Google Document AI

Google Document AI AI

Google Document AI AI

Google Document AI AI

Google Document AI AI

Google Document AI AI

Google Document AI AI

Google Document AI AI

Google Document AI AI

Google Document AI AI

Google Document AI AI

Built an operating system for my life managed by Claude

With the OS I can ask Claude "what did I spend on coffee in 2022" and get back "$847 across 213 transactions, mostly Blue Bottle and Verve". Name me one expense tracking SaaS that can do that! And its not just my financials, my OS contains everything about my life in one place so Claude can reason about it. I've been building this incrementally for a few months. Its just a small web app on Cloudflare that holds my entire life: * bank transactions from Chase, Apple Card, BoA business * every receipt out of Gmail going back to 2019 * legal filings for my green card (I-140 still pending lol), C-corp and LLC docs, contractor agreements * calendar with linked people and locations * notes and reminders the agent dumps in over time * health tracking (exercise stats, nutrition, sleep and other biometrics linked to my Aura ring) Whenever I have to upload something, I just throw it into Claude and tell it to do it. For refreshing financial connections to BoA for example, I click refresh once a week, complete the 2FA and it syncs up. any Claude surface (claude.ai, Claude Code, Desktop) talks to my REST API. one long-lived auth token, one line in CLAUDE.md saying "before answering anything personal, query <my operating system's URL>." Its f\*\*cking great for financial, taxes and legal stuff. Now that everything is in one place, I just ask Claude stuff like "status of my green card, next deadline?", "which LLC I used to sign the office lease?". I even have a dashboard showing a grid of all my subscriptions (Claude made it from reading my BoA account transaction history), and a giant money tracker at the top that shows my monthly income/expenses. This replaced a bunch of SaaS's I was using for expense tracking and whatnot. E.g. Claude blows RocketMoney's system out of the water - I can actually chat about my financials and get intelligent analysis. Its also nice not going Notion or Google Drive folders or a gazillion other places to find all the right files. I just ask Claude to add it to my OS instead. if there's interest I'll write up the full setup, it's a small backend plus loads and loads of integrations I've iterated on over months.

MarkdownAI v2.0, its a workflow engine, not a template parser

MarkdownAI is a workflow and runbook engine for AI. Yes, it’s also a templating language, but that’s the least interesting thing about it. The power is the MCP server. Claude never sees a stale file again. Every document resolves live, every time. Simple example: your frontmatter. Status fields, version numbers, last-updated dates, owner, the stuff that’s wrong within a week of writing it. With MarkdownAI, frontmatter becomes live. Claude doesn’t read “status: in-progress” from three weeks ago. It reads the actual current state, fetched at render time. No staleness. No verification step. No “is this still true?” check that costs a tool call. That same idea scales to everything in the document, DB record counts, branch names, env values, test results, file trees. Anything that goes stale becomes live. **The grunt work problem** Before Claude does anything useful, it does housekeeping. Verify the branch. Check CI. Query the DB. Hit the health endpoint. Read env vars. Confirm the image exists. Check migrations. That’s a real pre-deployment runbook, and Claude is doing all of it, one tool call at a time. Each check is roughly 2 seconds of dead time plus a context interruption where Claude has to re-orient. 15 checks = 30 seconds of grunt work and 15 quality hits before the first useful output. Splitting your runbook into multiple files doesn’t help, Claude still stops to Read. And every Read loads the whole file. If CLAUDE.md is 800 lines and Claude needs 40, it pays for all 800. MarkdownAI moves this out of the prompt entirely. Directives resolve in the MCP server before Claude sees anything. Need one section of a file? Inject just that section. Claude enters every turn with facts, not tasks. **@phase** A flat workflow loads every step into context upfront. Step 12’s instructions sit there during step 2, eating room Claude could use for actual work. \`@phase\` serves one step at a time. Claude sees what it needs for this step, nothing else. Session state persists across phases. A 20-phase runbook uses a fraction of the context a flat document would. \`\`\` >!@phase pre-flight!< >!@on-complete deploy /!< >!@phase-end!< >!@phase deploy!< >!@on-complete verify /!< >!@phase-end!< \`\`\` **Compaction stops being a failure mode** Long session hits compaction. Claude decides what to keep and what to discard. It keeps what it thinks is important, which is rarely the same as what actually matters. After compaction, Claude is working from a lossy reconstruction of your system state, with confidence. With phases, that problem is gone. The next phase re-injects everything live. Not a summary. Not what Claude remembered. Real env values, real DB results, real state, real constraints. Claude can’t misremember a \`@constraint\` because it was never stored in memory, it’s re-fetched every phase. Compaction becomes a non-event. 996 tests. Full docs at [https://markdownai.dev](https://markdownai.dev)

