GitHub Trending Today #58: herdr-app, Raven, rerun, cinecode, terminal-mac, latch, gputaskman Transcript, AI Summary & Key Points

Asm is a package registry and command-line tool for assembly-language projects. It creates, builds, runs, tests, and distributes projects with explicit machine targets, source dependencies, repeatable lock files, and release records containing source, commit, artifact, size, SHA-256, and verification information. Its commands support package search, version inspection, source fetching, executable installation, upgrades, and uninstalling; publication is submitted through reviewed pull requests. The CLI supports Linux x86-64, Linux ARM64, and Apple Silicon macOS 15 or later, with RISC-V Linux packages and additional target implementations supported in the project model. Asm is in early alpha and is licensed under AGPL-3.0-or-later; starter code and reference examples use MIT.

Atlas is a native, GPU-accelerated, local-first file manager for Windows, developed in Rust with GPUI. It renders its interface on the GPU and provides Grid, List, Details, Miller-column and Tree views, tabs, split panes, keyboard-driven commands, indexed fuzzy search, previews, archive browsing, tags, filters, private folders and disk cleanup. It can convert images, video, audio and documents locally, and supports cloud and network storage through rclone, SMB, FTP and WebDAV. Atlas is available as a free alpha for Windows 10/11, with optional one-time Premium limits; a Linux build is in development and macOS is on the roadmap. The source code is licensed under GNU AGPL v3.0 or later.

blank-page is a skill for Claude Code, Codex, and other coding agents that expands a one-line motion-graphics idea into a detailed creative prompt. Developed by Stewdio, it searches for visual references, selects an aesthetic, finds motion references in GIFs or videos, writes a numbered sequence of beats with a creative twist, and suggests real assets such as 3D models, footage, or an actual user interface. The resulting prompt includes the references and up to three creative decisions for review; short edit requests can instead return a concise revision note. The skill is invoked with /blank-page in Claude Code or $blank-page in Codex and can be installed with npx skills add StewdioAI/blank-page. Its repository includes the skill procedure, reference materials, examples, and a Python script for searching sources such as Are.na, Pinterest, and Sketchfab and generating contact sheets. It requires Python 3 and curl, with ffmpeg used for contact sheets; the project is licensed under MIT.

CardboardXR is a Unity mobile VR app for Google Cardboard-style phone viewers. It renders split-screen stereo with lens distortion and uses AR Foundation with ARCore or ARKit for six-degrees-of-freedom head tracking, camera passthrough, detected surfaces, anchored objects, and hand tracking from the phone's rear camera. Its hand interaction supports pinch, grab, and a pointing ray. The app provides a lobby and demos including 360/VR180 video playback, air painting, an orb game, a passthrough portal, and physics-based objects placed on detected surfaces. It uses a single persistent core scene with additively loaded demos; hand tracking uses MediaPipe BlazeHand models through Unity Inference Engine and feeds XR Hands and the XR Interaction Toolkit. The repository includes Android build scripts and an APK release path; iOS builds must currently be built from source. It works offline for tracking and processes camera images on the phone. The project is licensed under Apache License 2.0, with separate licenses for bundled third-party components.

cinecode is an MIT-licensed open-source kit for Claude Code and Cursor that directs an AI coding agent to create short films, long videos, advertisements, explainers, and music videos. The agent turns a brief into a script and production plan, presents style and scene approvals, builds frames as code using Canvas 2D or Three.js/WebGL, designs synchronized sound, renders previews and final video, and delivers an MP4. It supports a Quick mode for shorter productions and a Film mode with staged approvals, research, storyboards, animatics, revisions, and final rendering. The repository includes camera and transition vocabularies, style presets, typography and sound-design guidance, beat-grid music-video tools, text-to-speech options, asset and license tracking, quality-control scripts, and headless Chrome, FFmpeg, Node.js, and Python-based rendering tools.

