PH热榜 | 2026-06-09

@yoann_berno3 Love the focus on fixing the biggest bottleneck instead of overwhelming founders with endless advice. Matching with investors and generating personalized outreach from your own inbox is a nice touch. Curious, what have you found is the most common issue holding pitch decks back?

@yoann_berno3 Love this intro

@yoann_berno3 Congrats on the launch! While I'm personally not in the VC space at all, very cool to see this exists and something I can share with some of my Clients who are. You saw the problem and created a solution for it - that's true entrepreneurial spirit!

@chrismessina Chris, thank you for the hunt, and honestly for framing the why-now sharper than I did in my own post. 🙏

The "instant termsheet" line got a laugh out of me, and you're right to caveat it. VC Boom won't conjure a term sheet from thin air. What it does is make sure that when you're finally in front of the right partner, your deck isn't the reason they pass and your email isn't the reason they ghost.

And the squeeze is real: SpaceX, YC Demo Day, and the summer cutoff all landing in the same few weeks. Founders who tighten the deck and start outreach this week are the ones who catch the window. The rest will be emailing into a void come July.

So to anyone reading who's sitting on a deck telling themselves "soon": Chris is right. This is your sign. Score it, see your matches, get in those inboxes. 💥

@yoann_berno3  @chrismessina Love the urgency in this launch too many founders spend months perfecting outreach instead of actually talking to investors. Congrats on the launch!

@ben_d3 Thanks Ben. You can actually do that in Outreach. It acts as a CRM, tracks your sent emails, reply rate. If you connect your Gmail account, it does it automatically for you.

@ben_d3 Great suggestion, Ben. You’re right, getting the first meeting is only one part of fundraising. Tracking conversations and follow-ups could make the whole investor journey much easier to manage.

@mattaussaguel Thanks Matt. And congrats on success of SheCodes! 🚀

@lakshminath_dondeti Sure. There are def general trendlines though. VCs will always put extra emphasis on team, narrative/vision, product market fit, traction.
Those are the dimensions we evaluate and score.

That way you know in 90sec what needs to be polished before you outreach them and burn bridges.

@gabriella_anjani Thank you 🙏 I built VC Boom because I see founders not knowing how to access investors, or literally spamming hundreds without finding the best way to approach them.

Neither works. One leaves you stuck, the other burns the relationship before it even starts.

So the whole point was to fix both ends: get you to the investors who actually fit, and help you reach out in a way that earns a reply instead of a delete. If you're raising, run your deck through the free score sometime. Thanks again for the kind words. 💥

@nox_v Yes perfect for pre-seed. What's your company about? Happy to provide some pointers.
You should also definitely score your deck to figure out whether it's ready for investor outreach

@thedatadavis Thanks Chris, glad it's helpful. What's your company about?

@liam007 we don’t save any deck, super strict about privacy.

It doesn’t train our model, and there’s segmentation between all users.

Your data always remain your data

@antifreeze Great question, and fair point, that one’s on me. The description doesn’t make this clear enough.

Sweet spot is pre-seed through Series A. That’s who it’s built for and where it works best.

It adjusts in two ways. The matching filters to investors who actually write checks at your stage and size, so if you’re pre-revenue you get angels and pre-seed funds, not growth VCs who wouldn’t look at you yet. And the scoring weighs things by stage: pre-PMF it leans on team, market, and how clearly you frame the problem, since there’s no traction to judge yet. Once you’ve got revenue, your metrics carry a lot more weight.

Where it’s less useful: pure idea stage with no deck yet (nothing to score), or Series B and up, where raises run more on relationships and warm intros than cold outreach to a fresh list.

So who benefits most, honestly: a founder actively raising pre-seed to Series A who doesn’t already have a warm investor network to lean on. If VCs are already chasing your round, you don’t need me. If you’re staring at a blank list, that’s exactly who this is for.

Thanks for the sharp question 🙏

@you_x_you_i Hey Emma, best is to upload your deck to get reality check. It will tell you immediatley whether you're raise ready.
How much are you raising? and what geography?

@millwiller hey Will, it points you the areas that need improvements and make suggestions. It doesn't redo your entire slides though.
you can then do modifications yourself and then reupload to have rescore. You can do that as many times as you want

@oleksii_sekundant Hey Oleksii. I've worked in VC for 8 years. I've met hundreds of investors in my time, and I've helped hundreds of founders fundraise.
I've also been aggregating angels/family offices and VCs lists every time I get invited to an investor event, conference, whatsapp group, or private mailing list.

The data is my secret sauce, and it's the hardest part to come by. It needs to be up to date, accurate, and continually refreshed as people move on from organizations.

The matching boils down to 7 criteria, with different weights, which I've personally crafted based on my investor judgment. So when you get your score and the single biggest area you need to improve, it's like receiving a feedback straight from a no-BS VC.
Better to see it via VC Boom than hearing it in a decline email from your target investors

@vb_30 Great to hear Vaishnavi. How much are they raising and from whom (angels, VCs?)

