Leon

Quick Verdict

Leon is an MIT-licensed, open-source, self-hosted personal AI assistant. Current development targets the 2.0 Developer Preview on develop, with tools, environmental context, layered memory, agentic execution, smart, controlled, and agent modes, native skills, SKILL.md agent skills, Node.js and Python bridges, and Python services. master contains the more stable pre-agentic legacy architecture; it is not simply the same product with fewer bugs.

Leon suits developers willing to inspect source, operate the stack, and constrain tool permissions. It is not a turnkey Siri replacement. “Runs locally” also does not mean all inference and data remain local: Leon can use local models or remote providers, and every skill and toolkit has its own path. Compare Open WebUI for mature multi-user local chat, Ollama for a simpler local runtime, and Dify for governed business-agent applications.

Best For

• Open-source assistant enthusiasts and Node.js/Python developers.
• Contributors exploring personal-agent architecture and tool systems.
• Users able to deploy with Node.js 24+, pnpm, providers, and isolated test environments.
• Teams prepared to define model, tool, storage, and network boundaries themselves.
• Not ideal for buyers requiring stable APIs, complete documentation, enterprise identity, SLA, or zero maintenance.

Key Features

• Tool execution: Skills, actions, tools, functions, and optional binaries turn responses into actions.
• Three modes: Controlled favors deterministic native actions, agent supports planning, and smart selects a path.
• Context and memory: Environmental context and layered memory support continuity, subject to user governance.
• Local or remote providers: Model location is independent from Leon’s self-hosted application location.
• Extension layers: Native skills, agent skills, toolkits, bridges, and binaries separate capability concerns.
• Application stack: The 2.0 repository includes server, web app, Aurora preview environment, bridges, and a Python TCP service.
• Voice and media foundations: Audio and media capabilities exist, but availability should be verified against current develop code.

Use Cases

• Personal-assistant experiments in a disposable environment.
• Developing native skills for controlled actions or SKILL.md workflows.
• Comparing local-model quality, latency, and resource use behind one assistant.
• Testing hybrid providers only after explicitly approving data transfer.
• Studying tool schemas, approval gaps, context injection, memory retention, and recovery behavior.

Pricing

The Leon repository is MIT-licensed and has no software subscription. Self-hosting still costs compute, power, storage, backups, domains, and maintenance. Cloud models, search, speech, or other APIs charge under their own plans. MIT applies to Leon material its maintainers can license; it does not cover model weights, external services, skill dependencies, or user data.

Pros

• Inspectable and modifiable MIT-licensed architecture.
• Local context and local providers can reduce some external data transfer.
• Controlled and agent modes distinguish deterministic actions from open planning.
• Clear extension layers across skills, toolkits, and bridges.
• The project explicitly labels 2.0 and its documentation gap.

Cons

• 2.0 lives on develop and may change configuration, APIs, or behavior.
• New documentation is unfinished; public guides may describe legacy architecture.
• Node 24, Python services, provider configuration, and self-hosting require maintenance.
• A local app does not automatically resolve cloud-provider, tool, or model-license concerns.
• It lacks the SLA, support, enterprise controls, and secure defaults of mature commercial software.

Alternatives

FAQ

Is Leon 2.0 stable?

No. The project calls it a Developer Preview and develops it on develop. Expect breaking changes and documentation lag.

Should I use develop or master?

Use develop to study 2.0 tools, memory, and agent execution. Use master only when legacy pre-agentic behavior is specifically required. Do not apply master documentation blindly to develop.

Can Leon run fully offline?

It depends on the model, skill, and external services. A local provider with network tools disabled can create a more local path, while downloads, speech components, or some skills may still connect externally.

Are tools and previews safe by default?

Do not assume so. Tools may read files, run processes, or access networks, and previews can contain untrusted code. Use a dedicated account or container, restrict directories and egress, approve impactful actions, and inspect logs and diffs.

Does MIT mean every component permits commercial use?

MIT is permissive for Leon’s licensed code. Model weights, cloud APIs, npm and Python packages, media, and third-party skills retain separate licenses and terms.

Is Leon a production enterprise assistant?

That is not its current position. It is better treated as a developer preview and self-hosted experiment unless an organization adds identity, authorization, audit, backup, isolation, retention, and support controls.

Bottom Line

Leon 2.0 presents a compelling open-source direction for a personal agent by combining tools, context, memory, and several execution modes. Its honest current label is Developer Preview, not a free, private, turnkey replacement for commercial assistants.

Evaluate the branch first, then map three things: where the model runs, what each tool can access, and where data is stored or sent. Leon is valuable for learning and prototypes when develop-branch change and self-maintenance are acceptable. Mature alternatives are safer when stable team service is the priority.