Hugging Face or GitHub for new AI projects in April 2026. They answer two different questions, and a Mac agent pulls from both.

Both, but they publish different things. Hugging Face publishes the model weights, the tokenizer config, the chat template, and the quantized variants (GGUF for llama.cpp, MLX for Apple Silicon, AWQ and GPTQ for cloud GPUs). GitHub publishes the agent harnesses, the MCP servers, the inference engines, the bridges, and the apps that wrap all of that into something a human runs. A weight file on Hugging Face does nothing on its own. An agent on GitHub still needs weights to talk to. The two platforms are complements, not substitutes, and a useful AI project usually shows up in both places at once: the model on Hugging Face, the code that drives it on GitHub. April 2026 produced new entries in both columns, and a consumer Mac agent like Fazm pulls directly from both.

Four releases passed the agent filter: working tool calling, an Anthropic-compatible serve path, MLX or GGUF support on Apple Silicon, and a license you can ship. Google Gemma 4 (E2B, E4B, 26B A4B, 31B variants) landed early in the month under Apache 2.0. Alibaba Qwen3.6-27B shipped on April 22 with native mlx-lm and mlx-vlm support. Qwen3-Coder-Next added a 3B-activated mixture-of-experts variant tuned for tool calling. Multiverse Computing's LittleLamb 0.3B Tool-Calling published a tiny tool-router model built from Qwen3-0.6B with their CompactifAI compression, sized to fit on the Apple Neural Engine. None of these four ship as runnable agents. They ship as weights, and a separate piece of code on GitHub turns them into agents.

Look at acp-bridge/package.json in the Fazm repo on GitHub at github.com/mediar-ai/fazm. The bridge sub-process is a small Node project, and its dependencies are: @playwright/mcp at 0.0.73, @agentclientprotocol/claude-agent-acp at ^0.29.2, @zed-industries/codex-acp at ^0.12.0, and ws at ^8.20.0. Three of those four are open-source agent infrastructure that landed or moved during April 2026. Fazm 2.4.0 on April 20 upgraded the Claude agent protocol to v0.29.2 and Fazm 2.7.1 on May 1 added the Codex backend that uses @zed-industries/codex-acp. None of these are vendored or forked. They install at their published version on every clean build, which is what makes them visible to a user reading the lockfile.

Through one Swift-side env var. In Desktop/Sources/Chat/ACPBridge.swift, lines 468 and 469 read UserDefaults for the key customApiEndpoint, and if it is non-empty, set ANTHROPIC_BASE_URL on the agent process env. The agent process is the Claude Code SDK, which obeys ANTHROPIC_BASE_URL when it sends tool-calling messages over /v1/messages. Set the field to the URL of a translation proxy (LiteLLM, claude-code-router, or any Anthropic-compatible adapter), point that proxy at a Hugging Face model running locally on llama.cpp, MLX, vLLM, or Ollama, and the agent talks to the local model instead of Anthropic. The harness on top does not change. The wire shape on the front is still Anthropic. The weights on the back are now whatever you pulled from Hugging Face that morning.

Because the surface area of a desktop AI agent moves fast. The Anthropic agent client protocol shipped a backwards-incompatible model-name change between the version Fazm 2.3.x used and v0.29.2 in 2.4.0. The Playwright MCP server fixes a Chrome profile bug almost every minor release. Vendoring those would mean the consumer Mac app re-bundles every fix as a code change of its own. Pinning by exact published version means the agent infrastructure on GitHub upgrades on its own cadence, the consumer app picks up the version it tested against, and the user sees a single CHANGELOG entry that says, in plain English, what changed. The April 26 fix in Fazm 2.4.2 is a good example: a stored opus model preference had to be migrated forward to the new ACP alias, and the migration is one Swift function rather than a re-vendored npm package.

Default path is accessibility, not screenshots. The frontmost macOS app exposes its UI as an accessibility tree of buttons, text fields, table cells, and labelled values. Fazm's agent reads that tree as structured data and operates on it directly. A vision model from Hugging Face would be doing the same job in pixels, which is slower, more tokens, and brittle across themes and color schemes. The accessibility path is also why Fazm works with native apps that have no DOM at all (Mail, Messages, Notes, third-party Mac apps), not just with the browser. Vision models still have a place when the accessibility tree is empty or dishonest, but for the everyday case the answer is structured first, screenshots only when forced.

Open Settings, scroll to the MCP servers section, click Add, and paste the same fields you would put under mcpServers in a Claude Code config: name, command, optional args, optional env. Save. The bridge sub-process picks the new server up on the next session start. Or edit ~/.fazm/mcp-servers.json directly. The format is documented at the top of MCPServerManager.swift in the Fazm repo and is identical to the mcpServers shape Claude Code uses, so an entry from a public README on GitHub pastes in with no rewriting. Fazm 2.4.0 on April 20 was the release that made this user-facing; before that, MCP servers were hardcoded in the bridge.

Yes. The agent harness is the same in both cases: it serializes Anthropic /v1/messages with tool_use and tool_result blocks, expects a streaming response in the same shape, and routes whatever comes back to the same parser. The only thing that changes is the URL the agent sends those bytes to. With customApiEndpoint empty, the SDK uses Anthropic's hosted API. With it set, the SDK uses your translation proxy, and the proxy answers in the Anthropic shape regardless of what the actual model on the back end speaks. From the agent's perspective there is no difference. From a privacy perspective there is a big one: when the proxy is local, the bytes never leave the machine.

If your task is short, latency-sensitive, and reasoning-heavy, the hosted frontier models still win. Multi-step planning across noisy accessibility input is where local Hugging Face models from April 2026 still trail Claude Sonnet 4.6 and Opus by a meaningful margin. The reason to wire them in anyway is that local models excel at the boring tool-call-heavy parts of a workflow: looking up a CRM contact, filling a recurring form, batching ten near-identical document edits. A common stack on a Mac with 32GB or more is a local 27B-class Hugging Face model for the bulk parts of a routine, and the hosted frontier model for the planning. Fazm's customApiEndpoint can be flipped between the two without restarting the app, so you can A/B in real time.

Open github.com/mediar-ai/fazm in a browser. The Node bridge dependencies are in acp-bridge/package.json. The Swift code that maps customApiEndpoint to ANTHROPIC_BASE_URL is Desktop/Sources/Chat/ACPBridge.swift, around lines 468 to 469. The Settings UI that exposes the field is Desktop/Sources/MainWindow/Pages/SettingsPage.swift, around line 984 (the placeholder text suggests a local proxy URL on a non-standard port). The MCP server config file is created by Desktop/Sources/MCPServerManager.swift, which writes ~/.fazm/mcp-servers.json with the same shape Claude Code uses. The April 2026 changelog entries are in CHANGELOG.json at the repo root, with the 2.4.0 entry on April 20 being the release that made user-defined MCP servers visible.