protoAgent is an open-source agent runtime designed to run on your own machine and use the models you choose. Its developers say conversations, memory, tasks and knowledge stay in a local SQLite database, while model requests go only to the endpoint you configure.
Watch Desk analysis
What happened
The project is available under an MIT licence on GitHub and PyPI, with desktop apps for macOS, Windows and Linux. Its creators describe a long-running agent that can plan, use tools, remember information and delegate work to sub-agents. It can connect to local llama.cpp or vLLM servers, or to other OpenAI-compatible endpoints.
The developers say the agent can pursue a goal until a defined check passes, such as making every test pass. It can also create charts, small apps and PDFs, and install tools, skills, scheduled jobs and other extensions as plugins. The project’s launch post is by Mabry of artificial-citizen, published on Hugging Face.
Why it matters
This is a concrete option for people who want an agent to keep working across tasks without handing its memory to a hosted service. The local-storage design and ability to point it at a local model give users control over where data and inference live, provided they configure the endpoint accordingly.
There is a practical trade-off: running an agent on your own machine does not make setup or reliability disappear. The developers say streaming, context management and smaller models getting stuck in tool loops were among the hard problems they encountered. Useful candour, and a reminder that “local-first” is an architecture choice, not a magic wand.
Our read
protoAgent is worth a look for technically minded users who want a persistent agent that can work with local models and hand off specialised jobs. Its mix of local storage, configurable model endpoints and plugin support is more interesting than another chat window with an agent badge. The project’s own account is a description of its design and development experience, not independent evidence of how well it performs across different setups.
What to watch
- How reliably it handles long tasks and tool use with different local models.
- Whether its plugin and sub-agent ecosystem grows beyond the developers’ own examples.
- How clearly users can inspect and control what the agent remembers and does.
Discussion spark: Would you trust a persistent agent with more autonomy if its memory stayed on your machine, or is inspectable control more important than where it runs?
Sources and evidence
- protoAgent: an open-source agent you run on your own machine, with your own models (26 October 2026, 18:04 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.