Open source AI on GitHub, past day: one real repo, read end to end
Every page that ranks for this keyword gives you a listicle of names. llama.cpp shipped, Codex CLI shipped, Hermes Agent shipped. True, and useless. This page takes a single open source AI project, github.com/mediar-ai/fazm, and walks the past twenty-four hours of commits against the actual Swift source so you can see what those commits buy a Mac end user.
Open source AI projects shipping on GitHub, past 24 hours
Why the roundup posts are not enough
Open the top five search results for this keyword and you get the same shape: a bulleted list of repos, a release tag, a one-line description, and a star count. Useful if you are writing a newsletter. Useless if your question is, does any of this change what I can do on my Mac today.
The missing step is reading the repo. Not skimming the release notes. Pulling the tree, diffing the commits, opening the files those commits touch. When you do that on a single well-chosen project, you find the part of the release that actually moves. Everything else is packaging.
I picked Fazm for three reasons. It is open source (github.com/mediar-ai/fazm). It shipped a tagged release in the past 24 hours (v2.3.2+2003002-macos, April 16, 2026). And it runs on an ordinary Mac without a GPU cluster, so the commits translate into user-facing behavior the same day they land.
Where a past-day commit on an open-source AI repo actually lands
When commit 2152e6a (add custom MCP servers) landed on main on April 16, it did not live in a vacuum. Here is the inputs-to-user path that past-day commit moves through on its way to a running floating bar on your desk.
From a GitHub commit to a user's floating bar
The last 24 hours on github.com/mediar-ai/fazm, literal
Not a summary, not a digest. Real `git log` output from a fresh clone, shortened to the commits that matter most. Every hash here resolves to a public commit on the main branch.
“The past day on this single open-source AI repo was a 56-commit burst that turned into one tag, three user-visible features, and one screenshot-free way to add any MCP server to a Mac agent.”
github.com/mediar-ai/fazm, April 15 to 16, 2026
What those commits add up to, in order
Six commits out of the 56 do the structural work. The rest are polish: copy fixes, button hints, bar padding. These six are the release.
0536a6e
Remove unused lastTextContentBlockIndex tracking
9d7d36a
Fix mid-word splits by removing text boundary on block index change
2152e6a
Add support for custom MCP servers via ~/.fazm/mcp-servers.json
07a19f5
Add session warmup to initialize MCP servers before first query
6d25fb7
Refactor available models to support dynamic updates from ACP SDK
v2.3.2
Tag v2.3.2+2003002-macos shipped on April 16, 2026
Three feature clusters, one day
Ranked by how much they change day-to-day agent use, not by what sits highest in the changelog.
Custom MCP servers via ~/.fazm/mcp-servers.json
Commit 2152e6a plus the MCPServerManager, edit sheet, and sidebar search item. You drop a JSON file in your home directory, Fazm reads it on launch, and every tool in those servers is now callable by the agent in the floating bar. No recompile, no plugin system, no marketplace. The same shape the MCP spec already uses for Claude Code, Cline, and the Anthropic Desktop app.
Session warmup for MCP before first query
Commit 07a19f5 stops the 'I asked and nothing happened for five seconds' experience on first use by spinning the MCP servers up during session creation, not on first tool call. A small-looking change with a large perceptual payoff.
Dynamic model discovery through the ACP SDK
Commits 6d25fb7 and abfb5ec add a models_available event on the ACP session, plumb it through the ACPBridge, and let the settings UI render whatever models the underlying SDK reports, not a hardcoded list. Useful the first time Anthropic ships a new model mid-session.
The one thing none of the roundup posts mention
A past-day commit on a Mac agent only matters if the agent already has a real way to read the screen. Most do not. Most Mac AI agents screenshot the display, compress the image, ship it to a vision model, and hope. Fazm reads the macOS accessibility tree directly with AXUIElement calls. Here is the actual source.
