BongoCat is a native animated desktop companion that reacts to keyboard, pointer, and gamepad input. It is built in C and C++ with SDL3 and OpenGL, supports Windows, macOS, and Linux, and processes keyboard and mouse input locally rather than recording or uploading it. The runtime uses platform-specific input listeners, an atomic input queue, an SDL3 main loop, model-parameter updates, and OpenGL composition for the pet window and overlays. It includes built-in model assets and supports external model sources, including Mver packages, Tauri sources, `.model3.json` files, and image patches. Live2D rendering uses the Cubism SDK when available; without it, the project builds a diagnostic backend. The source code and native runtime are licensed under AGPL-3.0-only, while the default model mode and bundled model assets have separate MIT licensing.
Compositor is a free, open-source image editor for macOS built around Photoshop-style compositing and post-processing workflows. It uses layers and folders with blend modes, opacity, masks, clipping masks, adjustment layers, non-destructive transforms, selections, painting and retouching tools, content-aware fill, filters, and multiple project tabs. It imports JPEG, PNG, HEIC, and TIFF files, and exports JPEG images; Photoshop-style keyboard shortcuts are supported throughout. The repository provides an Xcode project and documents macOS 26 and Xcode 26 requirements. It is licensed under the MIT License.
DeskcommCRM is an open-source, self-hosted CRM and AI sales operating system for businesses that sell through WhatsApp and other chat channels. It combines a multi-tenant CRM and sales pipeline with inbox, Kanban pipelines, contacts, team governance, audit logs, LGPD features, webhooks, automations, scheduling, and human takeover capabilities, and supports WhatsApp through WAHA QR connections or the official Meta Cloud API. Its AI agents use tenant-specific RAG, organizational memory, executable skills, intent routing, sentiment analysis, follow-ups, and audited handoff to human attendants. Agents can operate CRM records such as leads and funnel stages, while the CRM exposes an internal MCP interface for agent operations. Tenants can receive leads through public webhook endpoints and process event-driven QUANDO/SE/ENTÃO rules through an event-log queue drained by scheduled workers. The application is built with Next.js, TypeScript, and Supabase/Postgres.
herdr is a terminal multiplexer and persistent runtime for coding agents, developed as a Rust binary. It runs a background server that keeps agent terminals and sessions available across terminal disconnections, SSH reconnections, network loss, lid closure, and machine restarts; users can reattach from another terminal and use tmux-style keyboard controls or mouse interactions to split, move, and manage panes. Each pane is marked working, blocked, or idle, and agents can control herdr through its CLI and socket API to spawn panes, prompt other agents, and wait for an agent that is blocked. It hosts existing tools such as Claude Code, Codex, Cursor, OpenCode, and Grok without wrapping or replacing them, and supports plugins for extending panes and workflows. The project provides installation scripts and package-manager installation, documents remote use and session state, and is licensed under the Apache License 2.0.
Hindsight is an agent memory system developed by Vectorize.io for storing, retrieving, and synthesizing long-term memories for AI agents. It organizes memories into world facts, experiences, observations, and mental models within isolated memory banks, with support for multilingual data and optional per-bank scanning for secrets and personally identifiable information. Its retain operation uses an LLM to extract facts, entities, relationships, and temporal data, then normalizes them into searchable representations. Recall combines semantic vector search, BM25 keyword matching, entity/temporal/causal graph links, and time-range filtering, merging results with reciprocal-rank fusion and cross-encoder reranking. Reflect performs deeper analysis of stored memories to form connections and answer questions requiring more than retrieval. Mental models and knowledge pages provide continuously updated answers or documents derived from a bank's memories. Hindsight can run as a Docker service, Python package, embedded database, Kubernetes deployment, or hosted Hindsight Cloud service. It provides Python, Node.js, Go, CLI, and REST clients, an OpenAI-compatible LLM wrapper, integrations for agent frameworks and coding agents, and a built-in Model Context Protocol server. Self-hosted storage uses PostgreSQL with pgvector or Oracle AI Database 23ai; the project is MIT-licensed.
