Mark Tofts

@cathrynlavery @mvanhorn Followed this pointer and got it working today 👌

@mvanhorn Thank god for terminus and TailScale

Fighting the urge to still bring my laptop everywhere for Mother’s Day. Staying strong. Keeping it home. Happy Mother’s Day Lauren!

I do find the Granola summaries a bit reductive. 90 min meeting > “here’s a handful of bullet points” My Hermes agent built a skill that pulls the full transcript via API, puts it in obsidian and emails a summary and follow up points from a custom prompt. Full transcript is then local and searchable.

Ripped out whisprflow and Granola is next. Loved both of those products and implemented them at our company but now they suck and aren't needed. Both are replaced with muesli which is open source and hosted on your machine. The real ROI will come in replacing granola. Instead of having Granola's shitty summaries, shitty api/mcp, and no ability to export transcripts—our raw transcripts will now feed into our database with all of our company data and business logic. Sensitive conversations will now stay on our servers and our agents will have access to raw transcripts with our company context on top of it. Will be a huge unlock for us.

@jphorism @Teknium I’m using agent mail - easy to set up and no issues so far agentmail.to

What's everyone doing for email in Hermes agent? Have seen a couple friends mirror their emails locally to md or sqlite. Curious what use cases benefit that vs plain using gws @Teknium

Using Hermes to assess potential sites from 6,000 postcodes. - Pull aerial image from postcode - Assess site characteristics - Calculate links to major roads - Calculate population within a 15 minute drive

@oliverhenry Best I've seen so far elevenlabs.io/blog/introduci…

I REALLY want voice models to start understanding tone of voice.

@ironcarbs which is the best solution you've seen so far to counteract this?

@zachtratar It is in English but I still think Granola or Notion make transcription mistakes that, when summarizing, can render the whole conversation useless because it missed a critical point. Maybe it will just get better as models improve but I prefer more accuracy for now.

Our retention on Notion AI Meeting Notes just hit a crazy high number. Growing like crazy as people switch from Granola and Fireflies. ONE top customer = 2500 hours per week. People are loving it! 😍 And we have a couple more features coming soon... then we can exit beta.

@thekitze Just saw that this is now an actual @ChatGPTapp ad

i asked gpt-image-2 to 90s-maxxx our boring ass kitchen and i just want to ask WHERE DID WE GO WRONG AS HUMANS WE USED TO HAVE IT ALL

@alexcovo_eth From my Hermes to yours 😄 Shoutout to @alexcovo_eth — followed your Hermes gpt-image-2-high setup notes, patched the config, tested it live, and it worked. Properly useful. Thank you.

Wasn’t me. One of my Hermes agents did it for me. I asked it explain. Copy/Paste to your agent and it will implement for you 👇 Hermes Agent image generation setup: gpt-image-2-high via OpenAI Codex auth Yes — in our Hermes profile, gpt-image-2-high is the default image generation model. Current config: image_gen: provider: openai-codex use_gateway: false model: gpt-image-2-high How it works: 1. Hermes uses the openai-codex image generation plugin. Plugin path: plugins/image_gen/openai-codex/__init__.py The plugin registers an image provider named: openai-codex 2. The normal Hermes image_generate tool checks image_gen.provider. Tool path: tools/image_generation_tool.py If image_gen.provider is set and is not fal, the image_generate tool dispatches the image request to the matching registered image-generation plugin. Since our config says: image_gen.provider = openai-codex Hermes routes image generation through the OpenAI Codex image plugin instead of the default FAL path. 3. The OpenAI Codex image plugin uses ChatGPT/Codex OAuth, not a normal OpenAI API key. It reads the existing Codex OAuth token through Hermes’ Codex auth machinery. That means the setup depends on Codex login/auth, e.g.: hermes auth codex It does not require OPENAI_API_KEY for this image path. 4. gpt-image-2-high is a Hermes tier alias. The actual OpenAI image model used by the image_generation tool is: gpt-image-2 Hermes maps: gpt-image-2-high to: gpt-image-2 + quality=high The plugin supports these tiers: gpt-image-2-low gpt-image-2-medium gpt-image-2-high 5. The plugin calls the Codex Responses API with the image_generation tool. Codex backend: chatgpt.com/backend-api/co… Host chat model used inside the plugin: gpt-5.4 Required tool: image_generation Actual image model: gpt-image-2 Selected quality tier: high 6. We verified it live. The plugin resolved as: provider = openai-codex available = True resolved_model = gpt-image-2-high quality = high We also previously generated a real StoryForge storyboard image through Hermes with: provider: openai-codex model: gpt-image-2-high quality: high size: 1536x1024 Summary: Hermes chat and Hermes image generation are configured separately. Normal chat model: gpt-5.5 via openai-codex Default image generation: gpt-image-2-high via openai-codex In practice, Hermes uses our ChatGPT/Codex OAuth session to call the Codex Responses API, forces the image_generation tool, requests the actual image model gpt-image-2, and sets quality=high through the gpt-image-2-high tier alias.

Just updated my hermes-agent to use gpt-image-2-high instead of the default. Now I'm able to render the best quality images directly from hermes-agent, using tui, telegram, etc.

@oliverhenry I know there’s extra usage atm which will go some way to this - but I still haven’t hit limits on GPT, but it was daily on Anthropic even on the £200 plan

I’m so happy with GPT 5.5 that I probably won’t go back to Anthropic if they release a new model. Larry with the codex subscription is peak computing.

@najmuzzaman @gregisenberg

open sourced WUPHF: an AI office that builds its own team and collaborates to run your workflow end-to-end, 24x7. each agent backed by its own knowledge graph (choose @garrytan's GBrain or @nexdotai) . 7x less tokens than Paperclip. benchmark in repo. github.com/nex-crm/wuphf

@najmuzzaman I’m thinking of going full circle and using WUPHF to launch a paper company.

surreal. WUPHF broke ShowHN and #1 within an hour, on a weekend. news.ycombinator.com/item?id=478998…

@morganlinton Which is the local model you’ve had the most success with so far? I think I’m running similar specs to you (M2 Max Studio 32gb) but haven’t experienced much with local

So yeah. DeepSeek's announcement yesterday was cool and all, but common - you need crazy hardware to run it. I still think Qwen takes the cake this week in doing something truly novel that could really accelerate the foundational shift that is happening in the local llm space. What DeepSeek did is only for super rich local llm ppl and companies that can afford to run something like this. I'm over here still trying to decided if I can spend $3k on a basic PC with a 3090 🫠

Codex via Hermes on AI: "The actual refresh pipeline should be deterministic software, not a hallucination engine with a god complex" Brutal self-takedown

Hermes / Codex "Here's the 12-week sprint backlog" Looks like 12 hours to me.

@johnennis This is similar to the Anthropic "we've achieved a lot for today! Let's revisit tomorrow"

@RoadworkUK @alanisnking

Extraordinary scenes! Rich Energy exists, in the flesh, in a Twickenham pub. (didn't check the expiry date)

In my head God’s Son is still one of @nas’s newer albums. But I saw today that it came out in 2002 and now I feel 100 years old.

@oliverhenry Grok 4.1 Fast

@mark_tofts What model was that?

xAI is a good alternative to anthropic for openclaw to be honest. It’s much cheaper and can manage long tasks. However, every now and then it tells me it does something when it hasn’t. Has anyone else experienced this? Apart from the odd blip, I’m very impressed.