
AI • Crypto • Agents • Models AMB @ZARGATES
Joined September 2021
- Tweets135K
- Following11.3K
- Followers34.8K
- Likes155K
Running an AI assistant locally gives you more control over your data, memory and logins.
Replit is exploring that idea with its new open-source project, Amuse.
Released under Apache-2.0, Amuse can research the web, control a live browser, manage reminders and build small apps.
Replit says its Missions agents built it in 3 days.
What makes it useful?
- A live browser you can watch, control or stop at any time.
- Memory stored as plain files you can read and edit.
- Approval cards before it sends, submits, pays or schedules anything.
Saved logins stay in an encrypted vault, while the model only sees stand-in tokens.
A few limitations:
Amuse is OpenAI-only, with "gpt-6.2-sol" as the default model. It is single-user with no login, so never expose it to a network.
The repository is brand new, with its first commit on November 19, so expect some rough edges.
GitHub: github.com/replit/amuse
What if your AI voice agent could stay connected while nobody is talking, without paying the full active rate?
OpenAI has introduced a cheaper standby mode for live voice sessions.
GPT-Live-1.1 is now available in the API, and idle time on a live call now costs $0.02/min instead of $0.05/min.
What it does:
- Better pronunciation, steadier accents and more accurate transcription across languages than GPT-Live-1.
- Full-duplex: it can listen while it speaks.
- Standby on WebRTC sessions: $0.02/min while paused vs $0.05/min active. It also works with GPT-Live-1.
- Keeps the same connection, session ID and history when it wakes up on speech.
How to start:
1. Set the model to "gpt-live-1.1" in your "POST /v1/live/sessions" config.
2. Add "session.wake_policy.type: "speech"" and "session.idle_timeout_s" (5 to 60 seconds).
3. Listen for "session.paused" and "session.resumed", then test a call with long silences and compare the bill.
Heads-up: The standby rate only applies to WebRTC sessions. Keep the mic track active during standby so it can detect speech and wake up. The first reply after a pause can also be slower.
Backend model and tool calls are billed separately.
Documentation:
developers.openai.com/api/do…
developers.openai.com/api/do…
32
34
607
Genie 4 is now live in Google AI Studio.
Explore interactive worlds, change things on the fly and experience native audio.
Try it here:
aistudio.google.com/live?mod…
What if you could cut AI API costs by up to 40% just by letting a task finish later?
OpenAI has introduced discounted completion windows for background tasks in the Responses API.
Instead of getting an instant response, you can choose a 2-hour or 8-hour window and pay less for tokens.
What it offers:
- 2-hour window: 30% below Standard token rates.
- 8-hour window: 40% below Standard rates.
- Supports GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra and GPT-6 Luna (public beta).
- Multi-step agent runs also qualify, including tool-call follow-ups.
How to start:
1. Choose a task that does not need an instant response, such as overnight summaries, weekly research digests or bulk tagging.
2. In "POST /v1/responses", set "background: true", "service_tier: "default"" and "background_options.completion_window: "8h"". Keep streaming off.
3. Save the response ID, then collect the result using a "response.completed" webhook or polling.
4. For follow-ups, reuse the same model, window and a "continuation_id".
Heads-up: The completion window is a target, not a guarantee.
Background mode temporarily stores response data, even with "store: false" or Zero Data Retention. Check the data retention implications before sending sensitive files.
The new GPT-6.2 models do not offer Flex processing, so this is the alternative discount route.
Documentation: developers.openai.com/api/do…
26
32
603
Running AI agents locally does not always mean building everything from scratch. These five open-source tools offer different ways to do it.
All of them can run on your own PC:
1. nanobot: lightweight personal AI agent with WebUI and chat apps.
2. goose: desktop AI agent for research, writing and automation.
3. Hermes Agent: personal agent that learns and remembers.
4. OpenClaw: self-hosted agent you can reach from Telegram, WhatsApp and other apps.
5. Agent Zero: general-purpose agent with its own desktop and browser.
You can explore, run and manage your own AI agents without relying entirely on paid cloud services.
Links:
nanobot github.com/HKUDS/nanobot
goose github.com/block/goose
Hermes Agent github.com/NousResearch/herm…
OpenClaw github.com/openclaw/openclaw
Agent Zero github.com/agent0ai/agent-ze…
3
122
A GitHub issue can become a pull request with AI agents handling the work from planning to code review.
That is the idea behind Mastra Factory.
The Mastra team says its agents opened and merged 238 pull requests last week. Now the tool behind that is open source.
