Diviso is a free AI agent that runs on your own machine, even an Android phone
A solo developer released Diviso, a Python agent with no account and no middleman server. It matters if you want an AI helper without handing your data to a provider.
The short version
- Diviso needs no account and no server run by its author; you use a local model or your own API key.
- The core uses only Python's standard library, so it runs on a GPU PC, an old laptop or an Android phone via Termux.
- Every run has hard caps on steps and cost, and its tools sit behind an approval gate and a workspace sandbox.
- It is v0.1.0 with no token streaming yet; its 74 offline tests use a mock provider, not real cloud services.
- Local Ollama is the path the author tests by hand; cloud providers are built from their documented APIs.
Why it matters
Most AI agents tie you to someone else's account, pricing and servers. A tool that runs on hardware you already own shows what the alternative looks like, and it keeps working if a provider changes its terms. It is still an early solo project, so treat it as something to try, not something to depend on.
What the developer built
As dev.to reported, a developer has released Diviso, a Python-based AI agent designed to run entirely on your own machine. An agent, in this sense, is a program that takes a goal, breaks it into steps and uses tools such as a code runner or a file editor to carry the steps out.
The author says the idea came from frustration. Most agents they tried wanted an account, a subscription and their data sitting on someone else's server. Diviso is the opposite: there is no Diviso account and no server of the author's in the middle.
You choose the model behind it. You can point it at a local model through Ollama, a popular tool for running open models on your own computer. Or you can plug in a cloud provider with your own key: OpenAI, Claude, Grok, Gemini, Mistral, DeepSeek or OpenRouter.
What is inside the agent
The agent works in a loop that reasons, acts and then looks at the result. This is a pattern known as ReAct. A planner sits alongside it, and every run has hard limits on how many steps it can take and how much it can spend.
Memory comes in two forms: a SQLite database (a small file-based database) and a plain text file called MEMORY.md. It can also search your local files using TF-IDF, an older keyword-matching method that scores how distinctive a word is in a document. There is also a small team of agents with the roles Researcher, Coder and Reviewer.
Its tools cover Python, JavaScript and Bash, file edits, tests, read-only git access and fetching web pages. All of them sit behind an approval gate, meaning you confirm actions, and inside a workspace sandbox that limits where the agent can work.
The list continues with an MCP client (a standard way for agents to connect to outside tools), a guarded scheduler, and checkpoint and resume after every step. You use it from a command line or a small web dashboard.
Running on a phone and old laptops
The part the author is proudest of is portability. The core relies only on Python's standard library, the set of tools that ship with the language, so there is little extra software to install. The same code runs on a GPU-equipped PC, an old laptop and an Android phone using Termux, an app that provides a command-line environment.
Diviso also checks your hardware and picks one of three profiles: lite, balanced or pro. Each comes with model sizes that the author says actually fit the machine. This matters because the usual obstacle to running models locally is memory, and a model that is too big for your device simply will not work.
Note what runs where. The agent itself is light enough for a phone, but the model it talks to still needs to be small enough for that device, or reached through a cloud key. The profiles exist to handle that choice.
Honest limits
This is a first release, v0.1.0, and the author is open about the gaps. There is no token streaming yet, which means you do not see the answer appear word by word.
The cloud providers are written against their public documentation. The 74 offline tests, which pass in the project's automated checks, run on a deterministic mock provider, a stand-in that gives predictable answers. That means those tests do not prove the real services work. The author says local Ollama is the main route tested by hand.
We have not tested Diviso ourselves, and this is a solo launch with no track record yet. Whether it holds up in daily use is an open question. The author asks for feedback, especially from people running it on unusual hardware.
Why local-first is worth a look
Local-first means your data and your tooling stay with you. If a provider raises prices, changes its terms or goes offline, a tool like this keeps working with whatever model you have on hand.
It also fits a pattern worth knowing about: a model has no memory of its own, so long-term memory has to be stored somewhere and fed back in. Diviso keeps that memory in files on your machine, where you can read and edit it. That is a sound design choice, although how well it works in practice is untested.
- Good fit: privacy-minded people who like to tinker.
- Poor fit: anyone who needs a polished, supported product today.
What to watch
- Will the cloud providers, tested only against a mock, work as well as the local Ollama path?
- Will token streaming and wider testing on odd hardware arrive in later releases?
Try it
If you try Diviso, start with non-sensitive files, and check which model it is using and what license that model carries before relying on it.
Sources and links
- Reportingdev.to: the article we sourced
First reported by dev.to. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.