@sabber_ahamed Hey Sabber, sure there's a 20% discount for PH users. Code in the slide ;)

Which country are you visiting the website from? can you try different browser

@alina_tyslenok_ Thanks so much 🙏

And yeah, exactly. The guessing is the worst part. You rewrite the same slide five times with no idea if it's even the right fix. The whole point of the 90-second score is to take that guesswork off the table so you can just get on with the raise. 💥

@maali_baali Thanks a lot for the support. Looking forward to your product feedback. You're raising funds right now?

@hamza_afzal_butt Thanks, appreciate it 🙏 Speed is the whole point. Upload your deck, and a couple minutes later you've got 100 investors who actually fit, each with a reason why, plus a cold email drafted in your voice. No week-long list building. You can start reaching out the same day. Some investors book meetings within 48h

On tracking: right now it's discovery and outreach focused. But pipeline tracking keeps coming up, so it's on my mind. What would you want it to do?

@mike_milord Honest answer is yes, it absolutely works outside the US. Europe is very much on our radar, not an afterthought.

The matching pulls from 47k+ investor profiles that include European and global funds, not just US ones, so you get surfaced people who actually deploy in your region and at your stage.

Best way to know for sure: run your deck through the free score and look at the matches it gives you. If you tell me roughly where you're based and what you're building, I'm happy to eyeball whether the coverage is strong for your space.

This map shows Today's users. All over the world. In Purple are paid users from last hour 💥

@khashayar_mansourizadeh1 Thanks for the support. Great timing indeed

@anthony_adams_ Thanks Anthony! Planning a fundraise soon?

@coinbase  @its_maddy_a This looks incredibly promising for infrastructure cost control. Do you offer an automated orchestration layer that dynamically determines whether a task requires a small model or needs to be escalated to a frontier model?

@its_maddy_a  Really interesting approach, most teams are focused on pushing model capabilities, while you're optimising the inference layer where a huge amount of production traffic actually lives. The latency and cost improvements are impressive.

Curious: how does ZeroGPU handle routing and fallback when a specialized model encounters low-confidence or out-of-distribution inputs? Do you dynamically escalate to a larger model, or is confidence management handled at the application layer?

@coinbase  @its_maddy_a The narrative around AI usually sounds like this: Build bigger data centers. Buy more GPUs. Consume more power. But the truth is, the world can’t build compute fast enough to keep up. So, we decided to take a different path.

Thanks to ZeroGPU team our partners — AI infrastructure powered by small language models running on a hybrid edge network. Instead of waiting for new hardware, we are unlocking and reusing the compute that already exists.

The reality of AI in production is simple: Not every task needs a frontier model. By using our purpose-built, edge-optimized models, companies can:

• ⚡ Run 10x faster than traditional setups.

• 💰 Cut costs by 50%.

• 📉 Offload 70–80% of production tasks to small models—all while maintaining frontier-level accuracy.

Here we don't need a bigger footprint; we just need smarter architecture. Welcome to the future of efficient AI. 🚀

@kevinzrzgg Great question. On the infrastructure side, failover is automatic. Requests are routed based on availability, capability, and load, so if edge capacity is unavailable, traffic can move to the next tier without you having to manage it.

Model escalation is different. We don't silently swap in a larger model because that makes costs and behavior unpredictable. Instead, our specialized models return confidence scores, and you decide the threshold for when something should be escalated to a frontier model.

So availability and latency are handled automatically by the platform, while quality and cost remain under your control. That separation is what makes the system both efficient and predictable.

You can learn more about it here - https://docs.zerogpu.ai/docs/how-zerogpu-works

Thank you!

@kevinzrzgg Fair points to press on. Let me take latency, cost, and reliability one at a time, because we handle each structurally rather than with per-request guessing.

Latency comes from running our models on an edge network — inference happens closer to where the work is, not in a distant data center. Cost comes from not depending on GPUs for these workloads, so the economics are fundamentally lower, not just discounted.

We're not trying to make a small model reliably do hard reasoning. We focus on the workloads where a specialized model is the right tool: summarization, scraping, data extraction, classification, PII detection, moderation. We're not competing with Claude on coding or complex reasoning — we handle the work where Claude is overkill. (use cases here)

For agentic workloads, our plugins only take a step when it's trivial enough for a small model to own cleanly — anything beyond that defaults straight back to your base model. The routing decision happens by task suitability up front, not as a recovery from failure.

Hope that cleared your doubts. Appreciate your support!

@mathew_chang Great question. The simplest way to think about it is that the structure of the task usually determines the model.

If it's something repetitive with a fixed output shape, like classification, tagging, PII redaction, or field extraction, a specialized model can typically do it much faster and cheaper without sacrificing quality. If the task requires deeper reasoning, synthesis, or judgment, that's where frontier models still make sense.

A lot of teams automate this by putting a lightweight ZeroGPU model in front of their frontier model. It handles the straightforward requests and only escalates the ones that actually need advanced reasoning.

Happy to go deeper if useful!

You can learn more at - https://docs.zerogpu.ai/

@mathew_chang Great question — you can achieve that following ways using Zerogpu:

One way is to use our MCP and Claude/Claw plugins which help decide which small model handles a given step on the fly. Say you're running a Claude agent to scrape and qualify potential clients — ZeroGPU handles the summarization and data extraction while Claude focuses on the higher-order judgment calls. Plugins here: https://docs.zerogpu.ai/integrations/claude-code-plugin

For production workflows - you identify the repeatable workloads like classification, summarization, data extraction, and integrate with our open AI compatible model endpoints.