Transform any document or url into a video inside Claude with this MCP

Connect Claude to the Ozor video API. Claude can generate animated videos from a prompt, turn a PDF/DOCX/PPTX/URL into a multi scene video with voiceover, poll long running jobs, export MP4 at 720p/1080p/4K, and return a share link and embed iframe. Tools: generate\_video, analyze\_document, generate\_from\_plan, export\_video, wait\_for\_export, get\_embed\_code, list\_videos, send\_message. \*\*How Claude Code built it\*\* I gave Claude Code the Ozor REST spec. It scaffolded the MCP server in TypeScript, generated tool schemas from the spec, wrote the handlers and the async polling layer. Most of the work was iterating on tool descriptions so another Claude instance picks the right tool. Roughly 3 days of work that would have taken me 2 weeks by hand. \*\*Install (Claude Desktop)\*\* Settings > Connectors > Add custom connector. URL: [https://mcp.ozor.ai/mcp](https://mcp.ozor.ai/mcp) \*\*Try it\*\* Ask Claude: "Generate a 16:9 video for my SaaS launch, 3 scenes, problem, product reveal, CTA. Export as 1080p." \*\*Free tier:\*\* 10 credits per month, no credit card, no watermark. Sign up at ozor.ai. Happy to answer questions about building production MCPs with Claude Code.

I had my agent use autoresearch over 8 iterations to improve my CLAUDE.md, measuring each version against tasks from real PRs. The best one still regressed on a holdout.

