Changelog

Release notes for Open Free Max — including the 4.0.0 AI Router and per-account isolation, 3.5.0 recurring work and the Antigravity provider, and the 3.0.0 local knowledge release.

Release notes for Open Free Max, newest first. Versions follow semantic versioning; MINOR bumps add features, PATCH bumps fix issues. Grab the latest build from the download page.

4.5.0 — Official agents, ready to install

Building a dependable agent workflow used to mean collecting prompts, scripts and conventions from different places, then hoping they would still work together. 4.5.0 adds an official library of reusable agents directly in OFM: browse a curated catalogue, see what each agent is for, and install it safely into the project where you need it.

New — official agent library

Curated reusable agents — an official catalogue of specialized agents, with a clear purpose, required capabilities and installation guidance for each one.
Safe, project-scoped installation — an agent is installed into the selected project; OFM verifies the package before writing it and keeps the installation separate from your other projects.
Compatibility shown before installation — the library checks the agent’s requirements against the CLIs and capabilities available in the app, so an unsupported setup is visible before you use it.

Improved

The agent bar is more stable when sessions change, and its state is preserved consistently across the workspace.
The Dashboard navigation is restored in the main application bar.

Security and sign-in

Remember me is now a real choice at sign-in. When it is turned off, the session stays in the running app and is removed at the next launch; authentication tokens remain managed by the native layer.
Service-side release operations now include rollback safeguards.

Windows and macOS

Version 4.5.0 is available as a Microsoft-signed Windows MSI and a signed, notarized Apple Silicon DMG.

4.0.0 — The AI Router: the right level, on the right subscription

You own several agent subscriptions and several models inside each. Choosing between them before every task is a decision you take dozens of times a day with no information — so most people take it once and never revisit it. 4.0.0 answers it for you: describe the task, and the router returns the CLI, the level of thinking the work needs, where that decision came from, and the model. Then you launch, or contradict it in one click.

The levels describe the work, not the trade it belongs to, which is why a writing or a data task is routed as confidently as a refactor. And nothing is ever re-routed mid-task: a running session keeps its CLI, its model and its transcript.

What changes when you update

The router is on. On new installs and on existing configurations alike. An agent launched from the router without a level of your own will no longer necessarily get the model you had set by hand in that CLI’s settings. You can turn the router off in Settings → Routing, and that choice is then kept.

Local state is now per account. Two accounts on the same machine used to see exactly the same projects, memory and orchestrator. Everything that belongs to you moves into its own folder, once, at the first launch after signing in. Switching account requires a restart. See Data & file locations.

New — AI Router

A level per task — quick task, regular work or hard thinking, decided per launch and stated in plain language, with the model it implies. See AI Router.
The choice is shown before anything starts, and one click sends the task to another CLI instead. The router proposes; you decide.
You can name the level yourself — write it in the brief and the router uses it as given.
It flags what the work needs — when a task calls for something in particular, such as producing or reading an image or live information from the web, the panel says so before you launch a CLI that cannot do it.
Several opinions at once — tick more than one subscription and the same brief starts on each, every agent with its own model. Marginal cost zero: those plans are already paid. See Several opinions at once.
It remembers your preferences per project — the times you overrode it, never the times you simply accepted. Shown in the panel, with a Forget button.
Attachments in the router — the screenshot strip and the file picker from the launch dialog, because a capture is often the reason you are launching at all.
The AI part of the decision is paid by OFM and never touches your subscription quota. It needs a connection.

New — quota, watched and preserved

A gauge per subscription on Mission Control’s permanent status bar, with the untouchable reserve drawn in. A preserved quota is invisible by nature.
Subscription switching for unattended work — when a subscription is spent, the orchestrator starts the next unit on another one. Only ever on a fresh session, and only towards a CLI you have already run yourself on that project. See Quota and subscription switching.
Fallbacks are visible and revocable — chips on the project header; one click excludes a CLI, and the exclusion outlives later launches.
A reserve nobody spends automatically. When everything is spent, the router steps down a level, or parks a unit that can wait, rather than dip into it.

