jerieljan

Hah, I agree. As soon as I can figure out how to replace batteries myself, I'd also be happily doing the same thing with mine.

Yeah, it's very confusing especially if you're not in the loop and their docs were only updated recently and was confusing during the v1 -> v2 transition weeks / months ago.

So far, "OpenCode" has gone through several possible definitions:

- OpenCode Terminal / CLI - This is what you get normally. Runs a service in the background and the CLI is your interface.

- OpenCode Desktop - Instead of a CLI, you get a desktop interface.

Most users usually stop here. You just install the app and you connect wherever.

So there's the services where they both make money and make things easier for users:

- OpenCode Console - Basically the serving platform for both Zen and Go, and has expanded to a hosted inference gateway and somewhat similar to OpenRouter or a managed LiteLLM. Has the option to BYOK if you have existing AI platform accounts so it charges those instead of Zen, and that's what got compromised.

- OpenCode Zen - the API service they host if you want to get access to providers but don't want to pay individually per API platform. Similar to OpenRouter. I noticed they're kinda slowly dropping the Zen moniker.

- OpenCode Go - the subscription service that they host that features open weight models and select low-cost models from frontier labs. Has different policies when it comes to privacy and retention compared to Zen.

- OpenCode Black - a limited-time, limited access subscription service that grants access to frontier models. Not sure if this is still alive or not.

I've been on the API-only mentality for months and all the open models were definitely the stuff I loved the most. Kimi K2.5, 2.7 and Deepseek v4 were among my favorites, while sparingly using Opus or whatever OpenAI had for specific situations.

But ever since I've switched to one of the $100-tier subs, I can see why a lot of the people on it don't really discuss the open models often. I'd still use it especially when it comes to sensitive inputs, but for most work, what you get on OpenAI or Anthropic is really more than enough.

It really got even better when they also made their cheaper models up to par if not better than the open models.

I do think the crowd for open models are out there, especially when you see trillions of tokens running for them on OpenCode or OpenRouter leaderboards.

I haven't even considered using MCP for locally-running apps! I mean I used to, but they almost always got associated with cloud service access and integrations nowadays, and most local stuff is trivially handled with agent skills.

Thanks for the tips. TIL about BTT having MCP support. Crank looks very promising too, since it looks like a more effective Shortcuts and Hazel replacement

Funnily enough, my first impression for this project was to use it for non-code and planning first.

I agree, I don't think people are going to remotely pair-prog with this when it's existed in the past and most work on their own and let their pull requests do the talking.

But I say non-code and planning because I've definitely strategized with AI before and have scrapped a good chunk of it because I overlooked details that my team shared. Being able to get that on an active session with context shared across all members as it's debated sounds good and would've reduced turnaround time.

I remember the Opencode folks mention about gangprompting before and I guess this is one take of it, besides adding an Opencode bot on Slack and hoping that's competent enough for the work that needs doing. I've never tried it myself, so I'm not gonna criticize it, but I'm sure it's an idea people want to see if it's effective.

Yep, I did similar stuff too! Used something like ntfy to push notifs to an Android or just whatever commands are available on the OS you're in (e.g., `notify-send`). And for state reads herdr has `herdr pane read` while you can happily accomplish similar results with `tmux capture-pane` or similar.

Herdr holds potential in stuff like their socket API ( https://herdr.dev/docs/socket-api/ ) and whatever enhancements they wish to bring (e.g., https://herdr.dev/docs/agent-automation/ ) so if that appeals to others, it's worth a shot.

But for most regular users, tmux and other multiplexers work just fine.

You can check their own AI-generated comparison here: https://herdr.dev/compare/

Realistically, what I use about its agent stuff is:

- reliable notification pings either via what your terminal app uses (e.g., ghostty's) or system or just within herdr only

- if absolutely necessary, be able to tell agent sessions to observe and run commands on a herdr pane (security beware, for obvious reasons)

- herdr --remote which is basically a quick SSH -> attach to herdr on a remote

- prefix+g (equivalent to prefix+w in tmux) presents itself nicer, has binds for agent state filtering and its /search imho is mildly nicer than tmux's

Arguably, you can have the same points in others if you know how to configure tmux or zellij already, so if you're comfortable with that then you really don't have to use herdr.

While we're at it, this should cover products and food in general too. I hate it when I see a poster that clearly has AI-generated images of food and the actual food you get is nowhere near what they generated.

Or even if it's similar, it's still clearly staged to look more appealing than what you should be getting.

Everyone knows deceptive advertising is bad, but somehow those who use AI for advertising turned their brains off on this topic because the images looked cooler than what they'd normally produce with a camera.

The Razer one is spot-on. I had a Razer mouse during the Windows 7 days and it kept running the installation setup despite declining it all the time, and it's so bad at times I recall seeing the setup trigger while on an ongoing Windows update when the screen was already locked up.

Logitech did something similar if I recall right.

