The native macOS harness for AI agents. Any model, persistent memory, autonomous execution, cryptographic identity. Built in Swift. Fully offline. Open source.
Replies
Best
@tpae Huge congrats on bringing Osaurus to PH! 🎉
For users running hybrid workflows (e.g., small local model like Gemma for simple file/routing tasks, but switching to BYOK Claude/OpenAI for heavy reasoning), can an agent dynamically route tasks between local and cloud models within the same workflow, or is an agent fixed to one specific model backend?
@franz_briones Yes! We recently launched Subagents, which allows your agent to use local or remote model to delegate tasks to another local or remote model. It doesn't do intelligent routing yet, but if your reasoning model is capable of identifying which task is good for which model, should be able to set up routing logic via system prompts.
I see you also made apple script models. What was the thought process there? I know many similar apps use applescript just as an escape hatch for terminal calls.
@conduit_design We built Computer Use first, and it was great for what it is (clicking around the screen). We wanted to push automation a bit further, and came up with a solution to run AppleScript using local models. Since AppleScript allows chaining apps, we thought it will produce a very rich experience in automating your Mac. It's a little bit too powerful though, so be careful. It emptied my trash bin one time and I got scared.
@tpae I guess gurdrails and escalation etc for some actions could be a powerful layer to add in the future? like different categories of actions. which wold justfiy applescript approche vs just using applescript for cli, like pplx computer does.
@conduit_design Yup exactly. We're also thinking about baking in the safety on the models themselves so it will short circuit it if it writes something dangerous
@tpae I think safety must be deterministic. The pain is already felt with the current LLMs. if your fixing an encyption issue on github you often get abrupt action not allowed in the LLMs with no explinations. Probably some security fuse went of for no reason. Other times its works fine. But its like okay, enough with the hand holding please. 😅 I think AFM also has some issue with that for some things you cant do.
Report
How do you plan to handle model updates and ensure the agents stay current with the latest AI advancements without requiring a full app update?
@aymnart We do have auto update enabled, and we do daily updates. Our builds are fully standalone, it doesn't require any servers and you can run it locally offline. You can choose not to upgrade, but your app will fetch our latest models catalog from our hugging face.
We quantize our own models and also tune our evaluations every time a new model is dropped. We've been one of the pioneers in releasing quants for smaller models. We're one of the few that is supporting range of models including the latest Hy3.
Report
You describe Osaurus as an AI agent platform rather than just an assistant. Long term, do you envision users interacting with a single persistent agent that accumulates years of context, or a collection of specialized agents with separate memory and identities?
@tarqiya_forgah There's use cases for both, having a single agent (or orchestrator) that accumulate years of context, and having multiple agents (or subagents) that are specialized with contexts that can perform specialized tasks.
What we're discovering now is, local models trained with specific set of tasks, they are capable of performing at the same level (if not better) than frontier models for those tasks.
We believe this is just the beginning of a local-first AI harness. We call it a platform because we expect others to join us in shaping what this future would look like.
Report
Nice. If Qwen (or some of its models) are still open-source, which is my guess, I can confirm it works fine on a Mac Mini M3 for doing automation tasks on websites, DOM manipulation...
Swift-native, fully offline, and open source is a rare combo — most agent harnesses assume you're fine piping everything to the cloud. The cryptographic identity piece is intriguing; I'd love to hear more about how persistent memory works across sessions and whether it's local-only storage. This feels built for people who actually care about owning their stack. Nice work.
@kelly_king3Â Thank you Kelly! Memory is stored locally on device, using local embeddings. It works fully offline as well. We believe everyone should own their AI.
Report
The RAM requirements are surprisingly low - very impressive!
@natalia_iankovych Yup! Took a lot of work but that's been our main focus.
Report
no python underneath, built your own swift mlx runtime instead — same constraint we're under for voice checkins, nothing leaves the phone. does the approval gate get old once you're chaining multi-step actions, or can you pre-approve a trusted flow?
@sabber_ahamed You can absolutely pre-approve a trusted flow, there's an easy "Always Allow" button per gate.
Report
@tpae that's clean. is "always allow" scoped per gate type or per session, so a bad actor prompt-injecting mid-chain couldn't just ask nicely and get through on an already-approved flow?
@sabber_ahamed It's per gate type, I would suggest setting "always allow" for ones that you know is safe to do. For destructive actions, always have the gate available.
Replies
@tpae Huge congrats on bringing Osaurus to PH! 🎉
For users running hybrid workflows (e.g., small local model like Gemma for simple file/routing tasks, but switching to BYOK Claude/OpenAI for heavy reasoning), can an agent dynamically route tasks between local and cloud models within the same workflow, or is an agent fixed to one specific model backend?
