The April 2026 news that nobody else covered: seven releases of a shipping Mac agent
Everyone reporting on LLM agents this month wrote about GPT-6, Claude Mythos, Llama 4, and research benchmarks. Fine, but that is the top of the stack. Underneath it, consumer agents kept shipping, one tagged version at a time, into the software that sits on users' laptops. This is a release-by-release log of what actually changed in one of those agents (Fazm) during the April 2026 window, with every fact pinned to a file path, a version tag, or a config file you can open.
Why the other April 2026 recaps miss this
Read the first page of Google for this topic and you get the same five stories: GPT-6 benchmarks, Claude Mythos preview, Llama 4's 10M-token context, a Hugging Face open-source post-training agent, and a Microsoft Research paper called Agent² RL-Bench. Every one of those stories is real. None of them is about agents as deployed software. They are all about the layer one level up: the model, the framework, or the paper.
A consumer on a Mac does not install GPT-6 or clone ml-intern. They install an app. That app has a version number, a changelog, a set of tool bindings, and a way of reading the screen. If you want to know what changed in LLM agents for normal people during April 2026, you have to look at those apps: at what got shipped, what got trimmed, what got rearchitected, and what new affordances the user gets when they open the app today versus on April 1. That is the report this page is.
The anchor fact: ~/.fazm/mcp-servers.json
The single most consequential shipping change in the April 2026 window was introduced on 2026-04-20 in Fazm v2.4.0. From the changelog entry, verbatim: "Added custom MCP server support via ~/.fazm/mcp-servers.json with Settings UI to add, edit, and toggle servers." That config file is where a user registers any Model Context Protocol server they want the agent to load at startup. It is not gated behind an app update or a command-line flag. It is a plain JSON file, owned by the user, reloaded by the desktop UI.
Most April 2026 coverage of "agent extensibility" focuses on enterprise governance: Databricks Unity AI Gateway adding MCP governance, fine-grained permissions, audit logs. All good. But on the consumer side, the story is the opposite: the interesting question is how to lower the bar for a single user to extend their own agent without reading documentation. A home-directory JSON file is the lowest-friction answer shipping today.
“Added custom MCP server support via ~/.fazm/mcp-servers.json with Settings UI to add, edit, and toggle servers.”
Fazm CHANGELOG.json, v2.4.0 entry, 2026-04-20
The architecture change underneath it: accessibility APIs, not screenshots
The April 2026 release cadence only works because of a choice made one cycle earlier. On 2026-03-27, v1.5.0 landed two entries that sit next to each other in CHANGELOG.json: "Bundled mcp-server-macos-use binary inside the app for built-in macOS automation support" and "Screen capture and macOS automation now uses native accessibility APIs instead of browser screenshot." Those two lines describe a full rewrite of how the agent sees the computer.
How the agent finds a button
// v1.4.x and earlier: screenshot-based
const image = await playwright.browser_take_screenshot();
const vision = await llm.visionCall(image, "Find the Send button");
await playwright.click(vision.x, vision.y);The bundled binary lives at Contents/MacOS/mcp-server-macos-use inside the signed app. It reads the AXUIElement tree of whatever window the user is focused on, returns structured nodes with role, label, and bounding box, and the agent proceeds deterministically from there. The April 2026 releases (2.0.x through 2.4.x) are the feature-level build-out of what this foundation allows: custom MCP servers, multi-window chat, native Mac integrations, and reliable control over non-browser apps.
What actually shipped, by date
A week-by-week timeline of the month, reading from the top of CHANGELOG.json. Each entry is one version, one real release, one ISO date. The items under each are the highest-signal changes, not the full changelog.
2026-04-03 - v2.0.1
Pop-out chat windows. The chat escapes the floating bar into a resizable native Mac window that survives app restarts. Push-to-talk microphone button next to Send.
Also in this release: knowledge graph capped at 100 nodes with texture-based rendering (performance), orphaned ACP bridge cleanup, paywall redesign with equal upgrade cards.
