and patching, then divide by the tasks the agent finished correctly. That figure also gives you a fair comparison against a paid platform.
Securing Open Source AI Agents Security is where open source AI agents get risky fastest, because they read untrusted content and then act on it.
The OWASP Top 10 for LLM Applications lists prompt injection first, supply chain third and excessive agency sixth. A typical agent combines all three.
IBM X-Force’s April 2026 analysis of OpenClaw describes a lethal trifecta. The three parts are deep access to private data, contact with untrusted content and the ability to communicate outward. It counted more than 255 security advisories at that time.
Memory is part of the attack surface too. An agent that stores what it learns, as Hermes Agent does, can carry a poisoned instruction into later sessions. Our Hermes Agent guide covers that memory layer and its governance gaps.
Skills and plugins carry the same supply chain risk. IBM X-Force links OpenClaw issues to malicious skills, unsecured endpoints and leaked credentials, so every extension needs review before it reaches production.
Controls That Work in Practice Give each agent its own identity with the least privilege it needs. Run agents in a container or sandbox with no standing credentials. Treat every plugin, skill and MCP server as third-party code that is reviewed and pinned. Require human approval before irreversible actions such as payments, deletions or external messages. Log every tool call so an auditor can replay a run. Our guides to agentic AI risks , the AI security framework and agentic AI governance go deeper on each control. Review them every quarter, because new skills, plugins and models change the risk.
Watch on YouTube
AI Agent Access Control: From Excessive Agency to Least Privilege
How to limit what an agent can read and do, using least-privilege identities and scoped tool access.
Tracing and Evaluation Tools for Open Source AI Agents Langfuse is MIT licensed apart from its ee folders, has about 35,000 stars and joined ClickHouse in January 2026. Arize Phoenix has about 11,700 stars and builds on OpenTelemetry instrumentation, but its license is the Elastic License 2.0, which is source-available.
Whichever you pick, build a fixed evaluation set before any model or prompt change. Our guides to AI agent observability and AI agent evaluation show how.
A useful evaluation set is small and specific. Include typical tasks, the answer you expect, known failure cases and a few hostile inputs, then rerun it after every model, prompt or dependency change.
Store the set next to the code, under version control, so reviewers can see exactly how it changes over time. Run it in continuous integration so that a regression blocks the release, and log the cost of each run.
Evaluation Habits That Catch Regressions Scoring outputs with a second model helps at scale, but that judge can share the blind spots of the agent it grades. Sample real runs for human review every week, and keep every failure as a new test case.
Track tool-call accuracy as well as final answers, because an agent can reach a correct answer through a risky step. A trace that shows each tool call and its arguments makes that visible.
Auditability depends on the agent showing its work. Kanerika’s compliance agent returns structured reports with citations and compliance mapping, so reviewers check evidence instead of redoing the research.
Case Study
3x Faster Expert Vetting with an AI Compliance Agent
Kanerika built an AI agent that profiles experts, screens news and social sources against compliance rules and returns structured reports with citations.
Read the Case Study → Build on Open Source AI Agents or Buy a Platform Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027. It cited escalating costs, unclear business value and inadequate risk controls. It also estimated that only about 130 of the thousands of agentic AI vendors are real.
Open source AI agents suit teams that have engineers, a need for control or a reason to avoid lock-in. A managed platform suits teams that need support and faster governance sign-off.
Hybrid setups are common. A team might run open source orchestration on its own infrastructure and buy managed models, identity and observability. That keeps the agent logic portable and hands the heavy operations to a vendor.
Managed options include Microsoft Fabric data agents , Databricks Agent Bricks , Snowflake Cortex agents and OpenAI AgentKit . Each ties the agent to its platform, which can be the right trade. Our Databricks vs Snowflake comparison weighs the agent features of those two platforms side by side.
Seven questions settle most build-versus-buy debates.
Do you have engineers who can own the agent after launch? Do you need service levels or vendor support for audit reasons? How sensitive is the data the agent will touch? How many systems must it integrate with? Which model providers must it support? What volume do you expect in year one and year three? What is the three-year cost, including staff time? Our guides to building AI agents , agentic AI tools and custom AI agents compare the paths in more detail. Revisit the decision each year, since project health and managed offerings both change quickly.