Either way, the principle is the same: frontier models for the complex reasoning, ZeroGPU for the high-volume repeatable work.

@shawn_idrees If you'd like to explore it before spending time on a full evaluation, the API docs are probably the best place to start: docs.zerogpu.ai/api-reference/responses.

ZeroGPU is OpenAI-compatible, so the request format should feel very familiar. There's also an interactive playground and dedicated pages for the classification and extraction models, where you can see example inputs, confidence scores, and response formats.

And of course, if you'd like a recommendation for a specific use case, feel free to share a bit about your workload. We'd be happy to point you in the right direction.

@nemanja_igic Its been a ride, but this is just beginning. We are on to something big! Thank you!

  @nishitha_t Yes rewarding and grueling. We are solving hard problems and thank you for being part of this journey. Upward and onward from here.

@peterbwf Totally! The associated cloud costs with increased latency is often overlooked. Every frontier request that takes more than >2 secs to respond is burning your lambda costs or any other provider you are using.

We did not account these cost savings in our numbers. If we do we will end up being 70% cheaper. :)

@omri_ben_shoham1  This is exactly the kind of workload ZeroGPU was built for.

The integration is pretty simple: point your existing OpenAI client at the ZeroGPU base URL, swap the model name, and your classification and extraction calls keep working.

One tip if you're processing at scale: use the Batch API instead of looping the sync endpoint. It handles up to 50k requests per job, avoids per-request rate limits, and is where the biggest cost savings usually show up: https://docs.zerogpu.ai/docs/batch/index

Would love to hear what the numbers look like on your traffic! 🚀

@alina_tyslenok_ Thank you! That's exactly the idea. Frontier models are incredible, but a lot of AI volume is repetitive work that can be handled much faster and cheaper with specialized models.

@praneethpike We actually don't need any GPUs. Our models are optimized and trained to run on CPUs. We also support models from hugging face that are optimized for edge and fine tune them to different domains and use cases.

So yes we are faster and cheaper. I see a lot of startups struggling to maintain AI features because of the token bill, this is especially true in developing countries where these costs cannot be passed down to the users.

We are here to make AI more accessible - this tweet by Brian Armstrong from @Coinbase sums up really well.

@its_maddy_a  @thisiskp_ very exciting stuff - congrats!

@akshay_arvapally Thank you @akshay_arvapally

@oleksii_sekundant Our models are super light weight making cold starts faster - since they are also edge network the cold start is not a huge latency overhead. For high token volumes we always reserve instances and adapt fast for high spikes.

We also support batch processing which ends up being even more cost efficient.

@sneha_reddy12 Fair catch on the name. It's not aspirational we optimized our models to run on CPUs and edge devices, so there's no GPU provisioning and no competing for scarce datacenter GPUs. Our models can run anywhere.

When we mention cloud fallback, it's about consistently delivering on our response-time promise and because our models also run on-prem and within VPC as well, it's how we support enterprise deployments.

But no GPUs were harmed in making ZeroGPU.

@maksim_ovsienko We do support BYO-model for high-volume customers. In those cases, we can help fine-tune, evaluate, deploy, and scale task-specific models. It is not part of our self-serve flow yet, but making that full loop self-serve is one of the top items on our roadmap.

We’re a small team moving fast, so we’re being intentional about not biting off more than we can chew. But the direction is clear: for companies running serious production volume, this often becomes a build-vs-buy question across fine-tuning, deployment, scaling, observability, and cost optimization. Our goal is to make that entire path much easier in one place.

Appreciate you sharing PJQ, that’s a great real-world example of why this layer needs to exist.

@gizem_ozturk Yes our models outperform frontier nano models in repeated workflows. For example our client @Dappier has seen 10x faster responses and 50% cost reduction with GPT-5.4 nano level intelligence. Our models also hallucinate less as they are specialized and are trained to do one task but do it perfectly.

Thank you for your support!

@dmitriychuta Thanks for your support! Yes, there will be an iOS version, and even an iPad version is planned.

@farrukh_butt1 Thank you! It's true, I wanted to create a unique and interesting new experience. Because most apps look pretty similar.

@craigcpaterson Craig, thank you so much! I really appreciate your support!

You're one of our earliest users, and that's incredibly inspiring for any indie developer!

@richie_giordano Thank you! It's worth it!

@rkuhcvark Thank you so much! I really like the name myself.

Of course, I'm a little concerned that it might be difficult for some users to pronounce, but that's something I'll figure out later.

@david_marko Oh my God, thank you so much! This is exactly the kind of experience I had in mind when I created this app.

It means a lot to hear that it helped. I'm really happy that I was able to share my experience and make a positive impact!

@varun1jan Thank you!

I use it for much more than just work. It's become part of my daily routine – whether I'm doing household chores, relaxing, or even reading a book while listening to nature sounds 😁

@oleksii_sekundant Yes, I definitely believe that sound placement has a significant impact on both perception and focus.