2026-04-04 - v2.0.6
Chat observer cards are now auto-accepted (deny to undo). The agent stops asking for manual approval on every read.
Fixed: AI agent typing its reasoning into user documents when controlling desktop apps. That was a real bug: the CoT leaked into a Word document during a live workflow because the model emitted the text stream into the focused text field.
2026-04-07 - v2.1.2 and v2.1.3
Upgraded ACP protocol to v0.25.0 with improved error handling for credit exhaustion. Added Manage Subscription in Settings.
This is the moment the protocol between the desktop app and the agent runtime gets first-class error handling for credit exhaustion, which matters because April 2026 is the month a lot of people blow through their free trial on new models.
2026-04-11 - v2.2.0
Custom API endpoint for proxies and corporate gateways. Global ⌘⇧N to open a new pop-out chat from anywhere.
Corporate-gateway support is the canary for April 2026's governance trend. Databricks shipped Unity AI Gateway on 2026-04-15; Fazm shipped a user-level equivalent four days earlier that routes through whatever endpoint the user's IT team approves.
2026-04-16 - v2.3.2
Privacy language tightened from 'nothing leaves your device' to 'local-first'. AI response stream isolation across pop-out windows.
Relevant because April 2026 also brought MIT Technology Review's piece on LLM-driven mass surveillance. Consumer agents got more careful about what they claim this month, not less.
2026-04-20 - v2.4.0
Custom MCP server support via ~/.fazm/mcp-servers.json. Dynamic model discovery. Referrals section. Trial shortened to 1 day.
The biggest release of the month. Rewrites how users extend the agent, how the agent discovers what models exist, and how the business model acquires users. Paywall now triggers after 3 messages for all users.
2026-04-22 - v2.4.1
Fixed paywall blocking users mid-onboarding. Fixed revoked sign-in causing retry loops. Fixed model label showing 'Smart' for Sonnet users.
Two days after the v2.4.0 monetization push, the follow-up release cleans up the edge cases that aggressive paywall triggering introduced. This is the pattern to watch in consumer agents: big feature releases followed within 48 hours by a ".1" that fixes the flows real users actually hit.
Month in numbers
Seven releases in 19 days is a ship-every-2.7-days cadence. That is not a research lab pace. It is the pace of a product-led consumer app treating agent features as a live target. Pair that with a trial cut from 21 days to 1 day and a paywall after 3 messages, and you have a clear statement of how April 2026 consumer agents intend to test their willingness-to-pay curve.
The shape of a consumer LLM agent, mid-April 2026
Zooming out one level, this is what the pipeline looks like for a user after the month's releases land. Voice comes in through push-to-talk. The runtime routes through the ACP protocol to whatever model the user selected from the now-dynamic list. MCP servers registered in the home-directory config file hand the agent its tools. The result is read back through accessibility APIs, not screenshots, and posted to whatever app is focused.
April 2026 consumer agent, end to end
What changed at each layer
Tool registration
Moved from compiled-in to user-editable. ~/.fazm/mcp-servers.json is now the source of truth. Settings UI reads and writes the same file. Any GitHub MCP server works without an app update.
Model discovery
Models populate dynamically from the ACP runtime's advertised capabilities. Newly released Claude variants appear without app update. No hardcoded allowlists.
Screen reading
Native AX APIs replace browser screenshots. AXUIElement tree gives roles, labels, bounding boxes. Deterministic click resolution.
Chat surface
Floating bar, detached pop-out window, or phone browser (chat.fazm.ai). All three share the same session. Global ⌘⇧N spawns a pop-out.
Business model
Trial cut 21 days to 1 day. Paywall after 3 messages. Referral credits ($49). Manage Subscription in Settings. Chat with Founder for a free month.
Privacy language
Tightened from 'nothing leaves your device' to 'local-first'. The agent does route through a runtime; the local-first claim is the honest one.
How to verify every number on this page
All of this is checkable. The repo is public. If you want to verify the April 2026 release log for yourself, here is the command sequence.