Magnitude is an open-source local inference server and CLI for running language models with AI agent harnesses. It profiles a machine's chip, memory, and bandwidth, recommends compatible models, downloads the selected models, tunes inference settings such as speculative decoding and concurrency, and runs models on demand. Models are loaded when needed and unloaded when idle or when memory is constrained; prompts, files, and models remain on the local machine, allowing offline operation after setup. It integrates with Pi, OpenCode, Hermes, OpenClaw, Codex, Claude Code, Oh My Pi, and Cline, and also provides a built-in harness. Magnitude supports macOS and Linux, with Windows support through WSL, and is licensed under Apache 2.0.
mini-AGI is an experimental continual-learning, byte-level language model that trains from scratch on a CUDA-capable GPU with at least 8 GB of VRAM. It reads 256 byte values directly rather than using a conventional tokenizer, processes data one chunk at a time, and uses the same forward path for training and generation. Its architecture combines dense prelude blocks with a recurrent block applied up to 24 times, adaptive PonderNet-style halting, and per-application top-8 routing through a dynamically growing and pruning expert pool. Expert weights and Adam moments are stored as files on disk; a working set is paged through RAM and VRAM as needed, allowing the pool size to exceed available VRAM. The project includes Python commands for building corpora, reading files, streaming training, serving a local web interface, and inspecting or replicating continual-learning experiments. The repository describes the current model as toy-level rather than frontier-capable, and its weights are generated in the local weights directory rather than distributed with the source.
Mobile MCP is an open-source Model Context Protocol server from Mobile Next for automating and scraping native iOS and Android applications on simulators, emulators, and connected real devices. It exposes a platform-agnostic set of MCP tools for device management, app installation and control, taps, swipes, text input, screenshots, screen recording, device logs, crash reports, deep links, orientation, location, and clipboard operations. The server primarily drives applications through native accessibility trees and structured UI-element data, falling back to screenshots and coordinate-based actions when required. It can run locally over stdio or as a Streamable HTTP server, and can use Mobile Next Cloud to reserve remote physical devices through the same tool interface. It is distributed as the `@mobilenext/mobile-mcp` npm package and requires platform tooling such as Xcode command-line tools or the Android SDK for device access.
NVIDIA Model Optimizer (ModelOpt) is an open-source library for optimizing deep-learning and generative-AI models for inference deployment. It accepts Hugging Face, PyTorch, or ONNX models and provides Python APIs for composing post-training quantization, quantization-aware training and distillation, pruning, neural architecture search, speculative decoding, and sparsity techniques, producing optimized or quantized checkpoints. The resulting checkpoints can be exported for inference frameworks including TensorRT-LLM, TensorRT, vLLM, and SGLang. The library also integrates with NVIDIA Megatron-Bridge, Megatron-LM, Hugging Face Accelerate, Transformers, and Diffusers for supported training and export workflows. It is distributed as the nvidia-modelopt package on PyPI and can also be installed from the NVIDIA GitHub repository or used through NVIDIA container images.
OpenBao is an open-source software solution for managing, storing, and distributing sensitive data such as secrets, certificates, and keys. It encrypts arbitrary key/value secrets before writing them to persistent storage, supports storage on disk and PostgreSQL, and can generate system-specific dynamic credentials such as AWS or SQL database secrets on demand. Generated secrets have leases that clients can renew through built-in APIs and that OpenBao can automatically revoke; revocation can also target individual secrets or groups such as all secrets accessed by a user or of a particular type. OpenBao also provides encryption and decryption without storing the data, along with audit and key-rolling capabilities. The project is community-run under open-governance principles and is implemented in Go, with a command-line server and web UI.