Mastra Factory (Apache 2.0) is a "software factory": connect a GitHub repo, and agents turn issues into plans, code and reviewed pull requests.
What it does:
- Pulls work from GitHub issues, Linear or Jira, and you can start sessions from Slack.
- Moves each task through investigate, plan, implement and review, and can wait for your approval at each step.
- Opens real PRs and runs a separate review session on each one.
- Runs on Mastra's platform, or self-hosted with your own auth, Postgres and sandboxes.
How to start:
1. In a separate folder (not your app repo), run: "npm create factory@latest my-factory"
2. Add a credential encryption key to ".env" ("openssl rand -base64 32") and back it up.
3. Run "npm run dev", then connect your GitHub repo and a model provider.
4. Start small with a docs-only issue, and check the plan, the diff and the PR checks before you merge.
Heads-up: The default setup signs you in to Mastra's platform. For full self-hosting, add "--no-platform" and bring Postgres with pgvector. Keep a human review before every merge.
27
26
501
Cloudflare now has one CLI for its whole public API, built with coding agents in mind.
cf just reached its 1.0 release candidate. It covers DNS, storage, security and Workers. It is also open source (Apache 2.0).
What it does:
- Commands come from the Cloudflare API. Agents can use "cf cli search" and "cf tools" to find commands and every flag they accept.
- Every command takes "--dry-run" to preview what it would do, with sensitive fields redacted.
- "cf observability tail" streams live logs as JSON.
- "cf docs search" finds Cloudflare docs from the terminal.
- Destructive commands refuse to run outside a terminal unless you pass "--force".
How to start:
1. Install: "npm install --global cf@latest"
2. Sign in with "cf auth login", then check "cf whoami".
3. Already on Wrangler? Run "cf migrate" in your project.
4. Tell your agent to run every change with "--dry-run" first, then run it for real.
Heads-up: It is a release candidate. A few less-used API commands may still be renamed before 1.0, so do not hard-code those in scripts yet.
developers.cloudflare.com/ch…
github.com/cloudflare/cf
COINBASE IS NOW AVAILABLE IN CHATGPT. READ THE APPROVAL RULES BEFORE CONNECTING.
Here is what you need to know:
- Trade crypto, eligible US stocks and dated futures directly from chat. Currently available to US Coinbase users only.
- Before every new order, edit, cancellation or USD-USDC conversion, you must approve it through a Coinbase confirmation card.
- The custom Coinbase MCP connection works differently. It does not require approval cards before every trade, making it suitable for automation.
- For safety, use a separate portfolio with only the funds you can afford to lose. Set Agent Controls such as trade limits, 24-hour limits and loss limits.
- You can revoke access anytime from Coinbase Account > Security > Connections.
These open-source tools can replace paid AI subscriptions.
All of them run free on your own PC:
1. LocalAI: run chat, image, speech and video models locally. It recently got a major 5.0 update.
2. Jan: clean ChatGPT-style app that works fully offline.
3. Linger: personal AI agent that runs on your own computer.
4. Llama for Windows: one-click access to local models from the llama.cpp team.
5. OpenGCUA: open-source framework for computer-use agents.
No subscription. No cloud.
32
35
744
Links:
LocalAI github.com/mudler/LocalAI
Jan github.com/janhq/jan
Linger github.com/simplexai-labs/li…
Llama for Windows github.com/ggml-org/llama.cp…
OpenGCUA github.com/Tongyi-MAI/OpenGC…
Check each repo for setup and hardware needs.
2
85
Pula (Earthshot 2026 Finalist) is joining NVIDIA Inception to use Earth-2 for climate modeling.
Faster, higher-resolution drought and extreme weather forecasts to help protect smallholder farmers, with a goal of 100 million by 2030.
This is AI climate tech moving from research into real insurance and farmer protection tools.
Pula aims to protect 100 million farmers from climate shocks by 2030. Climate modeling powered by NVIDIA Earth-2 will help it pursue that goal. 🌱
At The Earthshot Prize’s Impact Assembly in Mumbai, 2026 Finalist Pula announced it is joining NVIDIA Inception and will use Earth-2 to better understand how drought and extreme weather will affect the smallholder farmers it serves.
Faster, more detailed forecasts, alongside new support from AWS, will help Pula scale protection for farmers’ families and livelihoods.
Read the announcement: earthshotprize.org/news/the-…
#EarthshotMumbai
11
18
435
What if your coding agent could turn a folder of logos and screenshots into a 4K promo video?
That is what motlab does.
motlab is an open-source video tool (MIT) that creates videos using code instead of AI video models.
You provide images and a short description. Your coding agent builds the scenes, and you can review each frame and request changes.