Nice. I did transcriptions on a casual project before that went through something like this. Transcribing videos or audio files with Whisper? Very common. But having to swap it out with Qwen3 or a different family of ASR models? Oops, not as straightforward. For Qwen for example you gotta deal with the forced aligner or it won't be good as subtitles, and then gotta deal with some requirements and considerations if you want to make use of MLX on a Mac or something.

Will definitely check this out since it sounds like it eases through the pain of dealing with these.

I also quite like that it has a decent CLI and skill for agents to use too.

Quite handy to apply since you can instruct models to run devservers in herdr panes or workspaces rather than doing long running shell commands or scripts in the agent session.

(You could probably do the same with tmux and others but I've had better success with monitoring herdr state)

Default stickiness, tbh.

A few years ago, I started moving towards Podman when it got to that "good enough" point on both Linux and macOS and when Docker started to remind people about Docker Desktop that it needed a license for commercial use.

Even then, it took around a year or so to transition.

It's terrifying, yeah.

To some degree, the closest we have to these situations besides getting flagged with TOS violations (whether real or false-flagged) in these companies are residents of countries that are either trade or economically sanctioned by the USA.

Thankfully we haven't seen something like an account ban and deletion incident for such cases, but the severe ones I can remember usually prohibit access entirely and that'd be scary if it extended to primary services that others rely on for auth.

You will be effectively locked out to services if it's all that's linked and that identity provider just decided you'd be persona non grata.

I remember in the early days when I was just trying out ChatGPT on a phone for the first time (this was around GPT-3.5? GPT-4o?) and snapping a picture of our fridge that's full of magnet souvenirs and asked it to identify all the places we've been in and it gave a nice list of what it saw and the places that were featured.

Did it get it fully right? No. But it was one of those "oh wow, you could do that?" moments for me. There's obviously a lot more "oh shit" moments as time went on, but it was a neat little moment.

why do I feel like they basically added their AI support chatbot to the same group / mailing list that the human support belonged to along with the same permissions set and just called it a day?

I'll laugh even harder if they wrote tests for it and only made tests for the happy path and not the error cases or just ignored the latter.

The thing that intrigues me on this one is the Opencode roots.

I've currently host both Open WebUI and Opencode on a server setup and there's multiple things that Open WebUI has implemented very nicely over time but it was also prior to the emergence of coding agents / harnesses. As such, its strengths is more on the `/chat/completions` era and RAG and its own functions and ecosystem setup before tool calls were a thing.

If this stuff works for you already, I think that's fine. That's not a bad thing, and if anything, they've made attempts to improve esp. with the recent backend upgrades and projects like `open-terminal`.

However, Opencode has been built for the coding agent setup since its early days and Odysseus gives me more confidence to run sessions much better knowing that it runs on an Opencode core.

Stuff like MCP support and a clearer lifecycle for example are things I've never really seen implemented well on open-webui (they've improved and support this now, but I'm biased and am salty of the era from its early days of MCP support when they insisted you place a bridge (mcpo) in between just to make it do MCP.). In contrast, opencode does this very well since what they built that works on the terminal also applies to the desktop app and the ACP support for IDEs that use it to connect to the opencode harness.

At the minimum, I expect Opencode to work wonderfully against the frontier models, and all models they host on their Zen and Go service, so that adds some confidence as well on complex workflows.

In contrast, I still find model management a pain in the ass on open-webui (has always relied on LiteLLM or a third-party inference gateway) just to wrangle the models. Hell, they still tag the Responses API "experimental" today, a format that's been around since over last year.

Does that transfer to Odysseus? Maybe, maybe not. Seems it's based on it but a quick check on code says they merely adapted it so maybe it's the latter. I haven't tried it at depth and idk how well Pewds has implemented it, but if it improves on that angle and gets maintainers that accepts contributions from the community, I feel like it'll do nicely.

Surprisingly, there's a lot of items I'm seeing at first glance here at Odysseus that open-webui either doesn't have or takes extra effort to add in and I've used the latter for a really long time already.

Stuff like an agent mode, deep research and document work are things you'd be able to have with open-webui as well but Odysseus seems to have them thought out.

This is three steps back, one step forward kind of an approach imho.

Easy for everyday users to deal with, and effective for verifying humans vs bots.

But holy hell, if your phone is a requirement to access sites and you have to go through the security theater like a work device and setting this behavior as a default assumption to have? Ugh. The privacy and security implications of this is quite ugly to think about too, now that Google can link your devices to a stronger degree with this approach.

I always thought this was the case since they declared war against Opencode and its users.

The lines drawn by their consumer vs commercial TOS was clear and I never subscribed because of it.

That's a correct example, and I agree, it is disingenuous to just trivially call this an `is-odd` project.

Back in the days of GPT-3.5, LiteLLM was one of the projects that helped provide a reliable adapter for projects to communicate across AI labs' APIs and when things drifted ever so slightly despite being an "OpenAI-compatible API", LiteLLM made it much easier for developers to use it rather than reinventing and debugging such nuances.