A Short Roadmap From Shortlist to Production Whichever open source AI agents you choose, the same six steps keep the rollout controlled.
Define the task and the autonomy boundary, meaning what the agent may read, decide, change and escalate. Shortlist two projects using the license, health and protocol checks above. Build a pilot on a real workflow with real data and tools. Set an evaluation baseline for task success, error rate, cost per task, latency, tool-call accuracy and human intervention rate. Add identity, permissions, sandboxing, approval gates and logging before the agent touches production systems. Roll out with narrow permissions and widen them only when measured results support it. How Kanerika Runs Open Source AI Agents in Production Kanerika is an AI-first data and automation consulting firm, founded in 2015 and headquartered in Austin, Texas. Agent work sits inside a broader data and AI practice, so the framework is one decision among several.
Four Stages From Shortlist to Production Assess the decision the agent will make, the systems it must touch and the approval points a risk team will require. Design by checking each candidate against the license, health and protocol tests in this article, then build the data layer underneath it. Build and evaluate a narrow agent with tracing and an evaluation set from day one. Govern and scale by adding access controls, PII handling, audit logging and cost guardrails before widening scope. Four Controls No Framework Ships Whichever project wins, four controls must be added around it. Identity and permissions come from the enterprise directory, and data access follows sensitivity labels. Human approval gates irreversible steps, and logs stay readable to auditors.
On a recent deployment, Kanerika mapped Microsoft Purview sensitivity labels to an agent’s MCP tool permissions. The data layer decided what the agent could retrieve, and no prompt could override it.
Four pitfalls come up often in reviews. Teams pick a project by stars, hand an agent a broad service account or skip the evaluation set before a model swap. Some discover a license limit only after they have built on the project.
What We Check Before Adopting a Project Before a client builds on an open source project, we read the LICENSE file, scan the advisory feed and confirm it is not archived. We also pin versions, so an upstream release cannot change agent behavior without a test run. These checks give the client a written record of why a project was chosen, which helps when auditors ask.
We then plan for the fork. If the maintainers stopped tomorrow, the client should know who would patch the code. They should also know which tools sit behind MCP and how long a migration would take.
A Support Agent in Production For a network of expert members, Kanerika built an AI support agent that connects to knowledge bases and Zendesk. It resolves 65% of member queries and routes complex cases to live executives when its confidence is low.
The design choices matter more than the model here. Confidence-based routing keeps a person in the loop, and the knowledge base supplies the answers. Both choices carry over to open source stacks without change.
Case Study
65% Self-Service Resolution with an AI Support Agent
A Kanerika agent connected to knowledge bases and Zendesk resolves routine member queries and routes complex cases to live executives when its confidence is low.
Read the Case Study → Our agentic AI services and AI governance services apply the same stages to open source and managed stacks.
Wrapping Up Open source AI agents reward teams that check before they commit. The right pick depends on the job, the license and the health of the project behind it. Check the repository on the day you decide, because star counts and tags age quickly.
Then plan for the work around the agent, which is identity, governed data, approvals and tracing. That work decides whether the agent ever touches enterprise data.
Frequently Asked Questions
What are open source AI agents? Open source AI agents are software that plans, calls tools and completes tasks, with source code under a license that meets the Open Source Definition. They include frameworks such as LangGraph, ready-made agents such as goose, coding agents such as OpenHands and low-code builders such as Langflow. You can inspect, run and modify the code.
How do open source AI agents compare with proprietary and paid platforms? Open source gives you code access, model choice and no license fee, and you carry the work of hosting, patching, tracing and governance. Paid platforms bundle support, service levels and built-in controls, and they tie the agent to the vendor stack. Teams with strong engineering often pick open source, while teams that need faster sign-off often pick managed.
Are OpenClaw and Hermes Agent safe for company data? Treat both as high risk until hardened. OpenClaw has published 722 GitHub security advisories, 14 of them critical. IBM X-Force calls broad data access, untrusted content and outbound communication a dangerous mix. Hermes Agent lists no advisories on GitHub, but its persistent memory needs governance. Run either one in a sandbox with scoped credentials.
How much does it cost to run an open source AI agent? The license is free, and the bill comes from engineering time, model inference, hosting, patching and tracing. Retries and long contexts raise inference spend. The Hermes Agent README says it can run on a $5 VPS, but team deployments need more. Track the cost per completed task so you can compare it with a paid platform.