I have a confession: I vibe-coded my [`CLAUDE.md`](http://CLAUDE.md), and I'm pretty sure it's slop. I needed to make it better. Naturally, I asked Codex to do it. (I know this is a Claude sub, Claude could have done it as well!) The difference: this time, Codex used a benchmark on my repo to measure each change, and optimized [`CLAUDE.md`](http://CLAUDE.md) against the data, instead of on pure vibes. # Why We Should Take [CLAUDE.md](http://CLAUDE.md) Seriously Saying "`AGENTS.md` is important" is, at this point, a cliche. At risk of beating a dead horse, I'll say it again. Someone adds a rule that sounds smart, senior, and reasonable, commits it, and hopes the agent behaves better. But [`AGENTS.md`](http://AGENTS.md), [`CLAUDE.md`](http://CLAUDE.md), and shared skills are not normal docs. They are part of the runtime behavior of your coding system. **The shift is to start treating** [`CLAUDE.md`](http://CLAUDE.md) **like a tunable part of the harness:** holding everything else the same, how does agent behavior differ when I change `AGENTS.md`? That's what I measured. # The Results After eight candidate runs, one version looked useful on a five-task training slice. It fixed the task the baseline missed, improved footprint risk, and moved several craft scores up. Then I ran it on a clean ten-task holdout. The candidate regressed. Not catastrophically, but enough that blindly shipping would have been wrong. Footprint widened, tokens climbed, tool calls climbed, and code-review correctness fell, all while tests held even. *Caveat: one repo (mine), n=10 on the holdout. This is directional, not statistically significant.* *For this post, "equivalent" means the patch matched the intent of the merged human PR; "code-review pass" means an AI reviewer judged it acceptable; craft/discipline is a 0-4 maintainability/style rubric; footprint risk is how much extra code the agent touched relative to the human patch.* The pattern is the agent doing more work for mixed outcomes - better on local craft (clearer names, coherent implementations), worse on boundary judgment (scope, minimality, robustness). Tokens and tool calls confirm it: the candidate was spending more to get there, not less. "Better instructions make the agent cheaper" did not hold on the holdout. [best iteration and holdout vs baseline](https://preview.redd.it/9tgyk8gihq3h1.png?width=1854&format=png&auto=webp&s=8b5a5e42ba79ac554b143c92d091f0e4d8e25417) # Methodology The setup was Codex with `gpt-5.5`, medium reasoning, on real historical Stet tasks (dogfooding). Stet scored tests, strict publishability, equivalence, code review, footprint, total input/output tokens, duration, and craft/discipline rubrics like simplicity, coherence, robustness, instruction adherence, scope discipline, and diff minimality. The grader was `gpt-5.4`. 8 iterations on an n=5 sample set, and a n=10 task holdout. **I know sample size is small - the goal of this was to get directional analysis, and prove the methodology** Codex was set with a simple `/goal`: iterate [`AGENTS.md`](http://AGENTS.md) to improve performance on the benchmark. # Process The first round of iteration showed something I wish more people internalized: **plausible instructions are not necessarily good interventions.** Codex first tried a broad router rule: identify the work type, state a hypothesis before editing, read the right docs, and treat scope as part of correctness. It sounded good but exposed a failure mode: the agent could interpret "small scope" as permission to miss named obligations. The next candidate added an "obligation ledger". Before editing, the agent had to identify the named behavior, compatibility constraints, docs, tests, and non-goals. Before reporting back, it had to mark each as met, missed, or not checked. Here is the actual diff shape. First, the best candidate from the first loop replaced one generic "read the docs" rule with routing, hypothesis, obligation, scope, and evidence rules: - For nontrivial work, read the matching `agent_docs/` file first for current operational commands and conventions. + Route before acting: identify whether the work is implementation, eval/report interpretation, dataset/pipeline, Linear/Symphony, release, frontend, or GTM; then read the matching `agent_docs/` or skill file before changing behavior. + For nontrivial changes, state the smallest testable hypothesis before editing. After validation, report whether the evidence confirmed, refuted, or only weakly supported it. ... *Full details in blog post* [*https://www.stet.sh/blog/how-i-used-codex-to-improve-its-own-agents-md*](https://www.stet.sh/blog/how-i-used-codex-to-improve-its-own-agents-md) That obligation-ledger candidate was the first useful signal. Code review improved by `+0.75`, correctness by `+0.60`, maintainability by `+1.00`, simplicity by `+0.64`, coherence by `+0.60`, and scope discipline by `+0.36`. Tests stayed flat at 5/5. But

Can I leverage Claude in this way?

I’m new to Claude, have only ever used ChatGPT as a chatbot and DIY tasks with networking/troubleshooting things outside of my skill set. I was introduced to Claude and vibecoding by a friend. Now I run a business and I’m trying to leverage Claude for tasks through cowork and code using Pro/max. Can I use chat to understand the logic of a layouting software that gives me different layouts using dimensional inputs and a logic/maths to generate a visual and mathematical output? Essentially dimensions for a flat carton/box and it gives me the various multi-up flat layout options ? It would be software that I’d build out and I guess host on the web for a couple of users (minimal data hosting Would using chat the understand the task and then generate input for code to then code that out be the best approach? The other thing I’d like to do is automate some tasks. Would using cowork be the way? Can it reliably do this? I’d also like to automate extrapolating of client/purchase data from pdfs and sheets (google workspace) to then compile and organize on a daily and weekly basis. Parsing would also require some rules and understanding of different layout of documents from other organizations to pull relevant data. I would give the constraints and tweak and fine tune these tasks but not sure how to approach setting this up. Again do I use chat to understand the task then generate the prompt in cowork? Any particular attention to the folder structure needed? I’m sorry I don’t have any experience with cli or programming so it’s a bit confusing but I can generally pick things up well. Would skills be helpful in any of this? Can sonnet scrape data and compile, categorize and organize into sheets reliably? Atleast where the data to scrape is presented in different ways by each org and document? Sorry if this is too nooby. I only ask because I don’t want to go down this rabbit hole only to realize I’m in over my head and it won’t be reliable enough for day to day business function and other ways I’d like to leverage it as a tool to develop more. Atleast for someone like myself

Found a workaround or i didn't know you could do this

So whenever u generate a document with claude free plan(i don't have money 😔) so let's say u generated some notes from it always tell it to generate a doc based on that, most of the time if it's a small doc you'll get your output but if it's a pretty big doc you'll be soon out of free credits for that day. So what claude does is I don't know if it's intentional or not just before the command runs so that the js file is executed and you get your doc they cut you out. So what you can do is download the js file, open cmd in that directory, 1. Install node.js and npm 2. Run command - npm install docx 3. Run command - node file\_name.js The file will be as a word doc in that some directory seems pretty useful to me Note - at the end of that js file there will be something like ("some/directories/file-name.docx",buffer) change it to ("file-name.docx",buffer) it's some linux technicality which let's the AI download the doc file for you and provide it from their end.

