Jarvis · working name
Jarvis is the name we're using for something bigger than a chatbot.
A personal intelligence should remember context, understand priorities, work with real tools, and keep control in your hands. Jarvis is being built from that idea. Yes, the name is a nod to the obvious inspiration, and it may change before release.
Jarvis's core. It breathes, and it has something to say.
Task
“Compare our competitors' pricing and save a report in the workspace.”
Plan
- 1.Read the three pricing pages
- 2.Extract plans, limits, and prices
- 3.Draft the comparison table
- 4.Save the report to the workspace
Activity
parsing the request▍
How a task will move through the system: plan → tools → approval → result. This is the design being built, not a shipped product.
The feature map
What's in motion, what's next, what's a question.
Everything splits into three tiers so you never have to guess which claims are real. Nothing on this page is called “available”; the first release isn't out yet.
Being built now
The core of the first release. In code, not in slides.
- Multi-step planning from a single request
- Tool use: browser, files, terminal, workspace
- Controlled autonomous execution with approval gates
- Readable execution logs for every run
- Astro as consultable, local-first memory
- Model gateway: your API keys, or local models via Ollama
Planned
Designed, not yet started, or started but not yet usable.
- The Solen Model, shipping open-weights alongside the first release
- Recurring automations and scheduled jobs
- Deep terminal and repository integration
- Specialized sub-agents for research, writing, and code
- Reusable workflows you can share
- Project-level context and goals
Open research questions
The hard list. We don't know how to do these well yet. Nobody fully does.
- Long-term memory that stays reliable over months
- Learning your preferences without being creepy about it
- Agents evaluating their own results
- Multiple agents collaborating without chaos
Human in the loop
Jarvis can do a lot. It shouldn't do everything without asking you.
You choose how much it asks
Manual, Plan, Auto, or Always-approve: from approving every step to letting it run without questions. The leash length is a setting, and it's yours.
Your keys, for now
In the first versions, models run on API keys you provide and can revoke. Later, Solen's own services and keys will be an option too. Either way, you decide.
A sandbox when you want one
Code can run in isolated containers, or directly on your machine if that's what you choose. Isolation is an option you control, not a default you fight.
Traceability as a goal
Every run should leave a readable log of what was tried and what happened. It's genuinely hard, especially in the cloud, and we're building toward it rather than claiming it's done.
More detail on how this works (sandboxing, permissions, logs) is on the security page.
Open by design
Open source. Not just the first version. Every version.
You shouldn't have to depend on us to use your own assistant.
Free to download and run
Jarvis will be fully open source, and that's a commitment for every version, not a v1 marketing move: download it, run it on your own machine, modify it. It keeps working even if solenai.app disappeared tomorrow: no account, no lock-in.
Not MIT. Protective on purpose
The license won't be MIT. We'll use one that keeps the code free to use and study while protecting the project and its authorship: a copyleft-style license, named precisely at release.
The model ships open too
The Solen Model built for Jarvis will be open-weights alongside it, at least for the first version: download the weights, run them where you want. For other, future models we'll decide when we get there, and we'll say so here.
The first version is getting close.
We publish progress as it happens: no launch theater, just a changelog. If you want to know the moment it's usable, follow along.
Follow the buildAnd where does Jarvis point? At an old idea we take seriously: an operating system with AI in its bones. See it on the horizon.