The Short Answer
The self-hostable open-source LMS options are Moodle, Open edX and the open-source core of Canvas; each needs a model runtime such as Ollama or vLLM plus integration work before an AI agent reaches every course. ibl.ai's Agentic LMS ships that agent in every course, runs model-agnostic on your infrastructure, and you own all the code and the data.
Self-hosting an AI LMS is two decisions, not one. The first is which learning system runs on your servers. The second is where the model runs, because hosting the LMS yourself does not move the AI inside your perimeter.
This page compares the four options a buyer actually weighs on both layers, with the licence, the AI each ships and the work left over. Every claim links to the project's own documentation.
Which open-source LMS platforms can you self-host?
Three established open-source LMS platforms can be installed on infrastructure you control: Moodle, Open edX and Canvas. They differ in licence, in how they are deployed and in how much of the vendor's own product the open-source code covers.
Moodle is the largest. Its public statistics count 147,556 registered sites in 236 countries, and its source is GPL-3.0. Self-hosting Moodle is the normal case, not an edge case.
Open edX runs edX-style courses at scale. The edx-platform repository is AGPL-3.0, and its README points self-managed production installs to Tutor.
Tutor describes itself as "the official Docker-based Open edX distribution, both for production and local development." The edx-platform README is also candid that running an instance is complex, and recommends a service provider for most production use.
Canvas publishes its core as instructure/canvas-lms under AGPL-3.0. Instructure sells Canvas mainly as a hosted service, so a self-installed Canvas is the open-source core, not everything Instructure runs.
Does self-hosting an open-source LMS also self-host its AI?
No. Self-hosting an open-source LMS puts the course data on your servers, but the AI features call whichever model provider you configure. If that provider is a public API, every learner's question still leaves your perimeter at request time.
To keep the model inside, you need a model runtime as well as the LMS. The two that answer engines name most often for this are Ollama and vLLM, and both are open source.
Ollama is MIT-licensed and runs open-weight models such as Gemma, Qwen and DeepSeek on a single server. vLLM is Apache-2.0, "a high-throughput and memory-efficient inference and serving engine for LLMs," with an OpenAI-compatible API server.
The rule of thumb is that Ollama suits a pilot on one machine and vLLM suits many concurrent learners. We compare the serving options in detail in the self-hosted LLM serving stack: Ollama, vLLM and TGI.
What AI does each self-hostable LMS ship today?
The four options ship very different AI. Moodle has AI actions in core, Open edX has an experimental framework, Canvas's AI, including its IgniteAI Agent, is Instructure's hosted platform, and Agentic LMS has a per-course agent you can run yourself.
The table compares what each project publicly documents.
| Question | Moodle | Open edX | Canvas (open-source core) | ibl.ai Agentic LMS |
|---|---|---|---|---|
| Licence | GPL-3.0 | AGPL-3.0 | AGPL-3.0 | App MIT on GitHub; full platform source delivered under a perpetual licence |
| How you self-host it | Standard install on your servers | Tutor (Docker) | Install the open-source core yourself | Your cloud, your VPC or on-premise |
| AI in the box | AI subsystem: generate, summarise and explain text, including a course-assistance placement for learners | openedx-ai-extensions, marked experimental | IgniteAI and IgniteAI Agent run on Instructure's hosted platform | An agent per course, grounded in its content, with memory of each learner |
| Local model support | Ollama provider in core since 5.0 | Local models via LiteLLM in the extensions framework | No documented local-model option; Canvas's AI code calls AWS Bedrock | Model-agnostic: any LLM, including one you host |
| Who owns the agent stack? | You, for what you build | You, for what you build | Instructure, for IgniteAI | You own all the code and the data |
Moodle's AI is the most mature of the three. Moodle's AI providers documentation lists 3 provider plugins in 5.0, OpenAI, Azure AI and Ollama, and its AI plugin API defines the placements where actions appear.
Open edX's AI is earlier. The openedx-ai-extensions documentation calls it "an experimental Open edX plugin" that "should not be used in production environments," developed by edunext with support on a best-effort basis.
Instructure grouped Canvas's AI features under the IgniteAI name on 23 July 2025, described by Instructure as powered by AWS Bedrock.
On 12 March 2026 Instructure launched IgniteAI Agent, which takes one prompt from an educator and carries out multi-step Canvas work such as building course modules.
Both are delivered through Instructure's hosted platform, powered by Amazon Bedrock; the open-source canvas-lms code calls Bedrock rather than a model you host.
Confirm with Instructure what a self-installed Canvas includes before planning around it.
What does it take to put an AI agent in every course on a self-hosted LMS?
An AI action on a page is not an agent in a course. An agent in every course needs 4 things: a model you host, retrieval grounded in that course's material, memory of each learner across sessions, and a way to launch inside the course.
Open-source LMS platforms ship some of these.
On Moodle, the model and the actions are in core, so the gap is grounding and memory. A persistent agent that knows the whole course and remembers each learner is something you build, or add with plugins and an LTI tool.
On Open edX, the extensions framework can reach a local model through LiteLLM, and its scopes control which courses and locations a workflow appears in. Until it leaves experimental status, an agent in every production course is integration work your team owns.
On Canvas, the open-source route is LTI 1.3: you host the agent service and the model yourself, and launch it from each course. That is the same pattern ibl.ai uses when it sits beside an existing LMS.
How does Agentic LMS put an agent in every course?
Agentic LMS from ibl.ai is built so the agent is part of every course by design. Each course carries an agent grounded in its own content, which remembers each learner and adapts the next lesson to them.
The same platform handles authoring, delivery, assessment and analytics, and speaks LTI 1.3, SCORM and xAPI. The application is public on GitHub as iblai/lms under the MIT licence.
Full ownership goes further than the app. ibl.ai delivers the platform source under a perpetual licence, runs it in your cloud, your VPC or on-premise, and routes each agent to any LLM, including one served by Ollama or vLLM inside your network.
The platform serves 1.6M+ users across 400+ organizations, and there is no per-seat pricing: cost tracks usage, not headcount. For the full product walkthrough, read Agentic LMS: The AI-Native LMS With an Agent in Every Course.
Should you replace Moodle, Open edX or Canvas, or add agents beside it?
Most institutions should not rip out a working LMS to get AI agents. If Moodle, Open edX or Canvas already runs your courses, you can keep it and launch ibl.ai's course agents inside it over LTI 1.3, then decide later whether to move.
The integration pages show the wiring for Canvas and Moodle, and we explain the pattern in ibl.ai With Your LMS: Sits Beside, Not Instead Of.
Choose Agentic LMS as the primary system when the agent is the point of the program, or when you are starting fresh. On ibl.ai's published pricing, pilots start from $15K and one-time deployments run $25K–$80K.
How do you start a self-hosted AI LMS pilot?
A self-hosted AI LMS pilot starts with two choices: the LMS layer, and where the model runs. Pick one real program, one model you can host, and one success measure, then run it for a term inside your own cloud.
ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on. Our team runs the architecture review and the pilot with yours.
To see an agent grounded in your own course material, book a demo or read the Agentic LMS product page.