Nowadays, that gateway of theirs isn't also just a funnel for centralizing API calls but it also serves other purposes, like putting guardrails consistently across all connections, tracking key spend on tokens, dispensing keys without having to do so on the main platforms, etc.

There's also more to just LiteLLM being an inference gateway too, it's also a package used by other projects. If you had a project that needed to support multiple endpoints as fallback, there's a chance LiteLLM's empowering that.

Hence, supply chain attack. The GitHub issue literally has mentions all over other projects because they're urged to pin to safe versions since they rely on it.

I've been using it and so far yeah, a lot of the existing functionality and new functionality is effectively free and it generally only nags you when you use the clearly separate "Canva AI" panel.

Checking the settings also tells me that Segmentation (used in Object Selection) is provided but Depth Estimation (used in Portrait Blur and Select Sampled Depth tools), Colorization (used in Colorize filter, apparently intended to colorize B&W photos) and Super Resolution (used in Super Resolve filter) are all paywalled.

Honestly, I think that's fine. While I'd wish these are all available (and they could be if you looked hard enough for models that can do this) for a flat price (for parts that are not handled server-side), this is still imho mostly fair.

I have a similar setup in Todoist, it's just a reminder for scheduled recurring tasks like bills.

Funnily enough, I was quite savvy with the features several years ago but as my work changed and things aren't as easy to list down like a routine or in neatly defined projects and such.

And when regular tasks becomes freeform, it's no surprise that a plaintext file is sufficient.

I've tried the CLI app. A few warnings to those who'll want to do the same:

- The app didn't have any updates since July 2024, then got a handful of commits last June and no recent commits ever since.

- I've tried to calibrate with it and managed to discharge but failed to charge back up. It's not a big deal since I could force it back with a `battery charging on`.

- There's also [some issues](https://github.com/actuallymentor/battery/issues/322) reported which [I've also experienced](https://github.com/actuallymentor/battery/issues/301) on my end.

So it works, but it has some complications to keep in mind. Apparently [someone forked it](https://github.com/js4jiang5/BatteryOptimizer_for_MAC) and aimed to fix some of the issues, but the fork is err, opinionated and may or may not be ideal.

> quickly test and run different models in the cloud that you could run locally if you had the correct hardware.

I feel like they're competing against Hugging Face or even Colaboratory then if this is the case.

And for cases that require strict privacy control, I don't think I'd run it on emergent models or if I really have to, I would prefer doing so on an existing cloud setup already that has the necessary trust / compliance barriers addressed. (does Ollama Turbo even have their Trust center up?)

I can see its potential once it gets rolling, since there's a lot of ollama installations out there.

The trick to this is to run a LiteLLM proxy that has all the connections to whatever you need to connect to and then point Open-WebUI to that.

I've been using this setup for several months now (over a year?) and it's very effective.

The proxy also benefits pretty much any other application you have that recognizes an OpenAI-compatible API. (Or even if it doesn't)

> Masaki had the idea to register a three-diamond trademark, along with the "Mitsubishi" name, which means 'three diamonds.' (It may surprise you to learn that this was ten years before the much better-known Mitsubishi Group of heavy industry companies registered its name and identical mark. Mitsubishi Pencil has no connection to the numerous other Mitsubishi companies in Japan; it is and has always been a manufacturer of writing and drawing supplies.)

Well, I learned something new today. I always thought the pencils were part of the group but apparently they're not.

Seeing the iconic three diamond mark along with the name always made me think the pencils were related to the cars.

After seeing this and the stories after this comment, I... can relate to this.

I went from being unorganized to somewhat organized, then went back and now it's a case of "I'll keep things organized when it makes sense, but the rest is up to my memory, the natural way of doing things and wherever I left it."

I'm just going to try the next thing, see and adopt whatever works, but if it doesn't, I'll just stick to whatever does.

At the end of the day, it's up to our brains on whether to use systems or not and if they fit our needs or if it doesn't.

Every time I see magic links, I always think: "I thought we weren't supposed to click links in emails in the first place?".

When links in email come into mind, so does phishing.

I hate these magic links a lot.

I worked around this problem by adding an entry in my password manager with a username of "OAuth: Use Google" or something like that, so I'm informed when I habitually check my browser extension or when I attempt to auto-fill during login.

It's inelegant and could be better, but good enough.

It looks like a multi-purpose utility in the terminal for bridging together the terminal, your scripts or programs to both local and remote LLM providers.

And it looks very handy! I'll use this myself because I do want to invoke OpenAI and other cloud providers just like I do in ollama and piping things around and this accomplishes that, and more.

https://llm.datasette.io/en/stable/

I guess you can also accomplish similar results if you're just looking for `/chat/completions` and such if you configured something like LiteLLM and connecting that to ollama and any other service.