What the other April 2026 stories got right, and where this log fits in
This is not a takedown of the foundation-model coverage. It is a layer below it. The right way to read April 2026 is with both in hand.
| Feature | The common recaps | Shipping consumer agents |
|---|---|---|
| GPT-6, Claude Mythos, Llama 4 | Deep benchmarks and leaked specs | Show up in the dynamic model list after v2.4.0 |
| Agent² RL-Bench, ml-intern | Research paper, open-source post-training | Downstream: the models those produce get picked up by the runtime |
| Databricks Unity AI Gateway | Enterprise MCP governance | Mirror for consumer: ~/.fazm/mcp-servers.json |
| MIT Tech Review on surveillance | Policy and privacy framing | Why the privacy copy got tightened in v2.3.2 |
| Release cadence | Monthly drops, versioned frameworks | 7 tagged consumer releases in 19 days |
| How an end user extends the agent | Rarely covered | Edit one JSON file in your home directory |
What to watch for in May 2026
Leading indicators from this month
- Other consumer agents adopt a home-directory MCP config file, because ~/.fazm/mcp-servers.json demonstrably works.
- Trial length converges to 1 day plus a small message quota, following the v2.4.0 monetization shift.
- Screen-reading implementations move away from screenshots and toward native accessibility APIs on both macOS and Windows.
- Dynamic model discovery becomes a table-stakes feature, so users do not wait for app updates when new Claude or GPT variants ship.
- Privacy copy continues to shift from 'nothing leaves your device' toward 'local-first' or 'opt-in cloud' to match what agents actually do.
- Phone-as-agent-client patterns (like chat.fazm.ai) show up in more desktop agents, because the infrastructure is already there for remote control.
Version tags as a marquee
One line, scrolling. The complete April 2026 tag list for this consumer agent, in date order.
A single number worth staring at
public releases of one consumer Mac agent during the April 2026 window. Every one of them tagged, dated, and in the changelog of the public repo.
If you only read one takeaway
The research layer of LLM agents in April 2026 is loud. The deployed-software layer is quieter, but it is where the changes you will feel on your own Mac actually happen. A home-directory JSON file that registers MCP servers. A switch from screenshots to accessibility APIs. A dynamic model list that catches new Claude and GPT variants without an app update. Those are the consumer-agent news items of this month. They do not make a blog post title, but they change what the agent can do between the second you open your laptop and the second you close it.
Want a tour of the April 2026 build?
Fifteen minutes on a call, live agent, real accessibility tree, custom MCP registration in your own home directory.
Frequently asked questions
What changed in consumer LLM agents specifically during April 2026, not in research or foundation models?
Taking one shipping Mac agent as a worked example, Fazm tagged seven public releases between 2026-04-03 and 2026-04-22: 2.0.1, 2.0.6, 2.0.7, 2.0.9, 2.1.2, 2.1.3, 2.2.0, 2.2.1, 2.3.2, 2.4.0, and 2.4.1. The headline changes are user-editable MCP server config at ~/.fazm/mcp-servers.json, dynamic model discovery so newly released Claude models appear without an app update, a global ⌘⇧N shortcut to spawn a detached pop-out chat window, custom API endpoint support for corporate proxies, a shortened 1-day free trial, and a paywall that now triggers after 3 messages. Every one of these is in CHANGELOG.json in the public repo, tagged by date.
What is the most substantive engineering change of the month for consumer agents on Mac?
The shift from screenshots to native accessibility APIs as the primary way the agent reads the screen. Fazm's 2026-03-27 release (v1.5.0) bundled the mcp-server-macos-use binary inside the app at Contents/MacOS/mcp-server-macos-use and rewrote screen capture and macOS automation to use AX APIs (AXUIElement, kAXFocusedWindowAttribute, kAXRoleAttribute) instead of Playwright browser screenshots. This work is what makes the April 2026 cycle work reliably: the agent can now read Finder, Mail, Notes, Messages, Settings, Figma, Xcode, and anything else that exposes AX, not just the browser. The 2026-04-20 v2.4.0 release then layered custom MCP servers on top of this foundation.
