The open source AI projects on GitHub a real Mac app pinned in April 2026, with the file paths a reader can verify

Four. The acp-bridge subprocess inside Fazm is a Node project, and its package-lock.json declares: @agentclientprotocol/claude-agent-acp at version 0.29.2, @anthropic-ai/claude-agent-sdk at 0.2.112, @modelcontextprotocol/sdk at any 1.29.x or higher (transitively pulled by claude-agent-sdk), and @playwright/mcp at 0.0.68. The first three are the open Anthropic and ACP toolchain. The fourth is the Microsoft Playwright MCP server. All four resolve to public npm tarballs whose source repositories live on GitHub. None of them are vendored or forked; the consumer Mac app installs them at their published version on every clean build of the bridge.

It stores them at ~/.fazm/mcp-servers.json. The file is created by MCPServerManager.swift in /Users/matthewdi/fazm/Desktop/Sources/MCPServerManager.swift and uses an exact mirror of Claude Code's mcpServers JSON shape: each top-level key is a server name, each value is an object with command, optional args, optional env, and optional enabled. That means an MCP server entry that works in Claude Code's own .claude.json works in Fazm with no rewriting, and vice versa. If you publish an MCP server on GitHub with a one-line npx invocation in your README, that line is the value of command in both files.

Seventeen files at /Users/matthewdi/fazm/Desktop/Sources/BundledSkills/, totalling 4,523 lines. The largest is deep-research.skill.md at 856 lines. Then travel-planner at 534, docx at 481, doc-coauthoring at 375, pdf at 314, social-autoposter at 302, xlsx at 291, social-autoposter-setup at 274, pptx at 232, telegram at 203, find-skills at 148, google-workspace-setup at 132, canvas-design at 129, video-edit at 105, web-scraping at 58, ai-browser-profile at 47, frontend-design at 42. Each file is plain markdown with YAML frontmatter, readable by any human and any LLM, and free to fork or remix.

Three things. One: the Claude agent protocol was upgraded to v0.29.2, which is the @agentclientprotocol/claude-agent-acp package on npm and the agent-client-protocol repository on GitHub. Two: the available-AI-models picker stopped being a static array baked into the Swift binary and started populating from whatever the agent emitted at session start. So when Anthropic flipped a new model on the SDK side, Fazm picked it up the same day, with no App Store push. Three: a new ~/.fazm/mcp-servers.json file became the source of truth for user-defined MCP servers, with a Settings UI to add, edit, and toggle them, and the JSON shape was deliberately Claude-Code-compatible to keep the format portable.

SkillInstaller.swift in /Users/matthewdi/fazm/Desktop/Sources/SkillInstaller.swift does the work. On first launch it ensures ~/.claude/skills/ exists, then iterates every *.skill.md file inside the app bundle's Resources/BundledSkills folder. For each one it computes a SHA-256 digest of the bundled file and a SHA-256 digest of the installed file at ~/.claude/skills/<name>/SKILL.md. If the digests differ, the bundled version is copied forward. If they match, the file is skipped. There is no version flag and no manifest; the hash is the version. A skill update therefore ships in a normal app update without any user action and without needing the agent to be online.

Because Playwright MCP is the open-source bridge that lets a Claude agent drive a real Chromium tab through a stdio-spoken MCP interface. Pinning it at 0.0.68 means the bridge sub-process knows which build of the Playwright server is wrapped in this release of Fazm, so a regression in the upstream npm package does not silently move under the user's feet. Inside the Mac app the bundled web-scraping skill leans on it for headed runs, the bundled social-autoposter skill leans on it for actually posting in a logged-in profile, and the floating control bar uses it whenever a user asks the assistant to do something on a website that needs cookies and login state.

The screen-context path is structured. The floating bar can capture a window image when the user explicitly asks for one through ScreenCaptureManager.swift, but the default path that an agent uses to know what is on screen reads the live macOS accessibility tree of the frontmost app. That means the agent sees buttons, text fields, table cells, and table contents as labelled, addressable elements with roles and values, not as pixels. Compared to a screenshot-only loop, this is faster, smaller in tokens, deterministic across themes and color schemes, and works on any app that has an accessibility implementation, not just web pages running in Chromium.

Open Settings, scroll to the MCP servers panel, 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 it up on the next session start. If you prefer to edit JSON directly, open ~/.fazm/mcp-servers.json in any editor; the file format is documented at the top of MCPServerManager.swift and matches Claude Code's mcpServers exactly, so a config copied from a public README on GitHub is paste-compatible.

It nearly did. Before the upgrade the user's preferred model was persisted as the literal string 'opus' in UserDefaults. After the upgrade the protocol started using a new alias, 'default', for the dynamic Smart pill. Without a migration anyone who upgraded would have seen their Smart preference silently revert to whatever the new default was, and the pill would have lost its label. Version 2.4.2 on April 26 shipped a small migration in Swift that maps a stored 'opus' value forward to the new ACP alias on first launch after upgrade, so the pill stays Smart and the model stays Opus. The full release note in CHANGELOG.json reads: 'Fixed Smart (Opus) model preference not persisting after app update — now correctly maps stored opus to the new ACP model ID'.

Mixed. Some are original to this codebase (social-autoposter, travel-planner, ai-browser-profile, deep-research as written here). Others reflect well-trodden community patterns adapted into the same skill.md frontmatter format. All seventeen are stored as ordinary markdown so that a curious user can open one in TextEdit and read it, fork it, or replace it. Once installed at ~/.claude/skills/<name>/SKILL.md they are also visible to any Claude agent the user runs outside Fazm, including the Anthropic CLI, which is the same convention the open agent ecosystem on GitHub has settled on this year.