Which provider fits best for my needs?

Hi everyone, I’m looking to get more into experimenting with AI and considering a paid subscription, but I’m a bit unsure which direction makes the most sense for my use case. My main goals: \-Writing a technical book in the field of taxation \-Preparing presentations and structured content \-Learning and experimenting with programming \-Building automation workflows (e.g. n8n) \-Running or experimenting with tools like Hermes / OpenClaw (I know Claude doesn’t work everywhere there) \-Testing new AI features (e.g. Claude artifacts, coding tools, agents, etc.) From what I’ve read recently, opinions are all over the place: Some say ChatGPT (with Codex-style tools) is strongest for coding + general use Others argue Claude is better for writing and reasoning-heavy tasks Gemini seems strong for long context and Google integration And then there’s the API route (DeepSeek looks extremely cheap right now and seems attractive for experimentation) So I’m trying to figure out what actually makes sense in practice. Would you recommend: A ChatGPT subscription Claude Pro Gemini Advanced Or skipping subscriptions and going API-first with models like DeepSeek / others? Would really appreciate real-world experiences—especially from people doing a mix of writing + coding + automation rather than just one narrow use case. Thanks! (Ai generated as englisch is not my mother language)

11 months solo. dropped 3 tools after claude including the notion alternative i was paying for.

what i cancelled this year: * a $39/mo notion alternative i was using as a "smart" workspace. claude in projects does 80% of what i was paying for. * a $79/mo "ai assistant" platform. didnt do anything claude couldnt. * a $49/mo ai document generator that produced templates that looked like every other landing page. what i kept paying for: * claude max ($200/mo). carries half the value of my whole stack. * gamma ($20/mo) for client deck deliverables. * notion ($10/mo). yes still notion. claude is the brain, notion is the filing cabinet. savings $167/mo. 11 months solo, revenue this year \~$112k working \~32 hrs/week. the unlock isnt any single claude feature. its that the SaaS layer between me and the model is mostly value extraction. some real value exists. most is markup on a thin prompt. what have you cancelled this quarter that you do not miss.

non coder, 6 months on max, my favorite use of claude is honestly the boring ai content generator stuff

I am not a developer. I run a small training and content business. Have been on Max since february. Everyone in this sub talks about agents and skills and Claude Code. I do not use most of that. My favorite use of claude is the most boring thing imaginable. It is the ai content generator for the work I find tedious. Specifically: 80% of my client emails. I write a 2 sentence brief, claude drafts, I edit, I send. Training material first drafts. I dump notes, claude builds an outline with timing. The 800 word weekly update to my retainer clients. Claude drafts, I revise. Slack messages I have rewritten 3 times in my head. I tell claude what I am trying to say and what I am worried about. It writes the version I would have if I had 20 more minutes. I will not use claude for: my newsletter intro, instagram captions, anything that needs to sound like me to people who know me, anything emotional. I also use google docs ai for the surface polish on long documents claude has drafted. They are different tools doing different work. What I have not yet figured out: a use case where claude is actually replacing my judgment rather than my typing. I think that's correct? The judgment is the work. Curious what other non coders are doing past 6 months.

Solo bookkeeper. Claude paired with google docs ai is the only ai tool for writing client emails that hasn't burned me.

16 clients, mostly e-commerce and small services. 6 years in practice. What burned me with other AI tools: * One drafted client emails that sounded too smooth. Clients asked if I was sick. * Another categorized transactions with 60% accuracy which means I checked 100%. * A third "summarized" my client meetings and got tax facts wrong in the summary. What works with Claude: * I write a brief, claude drafts, I edit in google docs with google docs ai for surface polish only. * The voice stays mine because I edit every line. * I never let claude write a number that goes anywhere a client will read it. 5-6 hours a week saved on client correspondence. Same client relationships. Better turnaround. The ai tool for writing client emails is finally just a faster version of me, not a different version of me. That distinction matters for client trust.