New — elsewhere

A library of reusable recurring tasks — save a whole recurring job (CLI, sub-agent, cadence, active hours, turns per day, checklist, prompt template, pinned model settings) once and bring it back on any project in one click. Reusing an entry fills the form and stops there: a template three months old does not start running on a new project from a single click. See Recurring work.

Changed

A subscription can take a level even when no model can be forced on it. Cursor Agent exposes no model selector, so it now takes routed work and starts on its own setting. Claude and Codex come pre-mapped; tick the others in Settings → Routing for the levels you want them to take.
The router button has its own place in the agent panel bar instead of scrolling away with the seven providers, and it states the CLI, the level and its provenance in plain language, with the exact flags folded behind a disclosure.
Recurring-work modals close with Escape or a click outside; a single Escape closes one level of a modal stack, not two.
Installation notes are now per provider and per OS, with several notes possible for one pair.

Fixed

Rate limits and failures are read correctly on a unit that was moved to another subscription.
Interface labels no longer render as {Claude Code}.
A file path added by clicking now counts in the task description the router reads.

Windows

MSI and NSIS installers signed with Azure Trusted Signing, published as the verified APPS VELOCITY publisher: OpenFreeMaxIDE_4.0.0_x64.msi.

3.5.1 was built but never published. Its one user-facing addition — the library of reusable recurring tasks — is folded into this release.

3.5.0 — Recurring work: agents that run on a schedule

An agent is only useful while you are watching it. 3.5.0 changes that: you write a checklist, pick an interval or fixed times of day, and OFM wakes an agent to work through it — on your own subscription, in an interactive session, never the API meter.

The design decision that matters is what happens when there is nothing to do: the turn writes a silent acknowledgement instead of a report. A scheduled agent that reports every turn trains you to stop reading its reports within a week.

New — recurring work

Scheduled agent jobs — as many as you want, across as many projects as you want, running alongside the sessions you drive by hand. Interval or fixed daily times, with an optional active-hours window in your machine’s own time zone. See Recurring work.
Report or acknowledge — a turn ends with either a readable report or a silent acknowledgement. Only real work becomes something to open.
Ambiguity stops the turn — a genuinely undecidable question is never guessed at. It reaches the approval queue tagged as coming from a recurring job, with a system notification.
Per-job CLI settings — approval flags, memory and vault access pinned to one job, so your careful interactive setup stays careful.
Per-job journal and blocking causes — trace what a job’s turns did without wading through every other session in the project.
Programmatic flags refused — -p, --print, exec and --output-format are rejected. A job that runs while you are not watching cannot open a metered path.
Honest limits, stated up front — if the machine sleeps, no turn fires, and missed slots are never replayed.

New — elsewhere

Antigravity replaces the Gemini provider — Google’s agent CLI is now driven as agy. See Antigravity.
User-declared MCP server registry — register your own MCP servers and attach them to agents, alongside the built-in Obsidian and memory servers. See MCP servers.
OpenTUI capability probes — the embedded terminal answers them, so OpenTUI-based CLIs render correctly.
Machine time zone for active hours — read from the system, nothing to configure.

Changed

Plan mode removed — bulk campaigns, briefs, the assistant and recipes are gone. Recurring work replaces the one thing they were actually used for, without the ceremony.
Session actions moved to the left of the agent tabs; the separate Settings button is gone.

Security

Explicit CSP for the webview, and an inert Markdown preview.
API and orchd hardening, with vulnerable dependencies bumped.
Local execution traces are no longer written into the repository, and no secret is versioned.

Fixed

Codex sessions from earlier runs are found again.
Usage rates you typed are no longer lost, and writes are serialized.
Claude’s variadic launch flag no longer swallows the initial prompt.
The explorer tree refreshes in place instead of flickering under the cursor.
Orphaned confirmations are closed when a session dies.
A session is no longer parked on a welcome banner mistaken for a rate-limit window.
Removing a recent project no longer shifts the ones after it.
Native path separators, and URLs opened without going through cmd.
tauri dev state is isolated from the installed app’s state.

