The native macOS harness for AI agents. Any model, persistent memory, autonomous execution, cryptographic identity. Built in Swift. Fully offline. Open source.
Replies
Best
this is one of the more compelling "local AI agent" pitches I've seen, the approval gate before every action is the part that actually matters to me, most of these tools just yolo the agent loose on your filesystem. question on the self-scheduling piece you mentioned - does that still fire if the Mac is asleep or the lid is closed, or does it need to actively be awake/plugged in for scheduled runs to trigger?
@omri_ben_shoham1Â Thank you for your feedback! It does require your Mac to be on and awake. It will queue it up until your system comes active.
Report
Congrats on the launch. The approval gate feels like the right trust boundary for an agent that can touch files, Calendar, Contacts, and iMessage. Iâm curious how you think about permissions once someone has several agents or projects: are you aiming for per-action approvals only, or eventually a project-scoped policy where an agent can access one context but not another?
@wesc Right now we're keeping it simple, per-action approvals seems to cover the highest surface area. In the future, we can get more granular, as agents earn more trust.
Report
That makes sense. Per-action approval is a practical default when the trust surface is still broad, and the idea of gradually earning more granular permissions feels like the right path. It keeps the user in control without forcing a complex policy model on day one.
Report
Mac-native and local is the combo I keep wanting, the Firefox-of-local-AI pitch landed. Practical question: what's the real memory footprint running one of the mid-size models on, say, a 16GB machine, does it stay usable or does it eat everything and make me quit my other apps? That's usually where "runs locally" turns into "runs locally if you bought the maxed-out box." Nice to see something leaning open source instead of another closed wrapper.
@chielephant I would say 16GB is going to be a little tight, but if you use something like Gemma 4 E4B, you should be able to use it alongside other apps.
Report
Huge congrats on the launch, @tpae ! Love the offline-first approach. One thing I’d find useful: a visual activity log to scrub through what the agent did locally. Any plans for that?
@jun_rex_bungabong Codex and Claude Code are designed for development, we're designed for every day use. We also focus on getting local models to work well, and provide local inferencing.
Report
persistent memory + fully offline is where the loop gets tight — the summarizer that decides what to keep is the same small model that produced the noise. compression drift compounds on itself with no bigger model to arbitrate. that's the part i'd want to see the eviction policy for, not the model swap.
@qifengzheng For which small model? It comes down to the model itself
Report
@tpae fair — though the bit i meant is structural, not model-specific. whatever runs the agent also writes the summary, so the compressor inherits the exact blind spots of the text it's compressing. a stronger small model moves where that breaks — it doesn't decorrelate the two.
Report
The approval gate + local memory combo is the interesting bit for me. I’ve been nervous about agents touching real files; seeing every action before it runs feels like the right tradeoff. Curious how you handle rollback when an agent writes bad files?
@xiaosong001Â You can ask the agent to undo the file operation, and it keeps an internal log of changes that it can revert to
Report
If I have my own key for OpenAI and Anthropic type of cloud models, is my Calendar/Contacts/iMessage data also transmitted to this external server through the agent or not even in such case all information is kept exclusively on my Mac?
The thing that bit us running tool-calling on small local models wasn't the reasoning, it was the JSON. Gemma-class models would emit a tool call with an unescaped quote or a trailing comma often enough that roughly 1 in 15 steps just died at parse time. With approval-before-each-action, does Osaurus validate and repair the tool-call schema before it shows me the approval, or do I approve and then watch it fail? On a 16GB machine that gap is the difference between usable and a toy.
Replies
this is one of the more compelling "local AI agent" pitches I've seen, the approval gate before every action is the part that actually matters to me, most of these tools just yolo the agent loose on your filesystem. question on the self-scheduling piece you mentioned - does that still fire if the Mac is asleep or the lid is closed, or does it need to actively be awake/plugged in for scheduled runs to trigger?
Osaurus
@omri_ben_shoham1Â Thank you for your feedback! It does require your Mac to be on and awake. It will queue it up until your system comes active.
Congrats on the launch. The approval gate feels like the right trust boundary for an agent that can touch files, Calendar, Contacts, and iMessage. Iâm curious how you think about permissions once someone has several agents or projects: are you aiming for per-action approvals only, or eventually a project-scoped policy where an agent can access one context but not another?
Osaurus
@wesc Right now we're keeping it simple, per-action approvals seems to cover the highest surface area. In the future, we can get more granular, as agents earn more trust.
That makes sense. Per-action approval is a practical default when the trust surface is still broad, and the idea of gradually earning more granular permissions feels like the right path. It keeps the user in control without forcing a complex policy model on day one.
Mac-native and local is the combo I keep wanting, the Firefox-of-local-AI pitch landed. Practical question: what's the real memory footprint running one of the mid-size models on, say, a 16GB machine, does it stay usable or does it eat everything and make me quit my other apps? That's usually where "runs locally" turns into "runs locally if you bought the maxed-out box." Nice to see something leaning open source instead of another closed wrapper.
Osaurus
@chielephant I would say 16GB is going to be a little tight, but if you use something like Gemma 4 E4B, you should be able to use it alongside other apps.
Huge congrats on the launch, @tpae ! Love the offline-first approach. One thing I’d find useful: a visual activity log to scrub through what the agent did locally. Any plans for that?
Osaurus
@tmaleh_Â Yup! We're looking into this
How is this different with Codex or Claude Code?
Osaurus
@jun_rex_bungabong Codex and Claude Code are designed for development, we're designed for every day use. We also focus on getting local models to work well, and provide local inferencing.
persistent memory + fully offline is where the loop gets tight — the summarizer that decides what to keep is the same small model that produced the noise. compression drift compounds on itself with no bigger model to arbitrate. that's the part i'd want to see the eviction policy for, not the model swap.
Osaurus
@qifengzheng For which small model? It comes down to the model itself
@tpae fair — though the bit i meant is structural, not model-specific. whatever runs the agent also writes the summary, so the compressor inherits the exact blind spots of the text it's compressing. a stronger small model moves where that breaks — it doesn't decorrelate the two.
The approval gate + local memory combo is the interesting bit for me. I’ve been nervous about agents touching real files; seeing every action before it runs feels like the right tradeoff. Curious how you handle rollback when an agent writes bad files?
Osaurus
@xiaosong001Â You can ask the agent to undo the file operation, and it keeps an internal log of changes that it can revert to
If I have my own key for OpenAI and Anthropic type of cloud models, is my Calendar/Contacts/iMessage data also transmitted to this external server through the agent or not even in such case all information is kept exclusively on my Mac?
Osaurus
@saksham_salvi We have a Privacy Filter function that helps keep sensitive data out of the cloud https://docs.osaurus.ai/privacy-filter
The thing that bit us running tool-calling on small local models wasn't the reasoning, it was the JSON. Gemma-class models would emit a tool call with an unescaped quote or a trailing comma often enough that roughly 1 in 15 steps just died at parse time. With approval-before-each-action, does Osaurus validate and repair the tool-call schema before it shows me the approval, or do I approve and then watch it fail? On a 16GB machine that gap is the difference between usable and a toy.