Claude Code Game Studios is an open-source template that turns a Claude Code session into a structured game-development team. It provides 49 specialized agents across directing, design, programming, art, audio, narrative, QA, production, and release, plus 74 slash-command skills, document templates, path-scoped coding rules, and validation hooks. Agents use vertical delegation from directors to department leads and specialists, while cross-department changes are coordinated by a producer; the collaboration protocol has agents ask questions, present options, draft work, and wait for user approval rather than operating as an autopilot. The template supports Godot, Unity, and Unreal Engine workflows, with project.yaml controlling engine settings, workflow rigor, documentation, review, team size, and testing behavior. It is distributed under the MIT License and is intended to be cloned or used as a project template with Claude Code.

djs is a small JavaScript and TypeScript engine written in no_std Rust and built mainly for WebAssembly. It runs codemode scripts, including model-written programs that call tools, process their results, and return data; the project describes it as a QuickJS replacement for pi's codemode. The virtual machine stores all state in one byte array addressed by 32-bit offsets, allowing a new VM to be created by copying a prebuilt image and allowing the host to read results directly from that memory. Its interpreter avoids using the Rust call stack for JavaScript calls, so execution can be suspended between instructions and resumed inside callbacks, getters, and asynchronous functions. A suspended tool call can be serialized as a continuation dump containing JavaScript code that recreates the run. A single compilation pass produces stack bytecode while erasing TypeScript during parsing, and clock, randomness, and weak references are supplied by the host so identical inputs produce byte-identical heaps in native and WebAssembly environments. djs is distributed as the @earendil-works/djs npm package and also provides a Rust CLI for running scripts, starting a REPL, and disassembling code. Its behavior is tested against dquickts, a QuickJS fork, and the repository includes documentation, architecture notes, testing instructions, and a pi codemode adapter.

EQUAL is a Go development and testing tool that records real HTTP API traffic and turns it into a local offline sandbox, a portable JSONL cassette, and an OpenAPI 3.1 document. It runs as a local reverse proxy: an application uses its existing configurable HTTP base URL, EQUAL stores observed requests and responses with built-in redaction, and replay serves the recorded responses without contacting the provider. Replay matches methods, paths, query parameters, and request bodies, with optional strict interaction ordering; unrecorded requests fail with a 404. The tool can also verify a recorded cassette against another implementation and report differences in statuses, headers, semantic JSON, and binary bodies. It is distributed as a static binary for Linux, macOS, and Windows on amd64 and arm64, requires no runtime, account, database, or language plugin, and is licensed under MIT.

Gauge CLI is an open-source command-line tool for evaluating and improving how coding agents discover and use Skills, CLIs, SDKs, MCPs, APIs, and documentation. It runs real agent sessions, captures normalized transcripts and outcomes, identifies friction such as blocked interactive commands, deprecated usage, unused Skills, incorrect limitation claims, or unclear multi-command workflows, and supports comparing proposed fixes with a baseline. The local workflow runs in Docker using configured agent credentials or subscriptions. Users initialize a project, define evaluations or preference prompts with tasks and criteria, execute agents such as Claude Code, Codex, Pi, or OpenCode against working-tree inputs, and have an agent analyze the evidence into structured findings and judgments. The CLI supports rerunning evaluations against changed Skills, documentation, code, CLIs, or SDK artifacts and comparing batches; hosted Gauge adds managed execution, shared measurements, authenticated inputs, and automation. The package is distributed through npm as @withgauge/cli and requires Node.js 22.13 or later and Docker for local execution. The repository is licensed under MIT. Local measurements and analysis are stored in a private .gauge workspace backed by SQLite.