Osaurus
@franz_briones Yes! We recently launched Subagents, which allows your agent to use local or remote model to delegate tasks to another local or remote model. It doesn't do intelligent routing yet, but if your reasoning model is capable of identifying which task is good for which model, should be able to set up routing logic via system prompts.
DiffSense
I see you also made apple script models. What was the thought process there? I know many similar apps use applescript just as an escape hatch for terminal calls.
Osaurus
@conduit_design We built Computer Use first, and it was great for what it is (clicking around the screen). We wanted to push automation a bit further, and came up with a solution to run AppleScript using local models. Since AppleScript allows chaining apps, we thought it will produce a very rich experience in automating your Mac. It's a little bit too powerful though, so be careful. It emptied my trash bin one time and I got scared.
You can read about our process here: https://osaurus.ai/blog/applescript-models
DiffSense
@tpae I guess gurdrails and escalation etc for some actions could be a powerful layer to add in the future? like different categories of actions. which wold justfiy applescript approche vs just using applescript for cli, like pplx computer does.
Osaurus
@conduit_design Yup exactly. We're also thinking about baking in the safety on the models themselves so it will short circuit it if it writes something dangerous
DiffSense
@tpae I think safety must be deterministic. The pain is already felt with the current LLMs. if your fixing an encyption issue on github you often get abrupt action not allowed in the LLMs with no explinations. Probably some security fuse went of for no reason. Other times its works fine. But its like okay, enough with the hand holding please. 😅 I think AFM also has some issue with that for some things you cant do.
How do you plan to handle model updates and ensure the agents stay current with the latest AI advancements without requiring a full app update?
Osaurus
@aymnart We do have auto update enabled, and we do daily updates. Our builds are fully standalone, it doesn't require any servers and you can run it locally offline. You can choose not to upgrade, but your app will fetch our latest models catalog from our hugging face.
We quantize our own models and also tune our evaluations every time a new model is dropped. We've been one of the pioneers in releasing quants for smaller models. We're one of the few that is supporting range of models including the latest Hy3.
You describe Osaurus as an AI agent platform rather than just an assistant. Long term, do you envision users interacting with a single persistent agent that accumulates years of context, or a collection of specialized agents with separate memory and identities?
Osaurus
@tarqiya_forgah There's use cases for both, having a single agent (or orchestrator) that accumulate years of context, and having multiple agents (or subagents) that are specialized with contexts that can perform specialized tasks.
What we're discovering now is, local models trained with specific set of tasks, they are capable of performing at the same level (if not better) than frontier models for those tasks.
We believe this is just the beginning of a local-first AI harness. We call it a platform because we expect others to join us in shaping what this future would look like.
Nice. If Qwen (or some of its models) are still open-source, which is my guess, I can confirm it works fine on a Mac Mini M3 for doing automation tasks on websites, DOM manipulation...
Osaurus
@elias_iturri Qwen is a beast of a model. One of the best released for it's size
Just discovered @Osaurus on X
Used my ChatGPT Pro sub
Recognised my local MLX models from LM Studio
💜 Dino avatars
Osaurus
@karthikeyan_ranasthala Welcome to the community! Thank you for your support!
Does it take a lot of space in my computer?
Osaurus
@doganakbulut Depends on which model you download, but it ranges between 6GB ~ 20GB for smaller models, and could be more than 200GB+ for larger ones
Creatium
Swift-native, fully offline, and open source is a rare combo — most agent harnesses assume you're fine piping everything to the cloud. The cryptographic identity piece is intriguing; I'd love to hear more about how persistent memory works across sessions and whether it's local-only storage. This feels built for people who actually care about owning their stack. Nice work.
Osaurus
@kelly_king3Â Thank you Kelly! Memory is stored locally on device, using local embeddings. It works fully offline as well. We believe everyone should own their AI.
The RAM requirements are surprisingly low - very impressive!
Osaurus
@natalia_iankovych Yup! Took a lot of work but that's been our main focus.
no python underneath, built your own swift mlx runtime instead — same constraint we're under for voice checkins, nothing leaves the phone. does the approval gate get old once you're chaining multi-step actions, or can you pre-approve a trusted flow?
Osaurus
@sabber_ahamed You can absolutely pre-approve a trusted flow, there's an easy "Always Allow" button per gate.
@tpae that's clean. is "always allow" scoped per gate type or per session, so a bad actor prompt-injecting mid-chain couldn't just ask nicely and get through on an already-approved flow?
Osaurus
@sabber_ahamed It's per gate type, I would suggest setting "always allow" for ones that you know is safe to do. For destructive actions, always have the gate available.