Open Code Review is an open-source, AI-powered command-line code-review tool developed from Alibaba Group’s internal review assistant. It reads Git diffs and sends changed files to a configurable LLM through an agent with tool-use capabilities; the agent can read complete files, search the repository, and inspect other changed files for context before producing structured comments tied to source lines. Its hybrid architecture combines deterministic engineering with dynamic agent decisions: deterministic stages select and bundle files, match rules to file characteristics, and apply independent comment-positioning and reflection modules, while scenario-specific prompts and tools guide context retrieval and review. The `ocr scan` command audits complete files or directories without requiring a meaningful diff, and the project provides multilingual rules for issues including null-pointer exceptions, thread safety, cross-site scripting, and SQL injection. It supports configurable model endpoints, including OpenAI- and Anthropic-compatible endpoints, and requires Git 2.41 or newer.
OpenMuse is a personal-agent application built by CopilotKit for iOS, Android, and web. It combines a persistent Chromium browser with an optional isolated Linux container, file and PDF handling, visible task plans, action reviews, and browser or terminal takeover so a person can inspect or continue the agent's work. It also provides Gmail and Google Calendar adapters with review required for sends and calendar changes. The application runs an API, durable task worker, and browser worker. Tasks use stored plans, progress, approvals, pauses, retries, and SQL leases; the browser worker maintains persistent Chromium profiles, while the optional non-root Linux container provides a retained workspace volume with bounded commands and no host-directory mounts or credentials. The interface is built with CopilotKit headless chat and AG-UI events, with inline email, browser, PDF, plan, and finance results. The repository describes OpenMuse as an alpha for self-hosting and building on, with model and Google-account configuration required for live agent and mail/calendar workflows. It is MIT licensed; CopilotKit Intelligence is a separate required service for rich conversation persistence and is not included under the repository's license.
QM is a multiplayer agent harness for startups that operates through Slack and a web interface. It gives each employee and room an isolated scope with its own memory, files, keychain view, permissions, scheduled jobs, web apps, and durable sandbox, while supporting collaboration in channels, group messages, and projects. Shared skills can be granted by scope, promoted by administrators, or imported from Git repositories; background work can run through crons, watches, and inbound webhooks, and internal apps can be published to selected users. A headless TypeScript core runs directly on Node with Fastify and handles identity, policy, scheduling, persistence, and the agent loop. Deployments can select harnesses and models including Pi, OpenCode, Codex, and Claude Code. PostgreSQL stores sessions, memory, queues, and other durable state, while a fixed tool surface includes an execute tool that runs commands in each scope's isolated, persistent sandbox. Slack is an optional in-process Bolt plugin; the web UI, admin panel, and public portal are optional HTTP API plugins built with Vite and Lit. Each deployment keeps organization-specific configuration, tools, skills, sandbox images, and infrastructure in a deployment directory validated and deployed by the qm CLI. Administrators can set organization-wide configuration, available harnesses and models, and a security posture. Strict mode pauses harness tool calls for human approval, Auto mode screens provenance-labelled external data and tool results with a classifier, and Dangerous mode disables content screening and pauses; a predeclared command policy with approvals and hard denials for operations such as recursive deletion or destructive SQL applies in every posture. Deployments run in the operator's own cloud account and are initialized from the @yc-software/qm package without requiring a source checkout; the project documents its threat model, operator assumptions, and known limitations in SECURITY.md.
Reverse Skill is an open-source, client-neutral cybersecurity skills router for AI coding agents performing reverse engineering, authorized penetration testing, security research, and CTF work. It processes a task through structured routing rules and a primary master-routing workflow, initializes authorization and network scope before action, selects a scenario-specific skill, checks available tools, MCP servers, and scripts, and records a timeline, evidence-to-finding path, report, and field journal rather than relying on guessed commands. Supported scenarios include APK and mobile analysis, binary and .NET reverse engineering, frontend JavaScript and encrypted-parameter analysis, DSL-VM reverse engineering, HTTP capture and request replay, malware and YARA analysis, penetration testing, attack-chain orchestration, case review, CTFs, firmware and IoT security, patch-diff analysis, exploit development, EDR bypass research, API and GraphQL security, supply-chain and SBOM security, and LLM security. Its documented tool ecosystem includes JADX, Apktool, Frida, IDA Pro, radare2, Ghidra, BurpSuite, and YARA. The repository provides platform-specific setup and tool-index refresh scripts for Windows, Linux, macOS, and Kali Linux, with prerequisites including a JDK, Node.js, and Python. It supports Claude Code, Codex, Cursor, OpenCode, and other compatible clients while keeping client adapters separate from the routing core, and includes cross-platform CI, routing regression tests, structure and supply-chain checks, and generated skill-navigation indexes.