gputaskman is a terminal GPU task manager for Windows, Linux, WSL, and Apple Silicon macOS. It shows device telemetry and per-process GPU memory, GPU activity, engine usage, system-RAM spill or paging estimates, and detected runtimes and model information. On Windows it combines WDDM process and adapter counters with NVIDIA NVML or AMD ADL/SMI data; on Linux it uses NVML, and in WSL it streams Windows host counters and matches them to Linux processes using GPU-context evidence. On macOS it reads IOKit and Metal statistics, where Apple GPUs share memory with the CPU. The tool's model inspector combines mapped DLLs, command-line arguments, model files, GGUF metadata, compatible runtime logs, and local server APIs to identify candidate models and estimate weights, KV-cache, and other buffer sizes. These classifications and breakdowns are estimates or runtime-reported values rather than GPU-memory inspection, and WSL per-process ownership can be inferred or guessed. It provides an interactive Textual terminal UI, one-shot text and JSON output, snapshot comparison, history charts, and support for local workloads including graphics, video, training, and several LLM runtimes. The project describes itself as early experimental software. It is distributed under the MIT license, with portable Windows and Linux builds and source-based execution on macOS; Windows releases include native x64 and ARM64 ZIPs and require a supported GPU with a current driver. It does not monitor cloud workloads, dump model weights, or enumerate individual game textures.

Herdr is a desktop application for terminals and coding agents, written in Rust on top of Ghostty's terminal engine. It organizes work into spaces, tabs, and panes, while an agents sidebar collects agents from those spaces and shows whether each is working, blocked, done, or idle. Its core keeps terminals running after the window closes; panes and spaces can run on local or SSH-connected machines, with remote work continuing while the local window is closed or the computer sleeps. The app supports layouts and JavaScript mods for customizing panels, menus, and dialogs, and provides command-line commands for inspecting the window and agent-driving guidance. Herdr is distributed for Linux, macOS, and Windows, with installers and source-build instructions in its repository. Release builds check for updates in the background and enable usage telemetry by default, which can be disabled in the privacy settings.

Immich Gallery is an independent open-source native macOS client for browsing and managing an Immich photo library through an Apple Photos-style SwiftUI interface. It connects to an Immich server using an email account or API key and provides a month-grouped timeline, albums, people, places, maps, Live Photos, video playback, HDR images, OCR and CLIP smart search, metadata editing, uploads with duplicate detection, downloads, sharing, and administrative views. The app also integrates with macOS through multiple windows and tabs, Quick Look, Finder drag and drop, Shortcuts, the share sheet, widgets, and in-app updates. It targets the Immich v3.3 API and remains compatible with servers from v2.0, requires macOS 15 or later, supports Apple silicon and Intel Macs, stores credentials in the macOS Keychain, and is released under the MIT License.

Kairos is an open-source, real-time fraud and risk decisioning platform from Finic. It evaluates sign-ups, logins, orders, payments, payouts and other events, returning an approve, step-up, review or block decision with a risk score and the signals behind it. Its workflows can query decision models, score events with logistic regression or gradient-boosted trees, branch on results, call webhooks, open review cases and make later decisions. The Grayson decision model answers plain-language questions about structured and unstructured data; its answers can be combined with transaction fields as features for traditional machine-learning models. Kairos also supports experiments that train and compare model variants on historical data before publishing a winning workflow version. The platform provides an MCP server for Claude Code, Cursor, Codex and other agents, enabling agents to inspect decisions, build workflows and run experiments with actions recorded under an agent key. It includes TypeScript and Python SDKs, a CLI, an HTTP API, a web console and Docker-based self-hosting, with Postgres managed by the server. The repository is licensed under Apache License 2.0.

Kiri is a lightweight container runtime for macOS, Windows, and Linux, developed in Rust as an alternative to Docker Desktop and OrbStack. It runs Docker images and supports the Docker CLI, Compose, buildx, Docker login, amd64 images through Rosetta, and the Docker Engine API. On macOS and Windows it runs containers in a small Linux virtual machine; on Linux it can use the host kernel. Its guest agent starts containers directly with clone3, namespaces, and cgroups, without runc, a shim, or dockerd. Kiri also provides a native interface for containers, images, volumes, and networks, local .kiri.local domains with HTTPS, one-command single-node Kubernetes clusters, and Linux machines. It is distributed for macOS, Windows through WSL 2, and Linux, and is licensed under Apache-2.0; the guest kernel is GPL-2.0.

