One open source AI project, one day, twelve commits, one JSON file that mirrors Claude Code

Fazm, the MIT-licensed macOS AI agent at github.com/mediar-ai/fazm. The 24-hour window covered here is the evening of 2026-04-16 through 2026-04-17, during which the project landed 12 sequential commits, every single one in service of the same feature: user-configurable MCP (Model Context Protocol) servers, wired through a JSON file at ~/.fazm/mcp-servers.json. The first commit in the chain is 2152e6a (Add support for custom MCP servers via ~/.fazm/mcp-servers.json) and the last is 0536a6e (Remove unused lastTextContentBlockIndex tracking). Every other commit hash, message, and timestamp on this page came from git log --since="2026-04-16" --until="2026-04-18" in /Users/matthewdi/fazm.

Because the listing pages already exist and do not help a reader decide anything. GitHub Trending, Product Hunt, and the llm-stats daily ledger each show 30 to 80 projects ranked by star velocity, with one-sentence descriptions. The thing that is missing from every single one of them is depth: what a day of shipping inside a real open-source AI project looks like. This page trades breadth for resolution. One project, one day, one feature, every commit explained.

It is a JSON file on the user's disk, loaded at app launch by MCPServerManager.swift (/Users/matthewdi/fazm/Desktop/Sources/MCPServerManager.swift, line 1). The comment on line 4 of that file reads: "Format mirrors Claude Code's mcpServers: { \"name\": { \"command\": \"...\", \"args\": [...], \"env\": {...}, \"enabled\": true } }". Matching Claude Code's schema is a deliberate choice: a user who already runs an MCP server for Claude Code (a filesystem server, a Postgres server, a Slack server) can paste the same configuration block into Fazm and it works. No conversion, no adapter, no new config language. This is the interop decision the 12 commits are built around.

Yes, straight from git log, all Pacific time on 2026-04-16: 17:56:47 2152e6a Add support for custom MCP servers via ~/.fazm/mcp-servers.json; 17:58:30 8cb97bd Add MCP Servers section to advanced settings; 17:58:47 9840b05 Add MCPServerManager to manage MCP server configurations; 17:59:22 401d79b Add Hashable conformance to MCPServerConfig; 17:59:34 eb797da Move MCP server state properties to top of SettingsContentView; 17:59:57 6ca2185 Add MCPServerEditSheet view for MCP server configuration; 18:00:45 8acaa07 Remove settingId parameter from MCP server settings card; 18:00:55 d97845e Add MCP Servers search item to settings sidebar; 18:01:07 5dfbb25 Fix homeDir reference to homedir() for user MCP config path; 18:03:18 3066157 Update MCPServerEditSheet frame to use fixed height of 500; 18:19:12 9d7d36a Fix mid-word splits by removing text boundary on block index change; 18:20:13 0536a6e Remove unused lastTextContentBlockIndex tracking. First to last is 23 minutes 26 seconds of pure feature work, plus two streaming-cleanup fixes at the tail.

It is a 114-line ObservableObject (final class) that owns the list of MCP server configurations and persists them to disk. It loads from ~/.fazm/mcp-servers.json at init, decodes the JSON into a [String: RawServerConfig] dictionary, then maps each entry to an MCPServerConfig struct with fields name, command, args, env, enabled. Save is atomic (write(to: configURL, options: .atomic)) and uses sortedKeys + prettyPrinted so the on-disk file stays diff-friendly. The manager exposes addServer, removeServer, updateServer, toggleServer. Nothing in that file touches an LLM, an AX API, or a network socket; it is pure configuration state. The integration surface is the servers array, which the chat runtime reads to decide which tools to mount.

The MCPServerManager sits between two other pieces. Upstream, the SettingsSidebar and the new MCPServerEditSheet.swift let the user add and toggle servers in the UI (commits 8cb97bd, 6ca2185, d97845e). Downstream, the chat runtime reads the enabled servers at session start and launches each command with its args and env, then speaks MCP (Model Context Protocol) to the resulting child process. From the model's point of view, every MCP server the user has enabled shows up as a set of callable tools with the native accessibility-API-style tool surface Fazm uses for desktop control. No screenshots, no vision tokens, just structured tool calls.

Clone the repo (github.com/mediar-ai/fazm), then run three commands. One: git log --since="2026-04-16" --until="2026-04-18" --pretty=format:"%h %ad %s" --date=iso prints the 12 commits above. Two: cat Desktop/Sources/MCPServerManager.swift prints the 114 lines described in the previous answer. Three: grep -n mcpServers Desktop/Sources/MCPServerManager.swift shows line 4, where the Claude Code schema match is declared in a comment. If any of those three checks fail, the commit has been rebased or renamed, and this page is wrong. As of publication (2026-04-17), all three pass verbatim.

Commit 3066157 on 2026-04-16 at 18:03:18 renamed an earlier dynamic-height attempt into a fixed 500-point frame. The reason is visible in the sheet itself: it is a form with up to four text-and-array fields (name, command, args, env), and without a fixed height the sheet resized every time the user tabbed between fields, which shifted the Save button up and down mid-edit. 500 is tall enough to show all four fields with a 16-row env table without scrolling. This is the kind of tiny, physical UI decision that never shows up in a GitHub Trending blurb but is most of what shipping software actually is.

Both, but the engineering-log half is non-negotiable. Every commit SHA on this page is git-verifiable. Every file path resolves. Every line-of-code count was counted. Every timestamp comes from the committer date, not a made-up date stamp. If a claim on this page cannot be verified by running a command in the Fazm repo, it should not be on the page. The marketing half is the framing: why a consumer-friendly Mac AI agent having user-configurable MCP servers is a better story than a generic open-source roundup. The engineering half is the proof that the framing matches the code.

Two things. First, treat Fazm as a data point about what open-source AI feature work looks like in practice in April 2026: short, focused, feature-per-day commit chains, interop with adjacent ecosystems (Claude Code's schema), and a real native UI on top of a file-based config. Second, if you already run MCP servers for Claude Code, Cursor, or Continue, you can point Fazm at the same list: copy your ~/.config/claude-code/mcp-servers.json (or equivalent) into ~/.fazm/mcp-servers.json and open the app. Every tool you already have now runs inside a Mac agent that drives real apps through accessibility APIs, not a browser-only or screenshot-only surface.

Listing pages optimize for coverage. This page optimizes for verifiability. A Trending page on 2026-04-17 might include Fazm as one row among 25 rows, with a star count and a one-line description. This page spends its entire word budget on a single day of that same project. You cannot find anywhere else on the web a list of these specific 12 commits in order, with the exact feature they add, with the exact JSON path they wire up, with the exact Claude Code schema they mirror. That is the thing the page is for.

Yes. The README at /Users/matthewdi/fazm/README.md line 50 reads: "## License" followed by "MIT". The full source of the Swift/SwiftUI desktop app is in Desktop/Sources, the ACP bridge is in acp-bridge/, and the AutofillCLI and AutofillExtractor helpers are also in acp-bridge/. The repo structure block in the README (line 30 onwards) names all three: Desktop/ Swift app, acp-bridge/ TypeScript bridge for Claude integration, dmg-assets/ installer resources. The builds and signed DMGs shipped via Codemagic are the same code.