Latest AI news, Hugging Face + GitHub announcements, April 2026: what a shipping Mac app ingests from the release stream

Hugging Face shipped huggingface/skills (a public GitHub repo of reusable agent skills that plug into MCP-compatible assistants), expanded the Hugging Face MCP Server for Hub operations, and smolagents gained 4,100+ stars in its first two weeks as a lightweight CodeAgent + ToolCallingAgent library. LeRobot (Hugging Face's open robotics library) nearly tripled its stars year over year. Netflix shipped its first public Hub model, VOID (Video Object and Interaction Deletion). Gemma 4 launched across 2B-31B parameter variants. On GitHub the dense cluster was Google ADK (8,200+ stars), Meta llama-stack (6,400+), OpenAI codex-cli (5,800+) and Block's Goose (4,900+) for local-first agents with MCP support. Qwen3-Coder-32B, Codestral 2, SmolVLM2-2.2B, and Unsloth Llama 4 support also landed.

Fazm bundles 17 .skill.md prompt files inside its signed .app at Desktop/Sources/BundledSkills/. On first launch SkillInstaller.swift enumerates them via Bundle.resourceBundle (lines 31-40), then compares the SHA-256 hash of each bundled file against what is already installed at ~/.claude/skills/<name>/SKILL.md (line 120). If the hashes differ it removes and recopies; if they match it skips. A separate list obsoleteSkills = ["hindsight-memory"] at line 66 gets deleted from ~/.claude/skills/ on every launch. One of the bundled skills, find-skills.skill.md, is the glue: it teaches the agent to search both skills.sh (GitHub-hosted) and skillhu.bz in parallel and install matches into the same directory.

Trust path and cold-start latency. A bundled .skill.md that ships inside a notarized .app is signed, reproducible, and version-pinned to the Fazm build. A runtime pull from a public Hub would add a cold-start network call, a failure mode on first launch when the user has not yet configured anything, and a supply-chain trust boundary that Apple's notarization does not cover. The compromise in Fazm is exactly what April 2026's announcements model: ship a default bundle in the binary (the 17 skills), expose a discovery surface (find-skills + skills.sh + skillhu.bz), and let the user pull fresh skills over the network when they want new capabilities.

It does not sit behind a special integration; it sits in the same surface as every other GitHub-hosted skill repo. find-skills.skill.md line 11 registers skills.sh as the 'GitHub-based, community leaderboard' source, and line 47-49 maps `npx skills find [query]` as the search command for it. When an April 2026 Hugging Face skill ships to the skills.sh aggregation, it becomes searchable from inside Fazm without a code change. Installation uses `npx skills add <owner/repo@skill> -g -y` (line 86-87), which drops the skill into ~/.claude/skills/. The next launch of Fazm's agent picks it up the way any Claude Code runtime does.

Two shifts. First, skills as a durable primitive: a .skill.md file with YAML frontmatter is now a portable unit that travels from Hugging Face to skills.sh to skillhu.bz to ~/.claude/skills/ without transformation. Second, MCP as the default agent substrate: Hugging Face shipped an official MCP Server for Hub operations (huggingface.co/docs/hub/en/agents-mcp), Block's Goose ships with MCP as its spine, and Fazm's MCPServerManager.swift stores user-defined MCP servers at ~/.fazm/mcp-servers.json with the same shape as Claude Code's mcpServers dict. The effect is that 'install a skill' and 'wire an MCP server' are now the two knobs every consumer agent has to expose, and the April 2026 releases are what made it undeniable.

Not in the default control loop. Fazm's main agent loop sends a structured text payload (role, label, value, coordinates for each on-screen element, captured through AXUIElementCreateApplication) to a Claude tier, not pixels to a vision model. Hugging Face inference endpoints become reachable through Fazm's Custom API Endpoint field plus an Anthropic-shape shim in front, the same path used for GPT-5.2 or Gemini 3.1 Pro. Where Hugging Face specifically shines in this architecture is the bulk multimodal and open-weight models (SmolVLM2 for on-device, Gemma 4 for small-footprint assistants, Qwen3-Coder-32B for code-heavy skills) that can be self-hosted behind the shim for users who do not want their accessibility tree to leave their machine.

Seventeen. They live in Desktop/Sources/BundledSkills/ as .skill.md files and total 4,523 lines of prompt. They are: ai-browser-profile, canvas-design, deep-research, doc-coauthoring, docx, find-skills, frontend-design, google-workspace-setup, pdf, pptx, social-autoposter, social-autoposter-setup, telegram, travel-planner, video-edit, web-scraping, xlsx. The categoryMap at SkillInstaller.swift lines 14-23 groups them for the onboarding UI into Personal, Productivity, Documents, Creation, Research & Planning, Social Media, and Discovery.

It prevents bundled-skill updates from clobbering user edits that the user did not know they were making, while still letting Fazm push authoritative updates when a skill prompt actually changes in a build. Every launch the app hashes the bundled skill and the installed skill. If they match, skip. If they differ, overwrite. If the user never edited their installed ~/.claude/skills/<name>/SKILL.md, their local copy always converges on the latest bundled version. If they did edit, the next Fazm update that touches that skill's prompt overwrites their changes, because keeping a user on a stale prompt silently is worse than an overwrite they can see in a diff.

Not as the agent core, because Fazm's control loop is Anthropic-shape and routes through Claude Code / ACP, not through smolagents' Python runtime. What smolagents plugs into is the tools layer: a skill installed into ~/.claude/skills/ can shell out to `uv run` or `python -m smolagents` for specific tool-use steps. In practice that means smolagents is a great fit for skill-local sub-agents (e.g. a research skill that uses smolagents with Inference Providers to aggregate sources) while Claude drives the accessibility-tree control at the top.

The install path resolves by directory. ~/.claude/skills/<name>/SKILL.md is the unit of identity. If a user runs `npx skills add huggingface/something@deep-research` while Fazm already has a bundled deep-research.skill.md at the same path, the npx install wins at that moment (it overwrites). On the next Fazm launch SkillInstaller compares the bundled SHA-256 against the now-overwritten disk copy, sees they differ, and resets back to the bundled version. Collisions are therefore not destructive but they are clearly arbitrated: Fazm's bundle is the source of truth for bundled skill names. Users who want a HF-sourced override with the same name should install it under a different name or edit the skill file after Fazm's launch-time install completes.

The skill format has become the interface between 'stuff announced on Hugging Face and GitHub' and 'stuff your Mac agent can actually do.' You do not need to wait for a Fazm release to pick up a HF skill: find-skills can search skills.sh and skillhu.bz inside a chat session, install matches, and the next query can use them. You do need to accept the trust boundary, since a skill is a prompt that can embed instructions. Fazm's bundle is the trusted default. Anything installed at runtime is on you.