Kubai Icons is a Claude skill for generating consistent icon sets as animated JavaScript/SVG icons or self-contained static SVG files. It applies documented rules for canvas geometry, layered shapes, outside outlines, gradients, shadows, colour, and object-specific hover motion; its JavaScript engine provides colour derivation, outlines, gradients, shadows, interpolation, and palettes. The repository includes 66 example icons, a live gallery, and a headless self-check that renders icons at rest, hover, intermediate animation states, multiple sizes, and each palette. It supports seven palettes and single-accent recolouring, and requires Node.js 18 or later plus Chrome for the self-check. The repository is source-available under the Kubai Icons License rather than open source.

LATCH is a JSON/HTTPS contract between an AI agent and a service for performing consequential operations. It exposes discovery through `/.well-known/latch.json` and uses five operations—ask, quote, approve, commit, and status—to separate read requests from execution: a quote returns an exact offer and SHA-256 digest, approval binds a trusted-host or user decision to that digest, and commit requires the offer digest, approval, and an idempotency key before the provider performs the operation. Status returns a task and provider assertion receipt. The protocol requires no SDK or runtime dependencies; its Node.js reference implementation includes a client, server, and CLI, with an independent PHP provider and optional MCP bridge. It does not provide identity, payment processing, a registry, hosted proxy, or a guarantee of exactly-once external effects. The wire version is experimental, and the repository is released under the MIT license.

LivLiv Player is a lightweight browser library for rendering animated 2.5D characters with WebGL2. Its dependency-free core is about 30 KB gzipped; it loads LivLiv model packages, VRM avatars, and PNGTuber images through related APIs, with expressions, actions, and parameter controls. Optional browser modules provide lip sync from microphone, audio, or video and webcam face tracking through MediaPipe, with camera images processed on the device. The package is distributed through npm as @livliv-dev/player, and the source code is licensed under Apache License 2.0. Featured characters included in the repository have separate licensing and usage terms.

Mago Studio is an open-source, local-first desktop AI video editor for Windows, macOS, and Linux. Mago Director turns scripts into casts, shot lists, storyboards, and connected scenes; Mago Transform applies frame-accurate edits to selected video ranges, including masked edits, relighting, character replacement, and style transfer. Projects and video remain on the user's machine by default, with rendering through ComfyUI on a local GPU or through Mago Cloud when signed in. Cloud renders upload only the selected Shot, its source video, reference images, and masks. The application can install a bundled ComfyUI setup on Windows or connect to an existing ComfyUI installation; local rendering is optional, and cloud models are paid. The app is distributed under the GNU AGPLv3, while downloaded models retain their own licenses.

NextTransformer is an open-source research framework, benchmark, and auto-research platform for comparing next-generation Transformer and language-model architectures. It trains recurrent-depth, sparse-attention, linear-attention, residual-stream, and hybrid methods on the same data under a fixed compute-matched protocol, then evaluates them at shared checkpoints against a dense causal-attention baseline. The framework records results as audited, file-backed runs and provides an in-browser observatory with learning curves, replayable leaderboards, architecture diagrams, model cards, and CORE task evaluations. Researchers can add methods through a plugin, calibration, formal training run, audit, and pull-request workflow; the repository also documents an experimental arena for agent-assisted architecture research. It is implemented for PyTorch training on Linux systems with NVIDIA GPUs and is released under the MIT License.

open-bot-avatar is an open-source JSON specification and renderer for interactive agent avatars. An avatar definition contains one SVG path in a 100×100 coordinate box plus eye position, spacing, and scale; the renderer infers volume from the outline as stacked cross-sections, producing animated canvas frames, static SVGs, or an optional CSS 3D scene. Avatars can follow a pointer, look in a fixed direction, hold their gaze, display idle, thinking, or happy moods, and support color, depth, reduced-motion, and accessibility-label options. The project includes a browser mounting API, an editor for selecting, coloring, generating, and importing avatars, and a gallery of community definitions. It is distributed as a dependency-free package, is MIT licensed, and provides validation utilities and DOM, SVG, and frame-rendering APIs.

