Hire a few members and split them into groups: who writes, who delivers, who researches. Not one do-everything chat box, but a room of members with different jobs.
A distinct identity each
Every agent has its own role, skills and memory. Swap the member and you swap the whole approach.
Grouped by function
Product, growth, research — one tab each. A single click shows who is on what.
You decide who gets it
Hand the same goal to different members, or let them pass it along and each deliver a piece.
Their own assets
Knowledge, skills and memory live on the member
Work goes to them, and what they learn stays with them — the longer you run them, the more they feel like a team you trained.
Finished work becomes skillThe next time a similar task shows up, it reuses the method instead of relearning it.
Conclusions become knowledgeSearchable, editable, and it does not vanish when a conversation ends.
Preferences become memoryYou never say it twice. Move the member to another machine and all of it goes along.
A member's own assets · skills, knowledge and memory it built itself
Self-evolution
The same job, the first time and the fifth
The difference is not the model — it is what the agent keeps for itself.
First time
You spell out who it is and what it owns
You walk it through the method, over several rounds
You watch every result
Fifth time
One sentence and it starts
It runs the method it kept last time
It stops only where your decision is required
What sits between them: the changes you approved — methods settle into skills, conclusions into knowledge, preferences into memory.
Beyond the workspace
The team plugs into more than the app
You do not have to do everything inside one window.
Members hand work to each other
One agent passes a subtask to another and the result lands back in the original conversation.
Reachable in WeChat and QQ
Talk to your agent inside WeChat or QQ — no app to install on your phone.
Everything leaves a trace
Who did what, when, and what came out of it: all of it stays reviewable, changes included.
How it works
Four steps to a team of your own
From install to your first agent improving itself — in minutes.
01
Download & launch
Install the package, then open http://127.0.0.1:8091/ in your browser.
02
Build your team
Create an agent per role: who it is, what it owns, and whatever it needs to do the job well.
03
Set it to real work
The agent puts its capabilities to work — and keeps the process and the experience.
04
Approve its evolution
When it finds a lasting gap, it brings you a plan. You approve, it makes the change, and nothing takes effect until the checks pass.
Security & privacy
Local by default
Local deployment means your agents, sessions and data stay on your own machine.
Conversations and files stay local
Agents, sessions, attachments and model connections live on your own machine; reinstalls and upgrades do not lose them.
One default outbound call: update checks
No ads, no third-party tracking. It checks for updates every 6 hours by default (can be turned off in Settings) and never sends conversation content.
It reports before changing itself
When it finds a lasting gap it brings a plan first; it only acts on the paths you approve, and every change is on record and reversible.
Quick start
Two commands, two minutes
The runtime ships in the package: no Node install, no Docker, no account.
On Linux, extract the tarball, enter the folder and run ./start.sh; on Windows, just run the installer. Then add a model provider in Settings to start dispatching work.
terminal
# Linux (x64)
tar -xzf proto-agent-0.1.93-linux-x64.tar.gz
cd proto-agent-0.1.93-linux-x64
./start.sh
# → http://127.0.0.1:8091/
# Windows
Proto-Agent-Setup-0.1.93.exe
FAQ
Good to know
How is this different from other agent tools?
Those are tools you use and leave. Here every agent keeps its methods and experience — and it never changes itself without your approval.
Is Proto-Agent free?
Yes. Free for individuals and teams; enterprises can purchase customization and support. The product is source-closed.
Do I need to write code?
No. Describe who the agent is and what it owns and it can start; if you want to go deeper, its assets are managed like code.
Which model, and what does it cost?
OpenAI, Anthropic or any compatible endpoint, with your own key. The software is free; you only pay your model provider.
Does it need internet?
Not by itself. Network is needed only when you call a model provider, and that request is configured by you.
Is my data uploaded?
Conversations, files and attachments stay on your machine. The only default outbound call is the update check (version, OS class and a random local ID), which you can turn off. No ads, no third-party tracking.
Which systems are supported?
Packaged today: Windows · Linux · macOS. Once downloads open, packages are on the download page.
Hire the first member of your team
Free, runs locally, and your data stays on your own computer.