At Ubisoft, I worked as a UX designer, mainly focusing on accessibility features and player experience, so I wasn't directly involved in audio design, and my experience isn't really relevant here.

@adriano_villa_bascon Thank you so much!

This looks like an apple-side bug, I haven't seen this happen before. I'll take a closer look and see what I can find.

For now, please try reinstalling the app, as that might fix the issue.

@kay_larina Thanks a lot!😌

@alexkhlystova Thanks! Have you tried the headtracking feature?

@alina_tyslenok_ Thank you so much!

@olexiy_vasylenko Thank you so much for your support! Yes, it's very relaxing, especially when I turn on the noise-cancelling headphones😌

@nikitaeverywhere Thanks! Regarding headphones, Spatial Audio works with any headphones, but the HeadTracking feature only works with AirPods models that support it, but it feels really cool!

@ferdi_sigona Thank you! Yes, that's the whole point of the app. I wanted to create a tool that would make it easy to customize your personal space.

@leventbuilds Thanks for your support! I've tested it on many devices so far, and there hasn't been a noticeable load during long-term use, but I'll continue testing.

Regarding presets, they're already available in the Pro version, you can save an unlimited number of presets. There's no schedule yet, but that's an interesting idea, I'll add it to the backlog.

@asti_pili C++ SDK timeline? Any plans for Go, Rust, or mobile SDKs

@asti_pili This is impressive work—the fact that you've handled over a million minutes of production translation in regulated industries with zero incidents is a strong signal of real reliability. The 96% accuracy on live calls with real-world noise is the kind of number that matters way more than lab benchmarks.

@igorgurovich good questions.

We measure latency in Krisp as the time to first translated audio after a person speaks.

• Total time-to-first-translated-audio is approximately 1.5–3 seconds, driven by three factors: context window size, source language structural complexity, and amount of speech. AI inference latency is around 700-800ms here.

• Language structural complexity is the primary variable. Languages with word order parallel to the target language (e.g., Spanish) can be translated incrementally as words arrive, resulting in latency toward the lower end.

• Languages with high reordering distance — such as Japanese, Korean, or Turkish — are verb-final or agglutinative, requiring the model to buffer more context before producing a grammatically correct translation, resulting in latency toward the higher end

• The AI latency difference between transcription and translation is ~100ms.

@rrhoover yeah, way back to those working from home days during covid - when we launched Noise Cancellation.

@jared_salois great question.

Yes, we support locale-specific variants (US Spanish, French Canadian, Egyptian Arabic, etc.) and regional languages (Catalan, Galician, Basque)

@marija_pojasnikova The API supports custom vocabulary, so you can pass your specific terminology. We also support a custom translation_dictionary, where you can provide an exact word and its translation for each language — that word will always keep your translation.
Here's the documentation on how to pass these parameters: https://sdk-docs.krisp.ai/docs/voice-translation-api#initial-client-message

@christian_knaut We measure latency in Krisp as the time to first translated audio after a person speaks.

• Total time-to-first-translated-audio is approximately 1.5–3 seconds, driven by three factors: context window size, source language structural complexity, and amount of speech. AI inference latency is around 700-800ms here.

• Language structural complexity is the primary variable. Languages with word order parallel to the target language (e.g., Spanish) can be translated incrementally as words arrive, resulting in latency toward the lower end.

• Languages with high reordering distance — such as Japanese, Korean, or Turkish — are verb-final or agglutinative, requiring the model to buffer more context before producing a grammatically correct translation, resulting in latency toward the higher end

• The AI latency difference between transcription and translation is ~100ms.

@fujibee the runaway-loop case is what i'd stress-test first: A asks B to clarify, B bounces it back, and they burn tokens in a clarify-loop with nobody refereeing. the tic-tac-toe demo is the harmless version of the same loop. so the real question is whether the stop condition lives in the protocol, or it's left to each agent's prompt to know when to shut up.

@fujibee go go! 🚀

@fujibee Brilliant Koichi, congrats on the launch. As you can see, you're not the only codex-claude courier. How do handle context across sessions?

@busra_seker1 

Ha, glad it landed 😄 — that courier feeling is the whole origin story. It's built for N agents: a "team" is just a room, and you can drop several in. Right now it's running with 5 different agent types and multiple instances in one room. What gets messy past ~2-3 isn't the transport, it's turn-taking — everyone wants to talk at once, so you need light addressing (who's this for?) or a coordinator. The DB handles N fine; the social protocol is the hard part. 🙂

@webappski 

You zeroed in on the part I care about most. Everyone reacts to "agents talking," but the persistence is the actual unlock — `sqlite3 agmsg.db` and the whole thread is right there, no special viewer. That exact pain (hand-rolling a state file so decisions survive plan→implement→review) is what made me build it. A room that outlives the session = your state file, minus the bookkeeping. Glad it landed with someone who's felt it. 🙏

@eliot_argy 

"the human as the message bus" — that's exactly the phrase I wish I'd put on the landing page. 😄 That's the whole itch: two systems smart enough to just talk, and there I was hand-carrying their messages.