Poki is an open-source tool that lets AI agents operate a real iPhone through a local MCP server and agent, without jailbreaking or additional hardware. It connects an iPhone to a Mac, exposes device-control tools to MCP clients such as Claude Code and Codex, and sends screen contents to the configured model service. Poki can also run a standalone terminal agent and authorized iMessage or Telegram channels. It requires macOS 15 or later, full Xcode, Node.js 24 or later, and a physical iPhone paired with the Mac over USB; the project documents limits including portrait-only support for verified physical devices. It is licensed under AGPL-3.0-only.

Quiet.md is a local-first notes application for macOS and Windows, drawn natively with Rust and GPUI rather than a WebView or Electron. It stores each note as a Feishu/Lark DocxXML file in an ordinary local folder, with no account, cloud service, or network activity by default. The rich-text editor supports headings, lists, to-dos, tables, columns, callouts, syntax-highlighted code blocks, LaTeX math, images, attachments, and covers. It can import Markdown and Feishu/Lark DocxXML, export Markdown and PDF, print directly, and preserve structure when pasting from Feishu or web pages. The app also provides document trees, pinning, favorites, full-text search, trash, configurable keyboard shortcuts, and built-in or custom themes. It is distributed for Apple Silicon Macs running macOS 13 or later and x64 Windows systems running Windows 10 version 2004 or later; the source code is licensed under MIT.

Raven is an open-source, multithreaded GUI library in C for building editors and tools. It uses a Vulkan renderer, native windows on Windows, Linux and macOS, and an XML-based interface model that applications load and connect to code written in C, C++, Rust, Python, Go, C#, Zig, LuaJIT and other supported languages. Interfaces can also be created through the C API, and changes to watched XML and theme files can be reflected while an application is running. The library provides Slate-style layouts, docking panels, a syntax-highlighting text editor, dialogs, file browsing, node graphs, timelines, color pickers, Markdown labels, themes, translations, accessibility support, touch and pen input, and solid, clear, frosted and tinted glass materials. It includes the raven_designer visual interface editor, raven_theme_editor, an inspector, examples and language bindings. Raven has an application-controlled frame loop, a render thread and job system, runtime-loaded Vulkan, native Win32, X11, Wayland and Cocoa backends, and CMake-based builds. It is distributed under the Apache License 2.0, with no royalties or paid edition stated by the project; the README notes that the public API is still subject to change.

Rerun is an open-source, self-hostable application for setting up teams of AI coworkers. A Foreman interprets plain-language requests, plans the team, and creates coworkers with separate jobs, skills, memory, and computers. Coworkers can use browsers and connected tools, collaborate on tasks, run on schedules or webhook triggers, and pause for approval before sending, paying, or deleting anything. It supports MCP servers, Composio integrations, and a catalog of ready-made integrations, and can use API keys, existing Claude or ChatGPT subscriptions, or models served through Ollama, LM Studio, vLLM, or an OpenAI-compatible server. The application is built as a Next.js and React frontend with Supabase, a per-workspace engine, the Vercel AI SDK, Hono, libSQL, Patchright, and Docker in a pnpm monorepo. The repository is licensed under AGPL-3.0. Self-hosting requires Docker, Node.js 22, a Supabase project, and an OpenRouter key according to the setup instructions. Some directories covering billing, plans, credits, managed models, and the template marketplace use the Rerun Commercial License.

runner-up is an open-source GitHub Actions runner coordinator written in Go that self-hosts jobs across Linux, macOS, and Windows machines. It uses GitHub runner scale sets to learn about waiting jobs, schedules them to available places according to pools, resource budgets, priorities, and ordering rules, and starts a one-job runner on the selected machine over SSH. Each job can run in a fresh Docker container or KVM virtual machine on Linux, a fresh tart macOS clone or Apple container guest on macOS, or a fresh Windows account or Hyper-V virtual machine on Windows; the environment is removed after the job ends. Host places are also supported but are not fresh. The coordinator runs on Linux or macOS, while Windows machines run agents or a Windows service. It uses the GitHub CLI for authentication, stores no GitHub token in its configuration, supports resource reservations and usage readings, and includes a default guard against jobs from fork pull requests. The repository describes runner scale sets as being in preview and identifies Hyper-V, Apple container, Windows account places, Mac borrowing, and draining as experimental. Distributed under the MIT license.