Windows

MSI and NSIS installers signed with Azure Trusted Signing, published as the verified APPS VELOCITY publisher: OpenFreeMaxIDE_3.5.0_x64.msi.

macOS

Apple Silicon DMG, signed with the Apps Velocity Developer ID and notarized by Apple.

3.1.0 was built but never published; its changes are folded into this release. Cursor Agent support, added as the seventh CLI in 3.0.0, was missing from those release notes.

3.0.0 — Local knowledge, direct Obsidian access and signed release

OFM now combines two complementary knowledge sources: its own private, Obsidian-like project base and an optional direct connection to the vault open in Obsidian.

New

Native local knowledge base — structured notes, durable facts, links, provenance and semantic search per project, stored entirely on your machine.
Obsidian vault over MCP — connect Claude Code or Codex to the MCP server built into the Obsidian Local REST API plugin. Agents can search, read and write the live vault directly.
Clean separation — OFM never copies or ingests the external vault into its own base; both sources can coexist in the same agent session without ambiguity.

macOS

First production DMG signed with the Apps Velocity Developer ID and notarized by Apple.
Gatekeeper accepts both the DMG and application as Notarized Developer ID.
Apple Silicon (arm64) on macOS 10.15 or later.

Windows

Production MSI signed with Azure Artifact Signing and published as the verified APPS VELOCITY publisher.
x64 installer for Windows 10 and 11: OpenFreeMaxIDE_3.0.0_x64.msi.

2.5.0 — Persistent memory: your agents stop forgetting

Coding agents are brilliant and amnesiac — every session starts from zero. 2.5.0 gives Open Free Max a persistent, local, per-project memory so the work you’ve already done with one agent is available to the next one, in any CLI, without re-explaining anything. It all lives on your machine; nothing is uploaded.

New

Persistent project memory — OFM keeps a private knowledge base of what happened in a project: past conversations, decisions and durable facts. It’s scoped per project and stored locally in a single SQLite database in your config directory — no cloud, no account, no API.
Memory panel — a dedicated screen with two tabs: Search (semantic grep over the ingested history of a project) and To review (human-readable facts distilled from your sessions, waiting for your approval before they’re trusted). Nothing is remembered behind your back — you curate what sticks.
Distill the last session — one click summarizes your most recent conversation into durable facts, on your own subscription (an interactive CLI session, never the API meter).
Two ways memory reaches the agent — for Claude, OFM wires an MCP server (memvault) the model can query on demand; for every other CLI, the approved facts are written into a tagged block of the project’s guidance file (CLAUDE.md / AGENTS.md / GEMINI.md). The block is idempotent and only that zone is ever touched.
Cross-CLI ingestion — history is read from each agent’s own store (Claude, Codex, Vibe, Gemini), so memory is shared across the tools you switch between.

Improved

Installation assistant v2 — the in-app onboarding now tracks each agent CLI through three honest states — declared → installed → authenticated — so you can see at a glance which tools are actually ready to launch, with verified install steps per CLI and runtime prerequisite (Node / Python).
Dashboard refinements across the project bar and agent panels.

Notes

Memory is local and private by design: the store (memory.db) sits in your config directory and never leaves your machine.
Available for Windows (.msi) and macOS (.dmg).

2.0.0 — OFM Pro: the subscription launch

The first major release. Open Free Max becomes a product you can subscribe to — with a 7-day free trial, no credit card — while everything that made it useful stays exactly where it was. Your agent usage still runs on your own Claude / Codex / Gemini plan; this is the price of the IDE itself. See Pricing.

New

OFM Pro plan — sign in inside the app and a 7-day free trial starts automatically (no card required). At launch, the plan was offered at $9.90/month. The launch offer has since ended; the standard price is $19.90/month.
Account & billing in-app — create an account, manage your subscription and invoices, all from a native screen; secure checkout and customer portal.

Improved

Hardened licensing — your access is now backed by a cryptographically signed entitlement verified by the app, so the subscription check can’t be spoofed by redirecting the app to a fake backend. Nothing changes for legitimate users.

Notes

Available for Windows and macOS (universal .dmg, Apple Silicon & Intel).
This is a major version per semver: the move to a subscription is a breaking change in how access works, even though the feature set is fully preserved.