The SQLite "room" was almost the dumbest possible design, which is why I like it — no daemon, no network, you can `cat` the DB to see the whole conversation. Curious what agents you're running together — would love to hear what breaks. 🙏

@luca_capone 

Honest answer: agmsg only carries what the sending agent actually writes. It's a transport, not a semantic layer — so the "why" survives only if the first agent spells out its intent in the message. It won't magically preserve reasoning the agent kept to itself. In practice that's a feature and a trap: I prompt the sender to lead with intent ("here's WHAT I did and WHY") before the diff. Then the second agent stops confidently rewriting the right thing. Your "loses the why" pain is the exact failure mode I'd watch.

@nao_yukawa LET'S GOOOO 🚀🙏 thank you Nao!

@juhi_sharma6 not a dumb question at all 🙏 — agmsg is built for the CLI coding agents (Claude Code, Codex, etc.) messaging each other, so it's not for moving a normal Claude chat into another app; for that you'd still be copy-pasting yourself. But "why isn't there an easy way to carry a conversation across tools" is a genuinely good question — just a different problem than the one agmsg solves.

@qifengzheng no daemon was non-negotiable, glad that lands 🙏 and you've found the real gap: agmsg has no claim/lock primitive — two instances can absolutely grab the same task. Today I solve it socially (a coordinator agent hands out work, or an "I'm taking X" message + addressing). You could lean on SQLite for a real claim (an atomic UPDATE ... WHERE status='open' as a lease), but agmsg doesn't ship that — it's transport. A claim/lease table is the most-requested thing I'm chewing on.

@syntheticfounder interesting — agents critiquing each other's output to converge on something better is exactly what the shared "room" is good for. The tic-tac-toe demo is the toy version of that loop. Curious where you take it for the YouTube angle 🙂

@david_marko that courier feeling is exactly it 🙏 one tiny correction — agmsg isn't over MCP, it's standalone (just bash + one sqlite file), which is what kills the dependency hell you mentioned. But you nailed the rest: persistent across sessions, grep the history later. Thanks for following!

@oleksii_sekundant it carries whatever the sending agent writes into the message — so if it writes the diff + the why, that's what lands; if it only writes a one-liner, that's all that crosses. agmsg doesn't auto-capture diffs or context, it's transport. In practice I have the sender paste the diff + intent explicitly. Upside: it's all just text in a row you can grep later.

@artstavenka1 The latter, exactly. sqlite guarantees durability + ordering (no dropped messages, consistent order by rowid); agmsg does zero turn arbitration — no lock, no token. Whose turn it is stays the agents'/prompt's problem. Deliberate choice: keep the transport dumb, push coordination up to the agents (or a coordinator agent in the room). And agreed — "persistent room" undersells nothing, it's the part I care about most too.

@galdayan ha, the "embarrassing in hindsight" part is universal — that's literally why I built it 😄. 2-agent is the rock-solid case, exactly what you need. 3+ works on the transport side (sqlite handles N fine), but turn-taking gets messy — you start needing light addressing (who's this for?) or a coordinator agent. The DB scales; the social protocol is the hard part.

@anand_thakkar1 SQLite handles this one for free — writes are serialized (single-writer lock, WAL mode for concurrent reads), so two agents posting to the same channel at once don't corrupt anything; the inserts just get ordered. Each message is a row, ordering is rowid/timestamp. So "nobody drops a message, order stays consistent" is guaranteed at the DB layer. What SQLite does NOT decide is whose turn it is — that's still the agents' problem.

@retain_dev Great question — and honestly agmsg doesn't solve it for you. It's transport, so it'll happily store a message bigger than the receiver's window; managing that is on the agents. What works in practice: send a summary + a reference (a key/rowid) instead of dumping raw output, and let the other agent pull detail on demand — the queryable room makes that pull cheap. But agmsg itself doesn't chunk or window, deliberately.

@james_effarah1 Thank you so much!

@james_effarah1 You had me at "de-slop" lol

@peterclaridge Thank you! Would love to hear your thoughts after using it :)

@luca_capone The specific rules per design system include layout instructions, spacing and so on so they are quite broad. But it only works if you tell your agent to actually rebuild your UI, instead of just redesigning it. From experience, agents will usually not make larger changes until specifically instructed :)

@zaczuo Interesting combination of features, especially the ability to analyze large batches of files alongside real-time web search. The enhanced image understanding also caught my eye. Curious, which use case are users gravitating toward most right now: research, coding, or content creation?

@new_user___1462026db37a4cc3a5c424a Love the calm-surface idea. Quick question for Mac power users and devs: what are the tiny, repeatable interruptions in your current Mac workflow that steal your focus and which of those would you most want BooBar to surface as a single actionable card instead of a full app notification?

@oleksii_sekundant Thanks! BooBar is mostly system-wide today

It does have some app-connected actions already, but they are tied to specific integrations rather than being a full per-app action layer. For example, it can open related GitHub items, jump back to monitored web pages, show coding session status, handle download/file review actions, and surface mail links or codes.

The direction is definitely to make this more app-aware over time: richer actions per integration, better deep links back into the source app or page, and clearer controls for what each app can contribute.