How to create an AI version of yourself using your reddit history

I hate the way AI talks back to me. Its so proper, so robotic, every response feels like a help article. I wanted something that actually knew who i am, my beliefs, my history, what shaped me, the positions i hold and why. Not a generic assistant that treats every question like it came from nobody. So i got to thinking, who better to talk to than myself? So i built it over a weekend. Heres what I did and how you can do it too. **Step 1: Export your Reddit data** Go to [reddit.com](http://reddit.com) and click your profile icon in the top right, then hit Settings. Scroll down to the bottom of the page and youll see a section called "Data Request." Click "Request Data Export" and Reddit will email you a download link within a few hours, sometimes longer depending on how much history you have. The zip file will contain your posts and comments going back to when you created your account. Mine was about 21,000 comments over two years. Once you have it, open the CSVs in excel or just upload them directly into Claude and ask it to help you make sense of the structure. The raw data is ugly but everything is there, the text of every comment, the subreddit it was posted in, the date, all of it. One thing worth knowing: you can go way deeper than just Reddit. I looked into Google Takeout while i was doing this and it was honestly a little scary how much data they have on you. If you want to go deeper Google Takeout is wild, i didnt realize how much data they actually have on you until i went through it. Search history, location history, YouTube, Gmail, its all there and its all exportable. I thought about pulling my SMS history too but that felt wrong, those conversations are with real people who didnt agree to any of this so i left it alone. Reddit was enough for me and honestly if youve been on here for years and actually say what you think in the comments, you probably have more to work with than you realize. **Step 2: Build the personality document and this is where the real work is** Dont just tell the AI "write like me." That gives you nothing. You need an actual document, a living reference file the AI reads every single conversation. Mine is a markdown file sitting in a Claude Project so it loads automatically every time. Start by uploading your Reddit export and asking Claude to interview you. Literally tell it: "Read my comment history and ask me questions about anything it cant determine on its own." Let it go deep. Mine asked about my beliefs, my family, my history, my faults, things that happened to me, why i hold the positions i hold. You answer honestly, including the uncomfortable stuff, and then after the session you tell it to compile everything into a structured document. Then you iterate. Every time it gets something wrong you correct it and add it to the doc. Two weeks in and its already a completely different document than what came out of that first session. Heres what the document actually needs to cover: **Who you actually are.** Not the resume version. The real version. Your beliefs, your politics and why you hold them, your actual faults, your history, the things that shaped you. An AI that only knows your best self sounds fake because you sound fake when youre performing your best self. **Your actual positions on things.** Not just "im conservative" or "im liberal." The specific positions with the reasoning behind them. Mine has maybe 15 specific theological positions with the scriptural basis for each, because if the AI doesnt know why i believe what i believe it cant argue it like i would. **Your life context.** Family, relationships, the stuff that matters. Your context is constantly informing how you respond to things even when the topic isnt directly about your life. **Your faults and struggles.** This one people skip and its why their AI version sounds sanitized. Put in the real stuff. The AI needs to know the full person or it just sounds like your linkedin profile with apostrophes dropped. **Step 3: Set up the Claude Project correctly** Claude has a feature called Projects where you can upload files and write a persistent system prompt that loads every single conversation. Heres how mine is structured: The **project files** are the personality document and the Reddit exports. The personality doc is the source of truth for who you are. The Reddit exports are the raw data the AI can search when it needs to verify something or find a voice sample. The **project instructions** are where you govern behavior, not just describe personality. This is the part most people miss. Describing yourself isnt enough, you have to tell the AI how to behave. Mine has: Grammar rules shown as examples not descriptions. Side by side. Heres AI voice, heres my voice. Because "sound natural" is meaningless instruction. Showing it what natural actually looks like works. A banned vocabulary list. Words i never use. "Nuanced", "crucial", "delve", "it's worth noting", "at the end of the day",