Search is a small, fast WebKit browser for macOS developed by Office Commun. It provides a single address-and-search field, tabs, reading mode, picture-in-picture video, bookmarks, history, downloads, built-in ad and tracker blocking, per-site clutter hiding, and password storage in the macOS keychain. It has no account, sync, cloud storage, telemetry, or analytics; browsing data remains on the Mac. The browser uses WebKit and lazily creates each tab's web view, while its ad blocker runs as a compiled WKContentRuleList before network requests are made and its hidden-element rules are injected at document start. It can run Chrome extensions through WebKit's extension engine, adding shims for Chrome APIs that WebKit lacks; extension support requires macOS 15.4 or later. Search supports macOS 14 or later, is distributed as a free download of about 3 MB or through Homebrew, and is licensed under the MIT License. The repository contains a Swift and SwiftUI/AppKit implementation with no dependencies beyond Apple's macOS frameworks.
Soup is an open-source Python CLI for fine-tuning and post-training large language models from a YAML configuration and a single training command. It supports supervised fine-tuning, LoRA, QLoRA, preference optimization, quantization, automatic batch-size and GPU detection, and local training without a cloud service. Its optional layer-streaming mode keeps the frozen base model out of VRAM and feeds it to the GPU one decoder layer at a time, reducing memory use for low-VRAM systems. The project describes training an 8B model on a 4 GB laptop GPU with this mode; layer streaming is marked beta.
vphone-cli is an open-source macOS command-line tool for creating and running virtual iPhones with Apple's Virtualization.framework and PCC research VM infrastructure. It targets Apple Silicon Macs running macOS 15 or later and automates the VM workflow from IPSW download and patching through DFU restore, custom firmware installation, and first boot. The tool supports multiple firmware patch variants, VM creation and management, SSH and VNC connections, IPA installation, and export/import or APFS cloning of VM bundles. Its host control socket provides programmatic screenshots, touch and swipe input, hardware-key events, and clipboard operations, with each action returning an inline screenshot; the repository also describes an MCP server named vphone-mcp for AI-driven end-to-end testing. Operation requires Xcode and the iOS SDK, several Homebrew dependencies, and macOS SIP/AMFI relaxation to permit the required private virtualization entitlements.
WeKnora is an open-source, LLM-powered knowledge platform from Tencent for turning documents into queryable knowledge bases, retrieval-augmented Q&A, autonomous reasoning workflows, and self-maintaining Markdown wikis. It supports RAG-based quick Q&A and a ReAct agent that orchestrates retrieval, MCP tools, sandbox skills, and web search for multi-step tasks. Its Wiki Mode distills source documents into structured, interlinked Markdown pages with an interactive knowledge graph, manual editing, revision history, line-level diffs, and rollback. The platform ingests formats including PDF, Word, Markdown, HTML, images, spreadsheets, presentations, and XMind, and can synchronize sources such as Feishu, GitLab, Tencent IMA, Notion, Yuque, and RSS. Its modular pipeline supports interchangeable parsers, LLMs, embedding providers, vector databases, and storage backends, with dense, sparse, hybrid, parent-child, and graph-based retrieval strategies. WeKnora provides a web interface, REST API, command-line client, MCP server, website embed widget, and integrations with messaging channels. It can be deployed locally, with Docker, or on Kubernetes, including private and offline deployments. The repository states that it is licensed under the MIT License and supports workspace RBAC, scoped API keys, audit logs, and Langfuse-based observability.