@dhiraj_patel5 Yes, BooBar is built to track multiple Claude/Codex sessions at the same time.

The idea is that it can surface several active coding sessions in the island/task panel, show which ones are running, completed, or waiting for approval, and let you jump back to the relevant session when needed. So it’s not limited to a single session.

@daniel_henry4 

By default, BooBar does not scan your whole Mac. You choose the folders BooBar watches, and it only reacts to files in those configured locations.

For many files, BooBar can use local metadata, filenames, paths, extensions, and extracted text snippets to suggest an organization result. You can also run it in local-rules mode, where files are organized without sending content to any cloud AI provider.

For sensitive files, the current recommended setup is:

only add specific watch folders you are comfortable monitoring

exclude private folders and file extensions in settings

use local rules or local LLM analysis instead of cloud AI

manually confirm lower-confidence organization suggestions before files are moved

A future improvement we’re considering is a dedicated “Sensitive File Guard”: BooBar would automatically detect patterns like bank statements, IDs, contracts, private keys, medical files, payroll documents, and password-like content, then either skip AI processing entirely or force local-only/manual review mode. This would make privacy protection more automatic, while still keeping file organization useful.

@tina_chhabra Exactly. That chaos in Downloads is basically the problem BooBar is trying to fix.

The File Card is meant to make every new file explain itself first: what it is, a short summary, useful keywords, suggested folder/name, and whether it needs manual confirmation. So instead of throwing everything into one pile, BooBar turns new files into small, reviewable decisions.

@fberrez1 
Yes, partially. BooBar currently supports OpenAI-compatible AI endpoints, so if your local LLM server exposes an OpenAI-style API, such as Ollama, LM Studio, or another compatible gateway, you can usually point BooBar at that local base URL and use it that way.

That said, we still consider first-class local LLM support a work in progress. The current path depends on how complete the local server’s OpenAI compatibility is, and some model/tooling behaviors may differ from cloud providers. We’re planning to make this smoother with clearer presets, local provider docs, and better fallback handling soon.

@leventbuilds 
Thanks! BooBar is designed to stay lightweight: the menu bar watcher uses local timers and macOS-native event hooks where possible, and most monitors only wake up when there is something relevant to check or show. The current build is still early, so we are continuing to tune polling intervals, idle behavior, and resource usage, but the goal is for it to feel like a quiet background utility rather than a constantly busy app.

On privacy: the default direction is local-first. File watching, local file metadata, screenshots/notes, reminders, and the desktop context surface are handled on the Mac. Sensitive items like email verification codes are extracted locally for quick access. BooBar does not need to upload raw email contents or local file contents just to show those desktop highlights.

There are a few optional features that may contact external services, such as subscription/license checks, web page monitoring, or user-enabled AI features. We are working on making those boundaries clearer in the UI and docs, including exactly what is processed locally, what may leave the Mac, and how to disable network-backed features. Privacy controls and more transparent resource reporting are near-term improvements.

@ibl_prd congrats on the launch Ihor. How much control is there of effects after generation?

@oleksii_sekundant Yes, absolutely! The effect is completely dynamic. You can change or edit the shape at any point, and the liquid-metal effect will adapt to the new geometry. If you modify the shape or generate a new pattern, you just need to re-run or update it within the plugin, and it will automatically calculate and adapt the effect to the fresh form. No need to start from scratch!

@german_merlo1 Thank you so much! Truly appreciate the kind words and the support. 🚀✨

@tina_chhabra That’s 100% true. While it’s obviously not a replacement for full-scale 3D production, it’s a massive quality-of-life upgrade for product and visual designers. Being able to skip the render pipeline for quick, high-fidelity visual assets is exactly the shortcut we needed!

@keirodev To be honest, I originally built Fluido just to scratch my own itch for a project I was working on 😂, but during the process, I realized how much time it could save others too. It’s awesome to hear that it's helping solo devs bypass the visual struggle and get polished, high-fidelity assets in just a few clicks. Thank you so much for the feedback!

@szamski congrats on the launch Maciej. What would you say is the biggest friction remover in hora?

@gabriella_anjani That browser tab thing is exactly what broke my flow every day. The menu bar widget came from that frustration directly - one glance, next meeting, join button. No switching.

Glad the privacy angle lands. It's not marketing, it's just how the architecture works. No backend means nothing to store. After I developed my skills more and understood how stuff works not only looks I was astonished that we are giving so much of our data not only to the big players such Google, but also others totally for free, or sometimes we are even paying for this :)

@curiouskitty Good question, and honestly one I think about a lot.

My bar for AI in a calendar is simple: it has to reduce the number of taps or decisions, not add a new surface to manage. If an AI feature requires me to review, confirm, or fix its output more than once a week, it failed.

On privacy: anything that touches event content, titles, or attendees stays on-device. Apple Intelligence is the right tool for that, because it runs locally and I don't have to build a backend that sees your calendar data. That's not a philosophical stance, it's a product constraint I locked in early.
hora has no backend. It connects directly to Google Calendar API and that's it.

The three features on the roadmap reflect that:

1. Quick add via natural language: on-device, processes text locally before sending a structured request to Google Calendar API. No content leaves your Mac except the final event payload.