How to build an AI of yourself using your reddit history

I hate the way AI talks back to me. Its so proper, so robotic, every response feels like a help article. I wanted something that actually knew who i am, my beliefs, my history, what shaped me, the positions i hold and why. Not a generic assistant that treats every question like it came from nobody. So i got to thinking, who better to talk to than myself? So i built it over a weekend. Heres what I did and how you can do it too. **Step 1: Export your Reddit data** Go to [reddit.com](http://reddit.com) and click your profile icon in the top right, then hit Settings. Scroll down to the bottom of the page and youll see a section called "Data Request." Click "Request Data Export" and Reddit will email you a download link within a few hours, sometimes longer depending on how much history you have. The zip file will contain your posts and comments going back to when you created your account. Mine was about 21,000 comments over two years. Once you have it, open the CSVs in excel or just upload them directly into Claude and ask it to help you make sense of the structure. The raw data is ugly but everything is there, the text of every comment, the subreddit it was posted in, the date, all of it. One thing worth knowing: you can go way deeper than just Reddit. I looked into Google Takeout while i was doing this and it was honestly a little scary how much data they have on you. If you want to go deeper Google Takeout is wild, i didnt realize how much data they actually have on you until i went through it. Search history, location history, YouTube, Gmail, its all there and its all exportable. I thought about pulling my SMS history too but that felt wrong, those conversations are with real people who didnt agree to any of this so i left it alone. Reddit was enough for me and honestly if youve been on here for years and actually say what you think in the comments, you probably have more to work with than you realize. **Step 2: Build the personality document and this is where the real work is** Dont just tell the AI "write like me." That gives you nothing. You need an actual document, a living reference file the AI reads every single conversation. Mine is a markdown file sitting in a Claude Project so it loads automatically every time. Start by uploading your Reddit export and asking Claude to interview you. Literally tell it: "Read my comment history and ask me questions about anything it cant determine on its own." Let it go deep. Mine asked about my beliefs, my family, my history, my faults, things that happened to me, why i hold the positions i hold. You answer honestly, including the uncomfortable stuff, and then after the session you tell it to compile everything into a structured document. Then you iterate. Every time it gets something wrong you correct it and add it to the doc. Two weeks in and its already a completely different document than what came out of that first session. Heres what the document actually needs to cover: **Who you actually are.** Not the resume version. The real version. Your beliefs, your politics and why you hold them, your actual faults, your history, the things that shaped you. An AI that only knows your best self sounds fake because you sound fake when youre performing your best self. **Your actual positions on things.** Not just "im conservative" or "im liberal." The specific positions with the reasoning behind them. Mine has maybe 15 specific theological positions with the scriptural basis for each, because if the AI doesnt know why i believe what i believe it cant argue it like i would. **Your life context.** Family, relationships, the stuff that matters. Your context is constantly informing how you respond to things even when the topic isnt directly about your life. **Your faults and struggles.** This one people skip and its why their AI version sounds sanitized. Put in the real stuff. The AI needs to know the full person or it just sounds like your linkedin profile with apostrophes dropped. **Step 3: Set up the Claude Project correctly** Claude has a feature called Projects where you can upload files and write a persistent system prompt that loads every single conversation. Heres how mine is structured: The **project files** are the personality document and the Reddit exports. The personality doc is the source of truth for who you are. The Reddit exports are the raw data the AI can search when it needs to verify something or find a voice sample. The **project instructions** are where you govern behavior, not just describe personality. This is the part most people miss. Describing yourself isnt enough, you have to tell the AI how to behave. Mine has: Grammar rules shown as examples not descriptions. Side by side. Heres AI voice, heres my voice. Because "sound natural" is meaningless instruction. Showing it what natural actually looks like works. A banned vocabulary list. Words i never use. "Nuanced", "crucial", "delve", "it's worth noting", "at the end of the day",

PAID Gemini vs FREE ChatGPT

I recently subscribed to Google One Ai Pro and recieved Gemini Plus Plan... I've been using it for some days, and the difference between Gemini and ChatGPT is enormous... i feel like talking to an Ai model from 2022. I asked them both to generate an image using the EXACT same prompt, here are the results... The prompt: "Generate a creepy midnight image in an abandoned road and there is a scary woman with white - blue gown standing next to the road. Make the quality unremarkably iPhone-ish, slight motion blur, grainy quality as if it was taken in dark. The picture is taken from a car in motion, from it's window on the front right seat." Models used: Gemini 3.1 Pro GPT-4o (afaik this is the model used in image gen in the free ChatGPT version atm) Edit: Added the models used. https://preview.redd.it/bbou5fdr0p3h1.png?width=1340&format=png&auto=webp&s=46ff98af1e386f8a4da6b1c304e13d1b319b95e5