Searchable transcript of Top Open-Source GitHub Projects : OpenBao, Magnitude, herdr, Compositor, WeKnora & BongoCat #296 — ManuAGI - AutoGPT Tutorials (13:45). Search for a phrase, then click its timestamp to jump straight to that moment in the video.
Captions sourced from the original video on YouTube, published by ManuAGI - AutoGPT Tutorials. The video, its captions and all related intellectual property remain the property of their respective owners; AINotes claims no ownership. Provided for research, accessibility and search — see the Transcript Notice and Copyright Policy.
00:00 Every week, developers release new open-source projects on GitHub. And this is our weekly GitHub project update video covering the best of them. Top trending open-source GitHub projects. This week, you'll discover useful and trending developer tools from local AI agents to self-hosted platforms. Without wasting time, let's get started. >> Before we jump into today's project updates, here's a quick announcement for everyone.
00:24 We've launched a brand new YouTube channel called AI Agent Studio dedicated entirely to AI Agent projects, tutorials, and tools. So, if you're interested in staying up to date with the latest AI agent open source projects, learning how to build your own agents, or exploring cuttingedge agent frameworks, make sure to check it out. Subscribe now to get weekly videos, in-depth guides, and real time project breakdowns.
00:50 The link is right there in the description. Don't miss it. All right, let's get into today's video. >> Project number one, Hindsight, long-term memory for AI agents that learns. Hindsight is open-source long-term memory for AI agents that learns. For coding agents, one package builds a per repo memory bank from your git history and past sessions, injects it as the agent starts, and adds knowledge pages, SDKs, and a rest API, expose, retain, recall, and reflect.
01:20 With light llm covering 100 plus models, it integrates with claude code, codeex, cursor, langraph, and crew AI. A local Damon runs it all in one command. Install it and give your agent memory. Project number two, non- Nvidia model optimizer. Compress models for faster inference on NVIDIA GPUs. NVIDIA model optimizer or model opt is an open- source library of model optimization techniques for compressing models and speeding up inference on NVIDIA GPUs.
01:49 You feed it a pietorch or onx model and stack techniques through Python APIs, quantization in FP8, NT8, NT4 and NVFP4 with post- training and quantization aware training plus sparsity pruning and distillation. The checkpoint deploys in tensor rtllm, tensor art, vllm and sgl lang with hugging face export for transformers and diffusers. Install it with pip and optimize a model.
02:16 Project number three, Asian open bow open-source secrets management forked from Hashi Corp vault. OpenBow is an open-source secrets management platform for storing and distributing sensitive data, secrets, certificates, and keys under open governance forked from Hashi Cororp vault. It stores arbitrary key value secrets encrypted before they touch disk.
02:40 Generates dynamic short-lived credentials on demand for AWS and SQL databases and revokes them after their lease and offers encryption as a service so apps encrypt data without managing keys. It adds audit logs, PKI and transit engines, and HSM/KMSbacked keys. Deploy it and centralize your secrets. Project number four, Reverse Skill. Skill Router for reverse engineering and authorized pen testing.
03:07 Reverse Skill is an open-source skill router that turns AI coding agents into assistants for reverse engineering, authorized penetration testing, and security research. When an agent hits an APK, a binary, front-end js, crypto, CTF or a pentest task, it routes to the right methodology, then calls tools like gidra, radar 2, and fida. Routing is a client agnostic config with a readonly case review layer for scope and findings.
03:34 It works in cloud code and cursor. Install it and route your security work. Project number five, mobile next MCP server for iOS and Android automation. Mobile Next is an open-source MCP server for mobile automation and scraping across iOS and Android emulators, simulators, and real devices through one platform agnostic API. It's accessibility first.