That is three lines of real Apple AX SDK. No image buffer, no vision-model token cost, no OCR guesswork on a retina antialiased button. When commit 2152e6a adds MCP server support, the agent can route a tool call into Linear or Slack and still give the tool the structured context of the user's focused window from those three lines. Screenshot-based agents have to re-ingest the entire screen on every tool result; Fazm already has structured state in memory.
“Fazm's screen-read primitive is AXUIElementCopyAttributeValue against kAXFocusedWindowAttribute, not a PNG. That is why the April 16 MCP commit compounds: new tools plug into real state, not pixels.”
Desktop/Sources/AppState.swift:441
How to track open source AI releases in the past day yourself
Three working patterns that do not require a paid aggregator. All return the actual bytes, not a summary.
Three ways to pull the past-day release stream on any open source AI repo
Releases Atom feed
Append /releases.atom to any repo URL. Example: https://github.com/mediar-ai/fazm/releases.atom returns the full tag stream as XML, consumable by any feed reader.
GitHub Search API
GET /search/repositories?q=pushed:>2026-04-16+topic:llm+stars:>500 returns every repo tagged llm that received a push in the past day, ranked by stars.
CHANGELOG on default branch
For projects that keep a machine-readable changelog (Fazm's is CHANGELOG.json), curl the raw file from main and grep by date. No scraping, no rate limit you care about.
Aggregator roundup vs. reading the repo
Both tell you that something shipped. Only one tells you what it changes for your machine.
| Feature | Read the roundup | Read the repo |
|---|---|---|
| Tells you a release tag name | v2.3.2 | v2.3.2+2003002-macos |
| Tells you commit hashes | Not included | 2152e6a, 07a19f5, 6d25fb7 and 53 more |
| Tells you the files that changed | Not included | MCPServerManager.swift and friends |
| Names the new config file path | Not included | ~/.fazm/mcp-servers.json |
| Shows the SDK call that reads your screen | Not included | AXUIElementCopyAttributeValue |
| Time to write | Thirty seconds | An afternoon reading the tree |
What to do with this, if you build agents on a Mac
Reproducible actions from one past-day open-source AI release
- Clone github.com/mediar-ai/fazm and run git log --since='2 days ago' to replicate the 56-commit burst verbatim
- Open Desktop/Sources/AppState.swift at line 439 to see the AXUIElementCreateApplication + AXUIElementCopyAttributeValue pattern any Mac agent should copy
- Write a ~/.fazm/mcp-servers.json with one server entry and verify Fazm picks it up on next launch
- Subscribe to the Releases Atom feed at github.com/mediar-ai/fazm/releases.atom for the next past-day release
- Read CHANGELOG.json on main to see how small the surface area of a user-visible release actually is (three lines for v2.3.2)
One past day, in one number
Fifty-six commits. One tag. Three feature clusters. Zero screenshots. That is the entire usable payload of a past-day open source AI release on a Mac-native agent, read against the actual Swift source on April 16, 2026. If you want a benchmark that is less impressive than a changelog graphic and more useful than one, this is it.
Questions people actually ask about past-day AI releases on GitHub
Frequently asked questions
What open source AI project shipped a release in the past day on GitHub?
Fazm (github.com/mediar-ai/fazm), a Mac desktop AI agent written in Swift, tagged v2.3.2+2003002-macos on April 16, 2026, the day before this page was written. It sits alongside a much larger past-day release stream on higher-star repos like llama.cpp (Vulkan flash attention DP4A shader in build b8779, Gemma 4 audio support in b8775, Qwen3 audio support in b8769), the OpenAI Codex CLI v0.121.0-alpha.4 with Realtime V2 background agent streaming, and Hermes Agent v0.8.0. Fazm matters here because it is the only one of those on a Mac end user's laptop by default, and because its daily commit stream is readable end to end without GPU infrastructure.
How many commits did Fazm ship in the past 24 hours?