sand3d is a Python library for writing 3D objects as short programs designed for LLMs and human readers. Built over Blender's bpy, it represents profiles, solids, edits, frames, joints, groups, and assemblies as named parts; parts, edges, and faces are addressed by name rather than hand-computed coordinates, and the same program produces the same mesh deterministically. It can export meshes and articulated assemblies to GLB, URDF, OpenUSD, OBJ, and STL for web viewers, game engines, robotics and simulation tools, CAD workflows, and 3D printing. The library is distributed under the Apache License 2.0 and includes example programs, documentation, and datasets of procedural reconstructions.

SlideSense is a free, MIT-licensed native macOS menu bar app that turns the left and right edges of a MacBook Force Touch trackpad into brightness and volume sliders. It reads contact frames through Apple's MultitouchSupport.framework, interprets them with a gesture state machine, and changes system levels through CoreAudio and DisplayServices, with key-event fallbacks. It provides haptic step ticks, a Liquid Glass or frosted HUD, precise sliding, quick mute, cursor freezing, configurable edge bands and modifiers, and a bottom-edge media strip for track skipping and playback control. The app requires macOS 13 or later, Accessibility permission for media and key-event features, and uses no background daemon; its update check can be disabled and the project states that it collects no analytics or account data. Prebuilt releases are ad-hoc signed rather than notarised, and the repository includes build and test commands for compiling it with Xcode 15 or the Swift 6 toolchain.

SSHYard is a native desktop SSH and SFTP client for macOS, Windows, and Linux. It provides terminal tabs, split panes, search, shared host connections, automatic reconnection, an SFTP file manager with an editor and transfer queue, server metrics, port forwarding, jump hosts, and HTTP/SOCKS5 proxy support. For each host, its AI assistant can run commands and edit remote files through connected AI providers or locally detected ACP agents. It supports Ask, Review, and Full access modes, configurable review, and integrations through an open API, MCP, and Skill; agent access is disabled by default. Hosts and credentials can be imported from SSH configuration, stored in encrypted local storage, and backed up with password-encrypted archives to local storage, WebDAV, S3-compatible storage, or private GitHub Gists. SSHYard is developed as a Go application using MyGo and is distributed under the GNU Affero General Public License v3.0. Linux credential storage requires libsecret and an available Secret Service provider.

statechart-art is an educational visual resource explaining statecharts through four models of running a boba shop. Each chart is drawn in code with a pen-and-coloured-pencil style: an order moving through a shop, the shop's daily schedule, two concurrent barista and pot processes, and a bug showing old pearls being served. Each chart is backed by a TLA+ specification and configuration that TLC checks across reachable states and rules. A statechart-generation skill produces Mermaid charts and verification reports; layout and drawing scripts then render the charts and shop illustrations, with the site and images reproducible from the repository's Node-based tools. The repository includes the models, source scripts, generated GitHub Pages site, fonts and drawing assets, and releases everything under the MIT license.