03:57 An agent drives apps from the native accessibility tree with no vision model falling back to screenshots and coordinates when needed. An LLM can automate user journeys, fill forms, and extract data from the screen. It works with claw code, codeex, gemini, and copilot against local or cloud devices. Install it and automate your app. Project number six, Wick Nora.
04:20 Turn documents into rag, agents, and auto wiki. Wikora is an open-source LLM powered knowledge platform from Tencent that turns documents into knowledge three ways. rag based Q&A for lookups, a React agent that orchestrates retrieval, MCP tools, and web search for multi-step tasks, and a wiki mode where agents distill documents into a self-maintaining interlin markdown wiki with a knowledge graph.
04:47 It ingests 10 plus formats, autosyncs from FU, notion, and UK, works with 20 plus LLM providers, and swaps vector databases. Self-host it with Docker and query your documents. Project number seven, Open Code Review. AI code review CLI from Alibaba. Open Code Review is an open-source AI powered code review CLI from Alibaba. Its hybrid design pairs deterministic pipelines with an LLM agent.
05:16 It reads your git diff, and the agent reads full files, searches the codebase, and inspects changes, then returns line level comments. The multi- language rule set catches issues like null pointer bugs, thread safety, XSS, and SQL injection, and it can auto apply fixes. It works with OpenAI and Anthropic with plugins for Cloud Code and Codeex. Install it and review your code.
05:40 Project number eight, God's Eye View, live open-source spatial intelligence on a 3D globe. God's Eye View is an open-source browser-based spatial intelligence console, a spy satellite simulator with real data on a 3D globe. It layers live aircraft, ships, satellites, wildfires, and CCTV, so you move between a global view and a single object. You ride a tracked flight, lock onto anything to see its trail, and hand off to the nearest camera.
06:10 It runs locally on public data, and by design, it tracks assets and systems, not people. Build it and watch the world. Project number nine, Magnitude. Private offline AI agent with local models builtin. Magnitude is an open-source AI agent that runs on local models with its inference engine. So, it's private and offline with no token costs or API keys.
06:32 It profiles your hardware, recommends the models your machine can run, then downloads, configures, and tunes them. Its Rust engine built on llama.cppates CPP calculates memory before loading and keeps parallel agents responsive. It uses your shell and edits files and skills at Excel, PowerPoint, PDFs, and Chrome. Install it and run agents locally. Project number 10, Vphone KSLI.
06:59 Run a virtual iPhone on Apple Silicon. Vphone KSLI is an open-source command line tool for creating and running a virtual iPhone on an Apple Silicon Mac. It uses Apple's virtualization framework and research VM infrastructure and a bundle handles firmware preparation and restore. You point it at iPhone and cloud OS firmware and it prepares, patches, restores, and boots the VM.
07:20 The touch enabled Windows supports file transfer, IPA install, screen recording, location, and an optional guest API. It ships an optional Launchpad GUI, build it and boot a virtual iPhone. Project number 11, DeskCom CRM. Self-hosted AI sales CRM for selling by chat. Deskcom CRM is an open-source self-hosted CRM for businesses that sell by chat. An alternative to KOMO, Octtoesk, and Intercom.
07:50 It combines a sales pipeline with AI agents and WhatsApp through WHA, so leads flow from chat and forms. The AI replies and books appointments, and a human can take over to stop the bot. It's MCP ready and multi-tenant with automations, scheduling, Google calendar sync, and LGPD compliance. One command installs the stack on a VPS. Deploy it and sell by chat.
08:13 Project number 12, Lash Soup. Fine-tune an 8B model on a 4GB laptop. Soup is an open- source tool that fine-tunes LLMs from a YAML file. One config, one command, no SSH. Its layer streaming keeps the frozen base out of VRAM and feeds the GPU one decoder layer at a time. So, an 8B model fine-tunes on a 4GB laptop card. It auto handles batching and quantization, ships templates for chat, code, reasoning, vision, and preference tuning, and lets you chat with, merge, and export to GGUF.