Fifty-six commits over April 15 and April 16, 2026, verified with `git log --since='2 days ago' --oneline` on a fresh clone of github.com/mediar-ai/fazm. They cluster around three features: a new `~/.fazm/mcp-servers.json` custom MCP servers file (commit 2152e6a and a cluster of follow-ups through 0536a6e), dynamic model discovery through the ACP SDK (commit 6d25fb7 and the models_available event wiring), and an MCP session warmup that initializes servers before the user's first query (commit 07a19f5). The shipped tag that bundles them is v2.3.2+2003002-macos.
Why is the Fazm release interesting next to the bigger repos?
Because it changes what the end user can do on their own machine today, not what a backend engineer can do on a rented H200 next quarter. llama.cpp b8779's Vulkan DP4A flash attention is a real win for quantized inference parity, but to actually feel it on your Mac you still need a runner, a quant, a prompt loop, and an app on top. Fazm ships the app on top, and the April 16 commits mean you can now point that app at any MCP server you run locally (Linear, Slack, your own database, a local Ollama tool bridge) by writing one JSON file, with no recompile.
What does the accessibility-API angle have to do with GitHub releases in the past day?
It is the uncopyable part. Most roundup posts for this keyword list project names, star counts and release tags. They do not open any of the repos. If you open Fazm's tree you find Desktop/Sources/AppState.swift line 439 calling AXUIElementCreateApplication(frontApp.processIdentifier) and line 441 calling AXUIElementCopyAttributeValue(appElement, kAXFocusedWindowAttribute as CFString, &focusedWindow). That is the macOS Accessibility API reading the focused window directly. It is not a screenshot, not an OCR pipeline, not a headless browser. That design choice is what makes the past day's MCP commits land cleanly: Fazm already has real structured screen context, so a new MCP tool plugs into real state, not a pile of pixels.
How do I actually track open source AI releases in the past day on GitHub myself?
Three ways that work without a paid aggregator. First, GitHub's Releases Atom feed: append /releases.atom to any repo URL (example: https://github.com/mediar-ai/fazm/releases.atom). Second, the GitHub Search API with `pushed:>$(date -v -1d +%Y-%m-%d) topic:llm stars:>500`. Third, read CHANGELOG.json files directly on the default branch: for Fazm that is github.com/mediar-ai/fazm/blob/main/CHANGELOG.json and it lists v2.3.2 on 2026-04-16, v2.2.1 on 2026-04-12, v2.2.0 on 2026-04-11, and so on. Aggregator roundups give you listicles; these three sources give you the actual byte stream.
What does the new `~/.fazm/mcp-servers.json` file look like?
It is a plain JSON file that Fazm now reads on startup. The commit that introduced it (2152e6a on April 16, 2026) adds a new MCPServerManager in Desktop/Sources/MCPServerManager.swift, a search item in the settings sidebar (commit d97845e), and an edit sheet for per-server configuration (commit 6ca2185). The schema lines up with the MCP client convention: a top-level `servers` object keyed by server name, each entry holding `command`, `args`, and optional `env`. Because Fazm is already an ACP-based agent running the Claude Code SDK under the hood, any MCP server that works with Claude Code works with Fazm after this release.
What is the difference between Fazm and a screenshot-based Mac agent?
A screenshot agent captures the screen as a PNG, ships it to a vision model, and asks the model to find a button. That round-trip is slow, private data is pixel-encoded, and accuracy degrades on dense UIs. Fazm reads the accessibility tree directly through AXUIElement calls, gets back structured role-and-value data for every control in the focused window, and ships that structured context (not an image) to the model. The latency is microseconds to read the tree; the payload is small enough to include in a single tool call result; and nothing leaves your machine unless you explicitly ask the model to send it. The proof is the AppState.swift AX calls on lines 439 and 441.
See the April 16 release running on your own Mac
Fazm is a free download. The v2.3.2 build shipped yesterday includes the custom MCP servers file, the session warmup, and the ACP dynamic model discovery. Install it and you are reading the past-day release directly, no aggregator in the middle.
Download Fazm →
Comments
Public and anonymous. No signup.
Loading…