Stroq is a local security platform and action firewall for AI coding agents, including Claude Code, Cursor, Codex CLI, Copilot CLI, Windsurf and Antigravity. It uses local hooks, a plugin for OpenClaw and a stdio MCP proxy to inspect agent tool calls before they run. Its rules can deny commands such as `curl | sh`, block outbound calls containing known secrets, or ask for confirmation when a command repeats text the agent read; provenance links risky actions to the content that influenced them. The CLI can read recent local agent sessions, key files and working folders without sending data or exposing secret values. `replay` evaluates existing sessions against the current policy, while `sent` compares session records with local credentials to identify keys that were sent, possibly sent, seen by the model, used with an agent's own service or blocked before execution. The local console records decisions and their reasons, shows agent hooks and MCP or skill changes, and listens on `127.0.0.1` without an account or telemetry. Stroq is distributed as the `@stroq/cli` package and requires Node.js 22 or newer. The project is written in TypeScript and licensed under Apache-2.0. Its documentation says it is a filter rather than a sandbox: hooks see only what an agent sends them, text matching cannot trace commands written from scratch to their source, and scripts, files or variables may carry secrets without appearing in a command.

terminal-mac is a macOS virtualization tool for giving computer-use agents a separate macOS desktop on Apple silicon Macs. It runs a real macOS guest through Apple's Virtualization framework; a helper app in the guest captures the screen and receives input events, while a terminal viewer streams the display through Kitty, Ghostty, or iTerm2. Agents control the guest through an MCP server with tools for screenshots, clicking, dragging, scrolling, typing, keyboard input, and status checks, without moving the host Mac's pointer or changing its clipboard. The guest has separate apps, files, accounts, and permissions, with host-to-guest communication through shared control and tools volumes. The project supports multiple virtual Macs, lifecycle commands, configurable virtual CPUs, memory, disk, and display settings, and optional offline operation. It requires an Apple silicon Mac running macOS 14 or later; initial setup requires downloading a macOS restore image and completing macOS setup and privacy permissions inside the guest. The repository describes it as an early release candidate, with self-signed releases and no Apple notarization.

Tucano is an open-source desktop application for coordinating teams of AI agents on macOS, Windows, and Linux. Agents work in chat-like channels, have roles, tools, memory, and token limits, and can hand work to one another for tasks such as investigation, coding, and review. It uses existing harnesses and sign-ins, including Claude Code, Codex, Cursor CLI, Grok, OpenCode, and Pi, rather than requiring a Tucano API key or per-token billing. Each agent runs in an isolated harness session with context assembled from its workspace, instructions, memories, skills, and recent conversation. Tucano exposes tools through a local MCP gateway that gives each turn only that agent's permitted tools; sensitive calls become approvals showing the exact action and arguments. It supports MCP connectors, scheduled routines, agent computers in Docker, browser sessions, and optional control of local desktop applications through Vercel Labs' agent-desktop. The application stores conversations, memory, logins, files, and configuration locally, using the operating system's keychain or keyring for connector credentials. It is distributed for macOS, Linux, and Windows and can also run experimentally without a window on a server. The repository is licensed under Apache-2.0.

uni-robot-agent is an open-source harness for robot-use agents, providing shared robot-tool interfaces across simulated environments and runtimes including Codex and Claude Code. It can run a built-in agent, invoke robot tools from a custom Python strategy, and inspect live or saved episodes. Its executor schedules agent and SDK tool calls, while adapters connect the shared tools to native environment backends; Direct mode uses robot tools, and Hybrid mode can add a compatible action model such as Pi0.5. The package includes a Python SDK, CLI, prompts, skills, and a dashboard that displays camera streams, tool calls, execution timing, and task outcomes, with replay controls for saved runs. It currently supports simulation integrations including RoboDojo, RoboUse, and RoboLab; supported physical-robot deployment is listed as forthcoming. The package is installable with pip or uv and is licensed under the MIT License.

Winc is a desktop code editor developed by DEEIX AI, written in Rust with its own CPU rasterizer and rendering engine. It provides multiple cursors, find and replace, syntax highlighting, lightweight go-to-definition and usage lookup without a language server, whole-file formatting through external tools, and previews for Markdown, images, SVG, XLSX, and CSV files. Its Git interface supports side-by-side and word-level diffs, hunk staging, commits, pushes, a commit graph, and blame; it also includes a project file tree, .gitignore-aware search, and an integrated terminal. Winc runs on macOS, Linux, and Windows, and is distributed through platform packages, portable archives, and GitHub Releases under the Apache License 2.0.