08:47 Install it and fine-tune from one file. Project number 13, Compositor. Free open-source Photoshop alternative for Mac. Compositor is a free open-source image editor for Mac, a Photoshop alternative for compositing and post-processing. It gives layers and folders with Photoshop's blend modes, layer masks, adjustment layers like curves, levels, and gradient map, and GPU rendered layer effects like drop shadow editable anytime, moves, scales, and rotations stay non-destructive, and it imports PSD files with blend modes
09:21 intact. AI agents can build projects, too. A comp is a folder of PNG layers plus a manifest. Download it and start compositing. Project number 14. Dend. Search. A small, fast, minimal WebKit browser for Mac OS. Search is a fast open-source WebKit browser for Mac OS with no toolbar, start page, or account. Just tabs and the page. One field does it all.
09:46 Type an address and go or words to search, completing from your history and sending nothing until return. Because it uses WebKit, the engine in every Mac, the app is 3 MB and opens instantly with no Chromium in memory. It keeps passwords in the keychain and runs Chrome extensions. Download it and browse quietly. Project number 15, Dashen's QM, multiplayer agent harness for a whole company.
10:10 QM is an open-source multiplayer AI agent harness open- sourced by Y Combinator and used internally at YC built for a company. It gives each employee their isolated workspace, scoped memory, files, credentials, crons, and a durable sandbox while people collaborate with the agent in Slack channels and projects. Pi, open code, codecs, and claude code all drive the same core, so you pick your harness and model.
10:37 Security postures range from per call approval to screened autonomy. It's early. Deploy it to your cloud. Project number 16. Herod is the always running terminal runtime for coding agents. Herod is an open-source multipplexer. The runtime your coding agents live on. It's a background server owning their terminals. So agents survive a closed lid, dropped network, or restart, reattach from any terminal or over SSH.
11:03 Every pane is marked working, blocked, or idle. So you never hunt for the stuck one. Agents drive it through a CLI and socket API and it runs what you run clawed code, codecs, cursor and more. One Rust binary install it and heard your agents. Project number 17 open muse personal agent with a browser terminal and files. Open Muse is an open-source personal agent with a browser terminal and files from Copilot Kit.
11:29 You ask for an outcome, follow the plan, and come back to the result. It runs a server and workers, giving the agent a computer. Chromium plus a Linux workspace, so it browses pages, runs commands, and handles files. Take over its browser or terminal anytime. It speaks AG UI, so you can swap the agent harness. It's a template for iOS, Android, web, build your agent.
11:54 Project number 18, Bans Open Glean, self-hostable AI knowledge workspace over Hydro DB. Open Glean is an open-source AI workspace for knowledge work built over Hydro DB, an Open Glean alternative. Ask a question across your memories, files, and connected apps, and it retrieves context, streams an answer with inline citations, and a sources panel, plus optional web search.
12:18 A deep research mode plans a DAG of sub questions, runs them against Hydra in parallel, and merges findings into one-sided answer. You bring your own open AI compatible model. Deploy it and ask your knowledge. Project number 19. Bongo Cat. Desktop Bongo Cat that plays along as you type. Bongo Cat is an open-source desktop pet. The Bongo Cat overlay that plays along with your keyboard and mouse reacting as you type and click.
12:44 It's a native app built in CC++ with SDL3 and OpenGL and stays lightweight. Input is read locally only to drive animations and shortcuts. It never records or uploads keystrokes and there are no ads, analytics or tracking. It supports live 2D models and runs on Windows with Mac OS and Linux and testing. Download it and boop. Project number 20, Mini AGI continual learning model trained from scratch on 8GB.
13:12 MiniAGI is an open-source continual learning bite level language model that trains from scratch on a single 8 GB laptop GPU and keeps learning from everything it reads. It assembles its own architecture and stores weights as files on disk, paging them onto the card as needed. So, its size is bounded by free disk space, not VRAM. It grows capacity, prunes what's unused, and reads through the same code path it serves on. It's an experiment. Clone it and watch it learn. Thanks for watching. See you in the next update.