I spent hundreds of dollars across different AI tools and subscriptions. At the end of it, I owned none of it. Every one of them needed some platform to keep paying just to keep working. What actually bothered me wasn't the cost — it was realizing that without the cloud provider behind it, none of it would run. I never really owned my own AI.
So I started figuring out how to build something different.
What I ended up building:
AI Partner is a desktop app that runs an actual multi-agent AI system entirely on your own machine:
Four agents (Research, Learning, Memory, Action) that hand off work to each other — you can watch it happen live
Chat with real code fixes: attach a file, describe the bug, apply the fix with one click, automatic backups
An optional Business Coach that remembers your goals across every conversation, notices repetitive manual work and offers to build you a personal tool for it, and can generate a logo/business card/letterhead directly in the chat
Image, video, and music generation
A community marketplace for sharing agents — every submission goes through automated screening plus human review before it's listed
The technical/product decisions I'm most opinionated about:
One-time payment, not a subscription. This is the entire point of the product, not just a pricing choice — it's a direct answer to the problem that got me here.
Ollama by default, cloud keys optional. The app has to work without any account at all, or the "own it" pitch falls apart.
GitHub PRs as the moderation system for community agents, instead of building a custom review queue. Free code review infrastructure that already exists, and it's transparent — anyone can see the review history.
Where it's at now:
Just wrapped up 2.0 — a full UI redesign, the Business Coach agent, and the ability for it to build personal tools or generate brand assets on request. Getting ready to actually put it in front of people (this post included).
Happy to answer anything — architecture, why local-first over cloud, the packaging pipeline, or the honest tradeoffs of building on a local model vs. a hosted API.
https://minister-ballew.github.io/ai-partner-website
Top comments (3)
I really appreciate your frustration with renting AI tools and the subsequent decision to build your own solution, AI Partner. What resonates with me is the importance of owning the technology and having full control over the data and functionality. I've had similar experiences with cloud-based services, where the cost and dependence on the provider became a concern. I'm intrigued by your decision to use a local-first approach and the use of GitHub PRs as a moderation system for community agents - can you elaborate on how you see this model scaling and handling potential security concerns as the community grows?
Thanks — that really is the core of it for me. Cost was the trigger, but ownership and control over the actual behavior of the thing is what stuck with me.
On the GitHub PR model, a few honest specifics:
How it works today: Every community agent submission goes through automated screening first — scanning for suspicious patterns (prompt injection attempts, unusual network calls, anything that looks like it's trying to exfiltrate data) — before a human ever looks at it. Then it needs a human review and merge before it's listed. Trust tiers (New, Community-Tested, Trusted) are earned from real submission history over time, not something a contributor can buy or self-declare.
The part I'd call actually solved: Even a fully "Trusted" agent doesn't get a free pass at runtime. Every file change an imported agent wants to make gets staged for the user's explicit approval — nothing writes automatically, regardless of trust tier. That's deliberate defense-in-depth: I don't want the whole security model resting on review quality alone, because review will never be perfect. If something malicious slips past screening and human review, the execution-time gate is still there.
Where I'll be honest it's untested: The human-review step is a real bottleneck at scale. Right now that's fine — submission volume is low enough for manual review to work well. If this ever gets to hundreds of submissions a week, that step needs to change — probably tiered review effort based on contributor trust history (a first-time anonymous submitter gets full scrutiny, an established Trusted contributor's straightforward update gets lighter-touch review), rather than the same manual bar for everyone. I haven't needed to build that yet, so I don't want to claim it's solved — just that the architecture (screening → review → merge → staged execution) leaves room to add it without a rearchitecture.
Happy to go deeper on any piece of that — the screening heuristics specifically, or how trust tiers get computed.
Just wrapped up 2.0 — a full UI redesign, the Business Coach agent, and the ability for it to build personal tools or generate brand assets on request. Getting ready to actually put it in front of people (this post included).