What it does
- Creates scenes using your logos, screenshots and images
- Lets you leave feedback directly on video frames
- Keeps different versions so you can compare them
- Exports videos in 720p, 1080p, 1440p and 4K
- Adds music and voice to the final video
- Creates reusable assets like icons, characters and 3D objects
How to start
1. Install Node 24+, Chrome and ffmpeg
2. Run "npm install -g @motlab/cli && npx @motlab/cli setup"
3. For Codex, Cursor, Gemini CLI and other agents, use "--skill agents"
4. Give your agent a short video brief and the folder containing your images
Heads-up: motlab is still new (v0.2.2). The author's claim that it is 2–4x faster than Remotion comes from a single-scene test on one machine.
github.com/staryone/motlab
Tencent has open-sourced the weights of Hy4 Flash, a 220B parameter AI model with only 12B active per token.
It can understand text and images, handle up to 1M tokens of context, and work on tasks like research, document creation, slides and computer use.
What makes it interesting
- Three versions are available: Hy4-flash (instruct), Hy4-flash-FP8 and Hy4-flash-base
- On Tencent's internal WorkIndex test, it scored 42.1 compared to 39.7 for Hy4 preview, at around 1/13 of the cost per task (Tencent's numbers)
- In a blind test with 85 users, it scored 2.79/4 in satisfaction, compared to 2.81/4 for Claude Haiku 5.5
- Supports vLLM and SGLang, with a GGUF build also available for testing
How to start
1. Open the model card and choose the FP8 version if you plan to run it
2. Use the official vLLM image: "vllm/vllm-openai:hy4-flash". The example setup uses 8 GPUs
3. Connect an OpenAI-compatible client using the model name "hy4-flash"
4. Test it on your actual tasks. Tencent recommends the larger Hy4 model for long software engineering work
Heads-up: The model uses Tencent Hy License, which is based on Apache 2.0 but has additional conditions. Model-as-a-service companies earning over $50M annually, or products with over 100M monthly users, need a separate license.
Model: huggingface.co/tencent/Hy4-f…
GitHub: github.com/Tencent-Hunyuan/H…
You do not always need paid AI subscriptions. These 5 open-source tools let you run AI directly on your own PC.
Here is what each one does:
1. VoiceStudio: Clone voices, dub videos and convert text to speech locally. (110K+ GitHub stars)
2. whisper.cpp: Convert speech into text offline, even on a regular laptop.
3. Jan: Use a ChatGPT-style AI chat app privately, without an internet connection.
4. Ollama: Run AI models like Qwen, DeepSeek and Llama locally with one command.
5. ComfyUI: Generate images and videos on your own PC without paying for cloud credits.
No monthly subscription or API fees for local use.
You just need the right hardware.
Links:
VoiceStudio github.com/debpalash/VoiceSt…
whisper.cpp github.com/ggerganov/whisper…
Jan github.com/janhq/jan
Ollama github.com/ollama/ollama
ComfyUI github.com/comfyanonymous/Co…
Check each repo for hardware requirements.
4
90
Your coding agent just got 7 new design skills.
craft is an open-source toolkit (MIT) for Claude Code, Codex, Cursor, OpenCode and other coding agents.
Here is what the skills do:
- craft-setup: Asks about your design preferences once and creates a DESIGN(dot)md that future sessions can follow.
- craft-states: Lists all screen states, including empty, loading, error and offline, before designing.
- craft-critique: Renders the design and ranks issues by importance.
It also works with Paper, pen(dot)dev Figma and plain HTML.
Heads-up: The Figma adapter has not been tested in real runs yet.
Try it: github.com/diego3g/craft
35
38
631
You do not always need paid AI subscriptions to use powerful AI tools.
These 5 open-source tools can run locally on your own PC, without relying on paid cloud services.
1. MiniMind: Train your own small language model in around 2 hours (66K GitHub stars).
2. Tiny-Voca: Convert text into speech using your laptop CPU.
3. Mimvox: Clone voices and dub audio offline in 60+ languages.
4. Linger: Run a local AI agent without needing an API key.
5. FreeVideo: Generate AI videos locally with just 8GB of VRAM (3.2K GitHub stars).
No monthly subscription. No credits.
Just local AI tools running on your own hardware.
Links:
minimind github.com/jingyaogong/minim…
tiny-voca github.com/n4ze3m/tiny-voca
mimvox github.com/tarpetos/mimvox
Linger github.com/simplexai-labs/li…
FreeVideo github.com/FlashML-org/FreeV…
Check each README for hardware requirements.
2
133