2. Focus time planning: reads your existing calendar locally, suggests blocks. No server involved.

3. TL;DR meeting summaries: this one I'm still figuring out. Summaries require reading event descriptions and potentially linked docs. I'll likely keep it opt-in and on-device only, which limits what it can do, but that's the right tradeoff.
   

How I'll measure success: time from "I need to block time" to "done". If it's not faster than typing, it ships as an experiment, not a feature. I'm tracking session length and interaction counts in TestFlight already, so I'll have a baseline. Obviously Apple yesterday showed us some new things, so it'll help a lot, but it'll be invisible for the user as something that works "just fine" not is another AI hype tool, that after a while feels redundant.

@a_petukhov Thanks! That was exactly the friction I wanted gone. Jumping between the calendar and three different apps to find the right link was driving me nuts, so hora pulls the join link straight from the event and puts it one click away. Really glad it lands 🙏

@marianna_tymchuk Thanks. It's also just the simpler architecture - fewer moving parts, fewer things that can break.

@anastasia_chavdia congrats on the launch Anastasia. How do you solve the "actually right for me, on this trip" question and how long does that take?

@anastasia_chavdia Congratulations!!

The 'taste not reviews' positioning is smart — review fatigue is real, especially in travel. Would love to see this paired with live deal discovery so when TravelMind recommends a restaurant or hotel, it surfaces the best current price too. The discovery + savings combo is underexplored. Upvoted!

@artstavenka1 Yes — your taste graph travels with you. Think of it like a friend who already knows you, just landing in a new city with you.

@gabriella_anjani Thank you, Gabriella — that workflow you described is exactly what we're trying to make obsolete.

Would love to hear your honest take if you try it!

@denitsapenchevavaltchanova This is exactly the kind of feedback that matters — thank you, Denitsa.

Community is the most valuable asset in this space right now — and experience sofia is a perfect example of that. The best places will always be found by people, not algorithms.

You can already add your spots on TravelMind and share them with your audience. Sofia through your lens would be something special.

@jared_salois Thank you! That's exactly what we're working on. Stay tuned! 😄

@tina_chhabra  "Zero overlap with the average Google reviewer" — we feel that deeply too.

You've basically described the entire reason TravelMind exists. Ratings tell you what everyone thinks. We want to learn what you think — and get better at it every time you swipe.

The short answer: it gets personal faster than you'd expect. But we'd rather you find out than take our word for it.

@evakk You've described the problem better than most. The noise is the issue — not the lack of information.

We're live in a number of cities already and actively growing — the honest truth is that the recommendations get sharper the more people engage in a given place. So the best answer to your question is: try it where you are, and tell us what's missing. That feedback is genuinely how we prioritize.

@clemens_vogelhaus Kudos on the launch. Could you maybe give a walkthrough of how it analyzes Apple Health? Like how does it understand our level of fitness or position in our fitness journey and our goals? Do we get to edit it?

@marianna_tymchuk glad you like it!

@ferdi_sigona yes, that's the key difference. The Bevel release was really exciting for us. We've been working on Whistle for more than a year now, and seeing it confirmed a lot of our assumptions. As far as we can tell, Bevel focuses on helping you make sense of your fitness and health data. Whistle is primarily a workout planner. It lets you create workouts, schedule them and send them to your Apple Watch so you get real-time guidance during your session.

The health data we show in the app serves two purposes: it gives you guidance, and it gives the coach relevant context. The coach itself is really just an additional layer that can do all of this for you. It can help you understand your health data, but where it shines is creating plans and workouts. When it plans a run, it doesn't just message you to say "go for a run." It actually schedules the run in alignment with your goal, your training load, your recovery, the weather and more, and it generates structured workouts, so a run might include intervals with specific heart rate targets or paces whenever that makes sense. With the same approach it can build plans across multiple months to prepare you for a race or to work toward a bigger goal.

When you miss a workout or a key session Whistle can easily adjust your plan around it. It's actually the biggest benefit compared to rigid training programs. We also plan to add additional features that we hope will increase motivation and help people stick to what they set out to achieve.

Happy to hear it helped :D I am sure your cycling isn't that terrible

Hey @curiouskitty ,

Whistle handles this through Coach tools rather than a hidden black-box rule engine. For example, before changing a week it can read the current plan, recent completed workouts, planned workouts, training load, sleep, recovery, and overnight vitals. If Thursday has intervals planned but the last few sessions already pushed load up, sleep has been poor, and recovery/vitals look off, Coach can propose a smaller change: move the hard session, reduce the interval volume, swap it for easy aerobic work, or move strength away from another hard day.

The important part is that plan changes are explicit. Whistle can adjust an existing workout, create a replacement, or remove a conflicting one, but those write actions go through user-facing approval/permission flows. The training program also stores coaching context such as progression, load strategy, recovery spacing, adaptation rules, and “what not to do,” so future planning turns preserve the intent instead of just reacting to one metric.

We’re careful not to frame sleep or vitals as medical diagnosis. They’re used as wellness and readiness context, alongside training history and the plan, to make conservative adjustments the user can review before applying.