1.3.1 — In-app installation assistant

New

CLI installation assistant — when an agent CLI is not detected, the agent panel now shows a full step-by-step installation guide instead of a bare “binary not found” hint. The assistant is aware of your runtime prerequisites (Node / Python + uv), your OS and target (Windows, macOS, WSL), and the selected provider (Claude Code, Codex, OpenCode, Vibe, Gemini, Kimi…). If Node or Python is missing, the runtime install step appears before the CLI step — no more guessing why npm install -g fails.
Copy & Open-terminal actions — each command has a Copy button. A secondary action opens a managed terminal in the app with the command pre-filled; you confirm and run, nothing executes silently.
Re-check button — re-probes the target after installation, instantly updating the panel state without restarting the app.
Fully localized — all labels respect the active language (English / French / Spanish); shell commands are intentionally never translated.

Improved

Dashboard refinements — tool-presence badges and install-help entry points integrated into the project tile and agent-panel header.

1.3.0 — Recent projects, one click away

New

Recent projects in the + menu — the + button now lists your recently opened workspaces right under Open / WSL / Clone, so reopening a project is a single click. It works in both places you add projects: the IDE left rail and the Dashboard project bar. Projects already open are hidden from the list (nothing to reopen), and picking one moves it back to the top of the history.

Improved

Dashboard project bar: the + button is larger and accented so it’s easier to find, and project tiles on a row now align to a consistent height.

1.2.4 — Mission Control split view & a definitive Explorer fix

New

Split view in the Queue — selecting a pending confirmation in Mission Control now shows the agent’s live terminal in a right-hand pane. You see exactly what the agent is asking while you approve or reject it, without leaving the queue.

Improved

Auto-purge of the queue — answer an agent directly in its own CLI and its pending confirmation disappears from the queue automatically, instead of lingering as a stale item.
Navigation rename — the first tab is now labelled Dashboard (was Wall); the top bar reads Dashboard · IDE · Plan · Settings.

Fixed

“Reveal in file explorer” on Windows (definitive) — a regression had brought back the old “opens Documents” bug for paths containing spaces. The path is now passed to Explorer as a raw OS argument, so /select, is never mangled.

1.2.3 — Explorer fix & folder colors

Fixed

“Reveal in file explorer” on Windows — paths containing spaces in the file or folder name (e.g. Open Free Max IDE_1.2.3_x64_en-US.msi) now open the correct directory in Windows Explorer. Previously the OS fell back to the user’s Documents folder silently.
Folder icon tint — yellow folder icons in the file tree are reliably applied; the CSS class was not attached in the search-results view and is now consistent.

1.2.2 — App icon refresh

Improved

All platform icons (Windows, macOS, Linux) rebuilt from the latest source logo — crisper edges at every size from 16 px up to 512 px.

1.2.0 — Kimi Code support

New

Kimi Code added as a first-class agent provider alongside Claude Code, Codex, OpenCode, Gemini and Vibe — select it from the provider picker in Settings.

1.1.4 – 1.1.5 — Freeze-proof: a slow system can no longer hang the UI

A stability release targeting a whole class of freezes: any external call that spawns a process or touches WSL (which can stall when the WSL VM or Docker is wedged) now runs off the main thread, with a hard timeout — so the window stays alive no matter what the system underneath is doing.

Fixed

Frozen UI at startup when probing installed CLIs: on Windows the probe no longer spawns a console process at all (the PATH is scanned directly), removing a conhost allocation that could hang system-wide.
Every wsl.exe invocation (distro list, path translation, CLI detection, file search) now has a timeout and never opens a flashing console window.
Cloning a repository no longer freezes the window for the duration of the clone.
Scanning past conversations of a WSL project moved off the main thread.

1.1.3 — Crash fix on quota banners & a crash black box

Fixed

A hard crash (instant app exit) when parsing a CLI usage-limit banner framed with box-drawing characters — the quota parser now handles multi-byte text safely. This is the parser behind park & auto-resume, so bulk runs no longer die exactly when they hit the token-limit window.