What is ~/.fazm/mcp-servers.json and why does it matter?
It is a user-editable JSON file in the user's home directory where any MCP server can be registered, enabled, or disabled without an app update. The v2.4.0 release notes describe it as 'custom MCP server support via ~/.fazm/mcp-servers.json with Settings UI to add, edit, and toggle servers.' This matters because most LLM agent coverage in April 2026 talks about governance frameworks and enterprise gateways, but on the consumer side the interesting question is how a non-developer user extends an agent. A plain JSON file that is reloaded by the Settings UI is the answer: users can add servers from the GitHub MCP registry without rebuilding the app.
How does dynamic model discovery change the monthly release cadence for consumer agents?
Before v2.4.0, shipping a new Claude model meant bumping a hardcoded list in the app binary, signing a new build, and re-notarizing. The v2.4.0 changelog says 'Available AI models now populate dynamically from the agent, so newly released Claude models appear without an app update.' That unlocks a different cadence: the agent runtime (the ACP bridge) reports its capabilities, and the desktop UI renders whatever it sees. When Anthropic released Claude Haiku 4.5 in October 2025, or Opus 4.7 in the 2026 cycle, a user on 2.4.0 or later gets the new model without waiting for Sparkle to push an update.
What is the accessibility-tree vs screenshot distinction, in plain terms?
A screenshot is a picture. To find a button in a screenshot, the model has to do vision OCR, guess bounding boxes, and hope the button is labeled with text that matches what it was told to click. Accessibility APIs return structured data: a tree of elements with roles ('AXButton', 'AXTextField'), labels ('Send'), positions, hierarchies, and focus state. When an agent reads AXFocusedWindow on macOS, it gets an exact element graph instead of pixels. Click resolution becomes deterministic: match by label and role, not by pixel region. This is what Fazm's v1.5.0 switch accomplished and what every 2.x release this month builds on.
Did any consumer agent change its business model during April 2026?
Yes. Fazm v2.4.0 (2026-04-20) shortened the free trial from 21 days to 1 day and started triggering the paywall after 3 messages for all users, including built-in mode. It also added a Referrals section in Settings with a $49 credit banner on the paywall (the prior 1.7.1 release had introduced the referral program; 2.4.0 surfaced it more aggressively). This is a real shift in how consumer LLM agents monetize: not 'try the agent for three weeks,' but 'try three queries, then decide.' Worth watching because other 2026 consumer agents will likely converge on a similar pattern.
Where can I read the April 2026 release notes verbatim for a shipping consumer agent?
In the Fazm repository, CHANGELOG.json lists every version with an ISO date and a changes array. The April 2026 window covers versions 2.0.1 (2026-04-03) through 2.4.1 (2026-04-22). Each entry is one line per user-visible change. This prompt's angle is that this kind of log is richer source material for 'what happened in LLM agents' than any blog recap, because it is the ground truth of what actually shipped to end users.
Is the custom API endpoint feature a reaction to any April 2026 model release?
Indirectly, yes. When a lot of new frontier models land in the same month, teams want to point their local agent at different endpoints: OpenAI-compatible proxies for routing experiments, corporate gateways for compliance, or self-hosted inference behind a VPN. Fazm v2.2.0 (2026-04-11) added 'custom API endpoint setting for proxies and corporate gateways' precisely for that. Combined with v2.4.0's dynamic model list, a user can now point the agent at any endpoint that speaks the ACP protocol and have its model selector populate from whatever the endpoint advertises.
What got quietly removed or trimmed during the month?
The free trial got cut from 21 days to 1 day (v2.4.0). A buggy session-recording path was made less aggressive in 2.3.2: session recording now only captures when the user is actively interacting with the app (landed earlier, hardened in April). The v2.3.2 release also tightened privacy language from 'nothing leaves your device' to 'local-first' to more accurately describe what the agent does. These are small but informative: the month is not just additive, it is also a round of trimming assumptions that did not hold in production.