Should I choose MCP or A2A? They solve different problems, and many teams use both. MCP connects an agent to tools and data, while A2A lets independently built agents delegate work to one another. Both sit under the Linux Foundation. Start with MCP, because tools outlive frameworks, and add A2A when agents from different teams must cooperate.
How do companies use open source tools to build production AI agents? They pair an open source framework with governed data, a tracing tool and approval steps. A workable stack uses LangGraph or Microsoft Agent Framework for orchestration, MCP servers for tools, Langfuse for traces and a sandboxed runtime. The framework is the small part, and most effort goes into permissions, evaluation and monitoring.
Can open source AI agents run on local models? Yes. Most frameworks and ready-made agents accept any model endpoint, including local servers such as Ollama or vLLM. goose lists Ollama among its providers, and Hermes Agent lets you switch models without code changes. Local models cut API spend and keep prompts on your hardware, but quality and speed depend on your GPUs.
Are there open source AI agents for specific tasks? Yes. OpenHands, OpenCode, Cline and SWE-agent handle software engineering, browser-use drives web browsers, and LlamaIndex and Haystack focus on document retrieval. OpenClaw and Hermes Agent act as general assistants in chat apps. Specialized agents are easier to test and secure than general ones, so match the agent to one narrow task.
Are LangGraph, CrewAI and the OpenAI Agents SDK open source? Yes. LangGraph, CrewAI and the OpenAI Agents SDK all use the MIT license, which allows commercial use, modification and redistribution. Google ADK and Strands Agents use Apache 2.0. Mastra is Apache 2.0 with separate ee folders under another license. Always confirm the LICENSE file in the exact version you install.
What are the most popular open source AI agent frameworks? By GitHub stars on October 5, 2026, LangChain has about 147,000, CrewAI about 59,000, LlamaIndex about 52,000, and LangGraph and Agno about 43,000 each. The OpenAI Agents SDK, smolagents and Mastra follow near 29,000. Stars show attention rather than fit, so check the license, release history and your own use case too.
What are common use cases for open source AI agents? Common uses include customer support triage, document and contract review, research, data extraction, browser automation and software development. Support agents that read a knowledge base can resolve routine member queries without a person. Start with one narrow, well-bounded task, because agents struggle when the goal is vague or the tools are broad.
Do I need to be a developer to use open source AI agents? Frameworks need Python or TypeScript skills. Ready-made agents such as goose and Hermes Agent install as apps or command line tools and need configuration rather than coding. Low-code builders such as Langflow and Dify give business teams a visual canvas. Anyone running them still needs engineering help with security, updates and monitoring.
Are open source AI agents secure and reliable? They can be, when you add the controls yourself. OWASP lists prompt injection, supply chain and excessive agency among its top LLM risks, and agents combine all three. Use least-privilege identities, sandboxes, reviewed plugins, human approval for irreversible actions and full logging. Then test against an evaluation set before launch.
Can I use open source agents in commercial projects? Yes for MIT and Apache 2.0 projects such as LangGraph, CrewAI and goose. Read the license first for the others. n8n limits use to internal business purposes, Dify restricts multi-tenant hosting and the AutoGPT platform bars competing products. Ask legal counsel before you resell or host any of them for customers.
Which open source AI agent is best for enterprise use? No single project wins for every team. LangGraph suits stateful workflows with approvals, Microsoft Agent Framework suits Azure and .NET estates, and OpenHands suits engineering teams. Whichever you shortlist, check the license, release history and security advisories. Then run a pilot on your own data with identity, governance and logging in place.
Which popular agent projects are not truly open source? n8n uses the Sustainable Use License, which limits use to internal business purposes. Dify uses a modified Apache 2.0 license with multi-tenant and logo conditions. The AutoGPT platform folder uses Polyform Shield, and Arize Phoenix uses the Elastic License 2.0. Mastra and Langfuse keep some paid code in ee folders.
Which open source agent projects are no longer maintained? AutoGen is in maintenance mode and points users to Microsoft Agent Framework. Flowise froze development on July 29, 2026 and was archived on August 13. AgentGPT was archived in April 2025, and OpenAgents and SuperAGI have had no pushes since November 2024 and January 2025. Check the repository before you adopt any project.