New

Crash black box — if the app ever goes down, the panic message and location are written to panic.log in the config directory, so a crash in the field is diagnosable (and reportable) instead of silent.

1.1.2 — Wall context, folder colors, prose wrap & macOS screenshots

New

macOS screenshot auto-import — a default Cmd+Shift+4 capture writes a file to the Desktop without touching the clipboard; OFM now watches for it, imports it into the image workflow and injects the path into the active agent automatically.

Improved

Mission Control: agent tiles and tabs now show a contextual title (what the agent is working on) instead of a generic label.
File tree: folder icons are tinted yellow, instantly separating folders from files.
Editor: word wrap for prose files (Markdown, plain text) — long lines wrap like in VS Code, while code keeps horizontal scrolling.

1.1.1 — Guided onboarding, brief templates & AI-written briefs

The discovery release: the Plan tab now explains itself, and briefs become reusable assets.

New — Plan / Assistant

First-run onboarding — an empty Plan tab now shows what a campaign is, the 3 steps to launch one, and a big Start with the Assistant button (Advanced one click away). No more guessing where to begin.
Personal template library — save any brief as a named template, then reuse, edit in place or delete it from step 1 of the Assistant. Write a brief once, run it forever.
AI-written briefs — describe the repetitive job in plain words and an interactive CLI session (on your subscription, never the API) writes a conforming brief; the file path fills in automatically when it’s ready.
The downloadable brief template now comes in the IDE language (English, French, Spanish).
Project is optional in bulk mode — without one, the brief’s folder becomes the working directory; AI generation offers a folder picker.

New — IDE

Recursive file search in the explorer, computed natively in one round-trip (instant even on big projects; native find inside WSL).

Improved

Professional icon set — crisp monochrome SVG icons (Lucide-style) across the whole UI, replacing OS-dependent emojis; they follow the light/dark theme automatically, and WSL is marked with a clean text badge.
Mission Control: closing an agent tile now asks for confirmation before killing the session.

1.1.0 — Plan, the bulk orchestrator

The headline release: OFM gains an autonomous engine that runs repetitive jobs to completion.

New — Plan

Drop one repetitive job and a pool of agents runs it in parallel until it’s done.
Bridges the three gaps: crash (relaunch), reboot (resume from the on-disk spool), and token-limit windows (park + auto-resume) — all while staying on your subscription, never the API.
A guided Assistant (3-step wizard) and an Advanced composer; click a campaign card to edit it pre-filled.
Brief format with {unit} / {brief} / {output} / {agent} / {skill} placeholders; four sources (glob, range, file, brief).
Output & done-detection by file or non-empty folder — idempotent and resumable.
Global concurrency pool shared across campaigns, scheduled by priority; quota auto-resume with jitter.
Monitoring: progress/ETA, workers drill-down, quarantine with the captured last screen so you see why a unit failed; one-click retry.

Improved

Flattened navigation: a single top bar — Wall · Plan · IDE · Settings — with the confirmation-queue badge surfaced on Wall.
Reorder projects by drag-and-drop in the IDE rail (order shared with Mission Control).
“Remember for the whole campaign” to auto-resolve a repeated permission/credit prompt across a bulk run.
Delete-campaign confirmation; launch instruction localized to the IDE language.
Light-mode polish across the new surfaces.

Changed

Bundle identifier is now com.openfreemax.ide (see data locations).

1.0.6 — Mission Control & macOS

Mission Control — supervise every running agent on one screen, with a unified confirmation queue and an auto-approval rules engine (levels, hard deny-list, audit, kill-switch).
macOS support alongside Windows, with a small native footprint.
Reliable CLI resolution via a login shell (the right PATH on macOS and WSL2).

1.0.5 — Internationalization

Full UI localization — English (default), French, Spanish — switchable on the fly without killing any session.

Earlier

The core IDE: multi-project workspaces, the file tree, editor and terminal.
Multiple agent sessions per project with bulletproof persistence (resume from the CLI’s own store).
First-class screenshot workflow (paste/drop, thumbnails, path injection).
Native WSL2 project support on Windows.
Support for Claude Code, Codex, OpenCode, Gemini and Vibe.