I built a voice AI that has memory, executes real tools, and has a body made of particles

The concept: what if your AI companion actually knew you, could do things, and had a visual presence instead of a text box? Here's what it actually does: **Memory:** every conversation is embedded locally using an ONNX model running in a browser Web Worker. Semantic search surfaces relevant context from past sessions. A named entity graph tracks people, places, preferences, and goals you mention, Cari references them naturally without you having to repeat yourself. **Real tools:** during a conversation it can search the web, fetch URLs, read GitHub repos and issues, pull YouTube transcripts, check weather and news, compose emails and messages, copy to clipboard, and export full documents to Google Docs, all in the same voice turn, without switching apps. **Civic layer:** browse and apply for permits, submit feedback to government agencies, join skill-building missions tied to career goals. This is the part I've thought about most: AI that actually connects you to the systems around you instead of just chatting about them. **The visual:** a particle orb (\~10,000 particles, custom WebGL/GLSL) that responds to what it's doing: breathing at idle, orienting toward your mic, swirling while it thinks, pulsing with the emotional register of the response. When it describes something physical it morphs into a 3D mesh of it. The shape isn't decoration, it's the AI showing its work.

How I build my own zero cost Agent

I’ve spent the last few weeks obsessing over one goal: having a personal, self maintaining AI assistant that costs $0and can be controlled from my phone. It wasn't easy. I started with an AWS Ec2 with 50GB storage and t3.micro memory- minimal setup (using the free credits) and made Oracle Cloud instance ($300 free credits but just for a month so I used it for experimenting with local models) I was using Termius to SSH into everything from my phone At first I used OpenClaw. It was cool, but I spent more time fixing it than actually using it. I almost gave up until I saw a video about Hermes Agent. And i actually found Hermes while looking for how to fix an OpenClaw error on YouTube (thanks NetworkChuck 🙌🏽) He mentioned the exact same frustrations I was having, and that Hermes had been stable for a month. I didn't even finish the video before I pulled the repo. The best part? It had a "migrate from OpenClaw" feature. I was up and running in minutes. The hardest part is the rate limits. If you use cloud models especially for code, you hit a wall fast. My solution? The Fallback Chain. Initially I was using openrouter/owl-alpha (stealth models are usually flagships in testing, like big-pickle is deepseek v4) which has 1M context window and was on multiple rankings. Over time after I transitioned to Hermes, I wanted a bit more customization, while owl alpha was good at tasks, It’s nothing to talk about on roleplay, it just scrapes the surface of the character I set in SOUL md file. On my oracle instance I had been experimenting with local models (keep in mind, if you go local, you’ll be sacrificing speed but privacy. Ofc since the vms don’t have a gpu it would be slower, about 3-5 minutes for a simple response) The one I was most impressed with is Google’s Gemma-4-31b-it It played the role perfectly Buuut if you know Google, you’re familiar with their aggressive rate limiting. So I set up my agent to rotate through providers. I start with Gemma 4 for that perfect personality and roleplay via openrouter (add an ai studio api key in BYOK for longer usage). If that hits a limit, I’ve also set the same model via ollama cloud and using Google OAuth directly (basically Gemma 4 3 times lol) And if those all hit limits, it jumps to Qwen3-coder-next (Alibaba, 1M free tokens per model. There’s like 80), then Nova (AWS bedrock), DeepSeek v4 (Azure and Opencode Zen), and Claude Haiku (GitHub). If everything fails, I have Owl Alpha; which is an absolute beast, took almost 70M tokens before I got rate limited once, that too for a few hours. It lives in my Telegram and Discord. It manages my Spotify, handles my emails, and when I need real research done, I have it spawn three separate agents to work in parallel. It’s been 8 days and it hasn't broken once. If you're looking to get AI without spending a fortune, I highly recommend looking into this 🫪