Hope that answers your question,

Clemens

Hey @gsostak ,

thanks! Whistle works best for endurance sports, but it can also generate strength workouts and provide exercise routines for home or the gym. Right now, though, the app just tells you to do a certain exercise for a certain number of reps and can't yet explain or show how to perform it correctly.

Hope that answers your question.

Clemens

@jdilla Very cool concept. Giving coding agents the ability to validate geometry and catch their own mistakes before handing off printable files feels like a big step toward practical CAD automation. Curious, what kinds of parts are users generating most today: simple prototypes, mechanical assemblies, or production-ready components?

@zaid_mallik1 the agents will always generate something so producing quality is definitely the bigger challenge.

@willie_dev Love the focus on clarity during demos and presentations. The cursor spotlight and keystroke overlay seem especially useful for tutorials and screen shares. Curious, have you found users relying more on the spotlight feature or the on-screen annotations?

@willie_dev Nice work man!

@fatalerrorist Really like the idea of using diagrams as the source of truth and generating the repetitive architecture automatically. Offloading the boilerplate 80% lets developers spend more time on the business logic that actually matters. Curious, what kinds of projects are teams using Solarch for most today?

Saw the VS Code extension on the roadmap with bi-directional sync. That's the part I'd want to understand better. If I generate a service, then add custom logic in the generated file, and later go back to edit the node in Solarch, does the sync know to preserve what I wrote? Or does it treat the generated file as owned by the engine and overwrite everything?

Separate question on node types. The built-in set right now is tables, DTOs, services. If I need something the rules engine doesn't have a template for, like a queue consumer or a scheduled job, can I define that as a custom node type, or is the vocabulary locked to what ships with the platform?

@samir_asadov 
The entropy in financial models is the exact same problem as architectural drift in software. If the map doesn't match the territory, the system becomes a liability. I completely agree with the principle behind ModeLoop—structure must remain legible.

To answer your question directly: right now in v1, we enforce the diagram as the absolute source of truth. It is a strict one-way compiler. If a developer manually changes the generated boilerplate faster than the diagram, the next export will overwrite it.

However, we are fundamentally solving this in our upcoming VS Code extension (shipping in a few weeks). Instead of treating the generated file as 100% owned by the engine, we are implementing AST (Abstract Syntax Tree) parsing and a Git-based control flow.

This allows the engine to recognize manually injected custom logic blocks, perform a smart diff, and merge visual architectural updates without crushing the developer's manual code.

@zaid_mallik1 

You just saw the exact endgame.

Historically, diagrams are just dead pictures that rot within a week. It's a broken loop. We’re turning the canvas into a literal compiler.

And you are 100% right about AI agents. If humans and AI are going to build complex systems together, they need a strict, deterministic visual language—otherwise, it's just pure hallucination. This is exactly why integrating MCP is our next move. Solarch isn't just a devtool; we are building the interface where human architects and AI agents will collaborate in real-time.

@andreas_rubin_schwarz Thank you for your feedback!

@akgulkazim Really interesting vision. Treating AI agents more like a company with approvals, tasks, and runtime visibility feels much closer to how real teams operate. Curious, what workflows are users coordinating most successfully with AgentOS today?

@tina_chhabra Exactly — this is one of the core problems we’re trying to solve with AgentOS.

OpenClaw is powerful, and our goal is to make it more manageable, transparent, and accountable for real operations.

Today, every agent already has visibility around which model it uses, task/session activity, token usage, and task-level execution details. From there, the natural next layer is operator controls: per-agent budgets, task-level cost tracking, spend alerts, loop detection, and kill switches before anything gets out of hand.

So yes — the goal is not just “run more agents.”

It’s to make agent work visible, measurable, and controllable.

@kohlhofer "Congrats on launching! With the local MCP server integration, what kind of prompt schemas can tools like Claude Code use to query the local recording archive?

@kohlhofer Hey, congrats on the launch!

Curious, how are you collecting feature requests and feedback from users right now?

@qifengzheng the MCP has a decent set of tools to help with that. An agent can:

• Run searches

• Understand the project structure

• Get an overview of a folder

• Access and edit one more recording/transcript

• ...

So it can work fairly efficiently. But in most cases, the markdown files on disk are the most performant.  Claude Code, for example, can directly access them and use its own tools and patterns to manipulate them on a system level.

I have a journal folder, and once a week, I process the various daily entries into a weekly summary with Claude. It even puts them back into SR-7 in a dedicated project for the weekly entries.

Some of the tools:

@tony2742 Love the privacy-first approach with local AI. Automatic organization and duplicate detection sound especially useful for keeping messy folders under control. Curious, what has surprised you most about how people are using NudgeFile so far?

@naimz Absolutely 👍

NudgeFile is designed to keep you in control.

✅ Every AI suggestion can be reviewed before it's applied.

✅ Workflow Automation supports approval-based actions, so you can choose whether files require confirmation before being renamed or moved.

✅ All file activities are logged, making it easy to see what happened.

✅ Rename operations are reversible, allowing you to restore the original filename if needed.

For users who prefer full automation, approval steps can also be disabled so workflows run automatically in the background.

The goal is to save time without taking control away from you. 🙂