What’s new in each release.

• The welcome takes you to your first message. Your first project comes right after connecting your agents. If you have none at hand, MADRE makes one to try, and at the end one button opens it with three first messages ready. If you update with no project, Home invites you once to add one.
• Checking a module no longer runs it. To tell you what a module does before you install it, MADRE loads it apart, in a process that cannot write files or open programs.
• The copilot only runs what reads. Tests, builds and saving versions in git wait for your button: they use your project’s code, and the copilot may have changed it.
• The memory respects what you chose. With the memory off or on Ollama, your AIs’ searches no longer go to Google, and INCOGNITO sends and writes down nothing. What you discarded no longer reaches them.
• More fixes. A Market update that asks for more waits for your permission; MCP connectors only let through what MADRE knows; protected folders, protected through links too; and the Windows installer no longer looks suspicious.

See this release in the CHANGELOG

• Windows with a double click. Download MADRE-Instalador.exe from madre.run/en/windows and open it. It turns WSL on if needed, prepares MADRE’s own Linux and installs Node and MADRE inside, with no terminal. If Windows says “Windows protected your PC”, press More info and then Run anyway: the installer has no digital signature yet, and the certificate is on its way.
• Uninstall MADRE uninstalls again. In 1.1.0 the button closed everything and crashed before removing anything.
• The AI on your computer installs on Ubuntu and WSL. Ollama’s installer now needs zstd; MADRE installs it first, or gives you the whole line.
• Two failures MADRE now explains. Codex out of your plan’s usage tells you when it comes back, and Gemini with no account takes you to connect one.
• A plan’s closing, in plain words. MADRE’s internal instruction to close a plan no longer shows in grey.
• More fixes. Sharing memory with a single project opens again, the Map no longer gets lost, and if MADRE changes version while open, the page tells you.

See this release in the CHANGELOG

• Home, everything in one place. Opening MADRE takes you to Home: your projects, the Coordinator and what is waiting for you. Your chats and projects are on the left, like in chat apps.
• Your chats, to one side. Each project keeps its chats in the left column. You can open two side by side, or one in its own window. Files, the browser and the map open on the right.
• The Market. Add features in one click: Notes, Reminders, Preview, Before you push and more. Every module tells you what it does before you install it, and anyone can propose their own.
• A step-by-step welcome. Connect your agents, put a free AI on your computer and get the memory ready, without the terminal. If you come from an earlier version, a short summary of what is new greets you.
• Windows, with WSL2. MADRE runs inside WSL2, the Linux that comes with Windows: wsl --install, open Ubuntu and npx @jossuealcala/madre. Opened on Windows directly, it tells you those three steps. From WSL it finds the local AI you have on Windows.
• Anonymous errors to improve MADRE. Once a day it sends what failed and how much it is used, with nothing you write and nothing from your projects. Turn it off in Settings › Updates and errors.
• Subagents in sight. Claude, OpenCode and Codex may split their work among their own subagents, from #1 to #3 and with your same permission; in #4 only the AI works. While they work you see a chip for each subagent, and at the end «Helped by 2 subagents» with each one's task and report.
• One question to several AIs. Start with several, «@opencode @madre …?», and the question goes to each one: it shows once and each answers on its own. The conversation no longer ends under the open questions or a reminder.
• Running now: what MADRE is doing. A key on the bar of every project and of Home says how many things are running in all your open projects. Open it to see each AI answering, with what you asked, how much of its limit it has used and its subagents; each plan with its steps; the servers and what is installing; what ended a moment ago and the reminders coming up. From there you go to the chat or stop it, one thing or everything, without opening a terminal.
• Search your chats. A magnifier beside «Chats» searches every conversation of the project, by its name and by what was said, accents and capitals aside. A result opens the conversation at that message and lights it for a moment.
• Tested the way people get it. Before release, 1.1.0 was installed with npx on a clean Linux and on a Mac with no sessions, and updated from 1.0.5 with the button. What turned up is fixed: the first message no longer goes to an AI without a session, OpenCode answers again in Home and says at once when its free use runs out, uninstalling also removes the restore points in your projects' git, the Files editor no longer shrinks, subagents report in your language and the terminal speaks yours. The package weighs half as much.
• Uninstall with one button. In Home › Settings › Version, «Uninstall MADRE» first shows what goes and what stays. It closes everything and takes away what MADRE put on your computer: its conversations, memory and settings, its .pulse folder in your projects and its restore points. Your projects and your AIs stay, and so does what the AIs created in .pulse/out, unless you ask for it to go.
• A new license. From 1.1.0, MADRE uses the Business Source License 1.1: free for individuals and teams of up to 10, also to work for clients; larger organizations or anyone offering it as a service need a commercial license. Each version moves to the GNU AGPL v3 after four years. Versions up to 1.0.5 keep their license.

See this release in the CHANGELOG

• A manual you can follow. Real pictures of MADRE with every step marked, from installing it to what to do when something fails. It is in Settings › System › User manual, as a PDF and on this site.
• Tips that teach you in minutes. The first time, MADRE walks you through Home and your first project («1 of 4», «Next»). If you used it before, you only see what is new. Each tip shows once.
• Notes and Reminders. Coloured notes on your project and reminders that tell you on the day and at the hour you said. Ask any agent: «remind me tomorrow at 10 to check the banner».
• Ask about any line. In Code, click a line number and press ASK: explain, improve, write tests or fix.
• Connectors in one click. GitHub, Notion, Linear, Stripe and more, from a catalogue in Settings › Connectors (MCP).
• Cursor and Gemini with the models of your plan. Each one shows the models you have, and Cursor now works with all four permissions.
• Diagnostics understands you. Write what you see, such as «Gemini does not answer», and it finds the case and the fix.
• Clearer and steadier. Plainer words, four new accent colours, less flicker, and closing MADRE from the Terminal closes all your projects.

See this release in the CHANGELOG

• Home's bar fits again. With Market and Modules, the pinned projects shrank to «+5» with no room to move them. Now, when it does not fit, the keys keep their icon and the projects keep their room.
• Your servers open. The ones listening on your whole network (Node, Express or Vite with --host) now open in the preview, inside your computer, and the list says your network can reach them too.
• The state dot always shows. Beside the project's name, when the bar is tight, the name gives way with «…» and the dot stays whole.
• Cursor in the welcome. The «Connect your agents» animation now shows all five.

See this release in the CHANGELOG

• The Market. A full-screen store of modules, in Home and in every project: featured ones with their illustration, shelves and a page per module. Opened from a project, you get it for that project; from Home, for all or only for one. Home also has «Modules», to see where each one is on.
• Modules that arrive on their own. The Market updates from madre.run every time you open MADRE, with a catalog signed by the author. New modules are announced once, and the ones you have update themselves and tell you what changed. The first one to download is «Donate».
• Cursor, the sixth agent. If you have Cursor and its terminal program, MADRE finds it: you write to it with @cursor and it joins plans and round tables. For now it reads and answers.
• Code alongside. The CODE key opens your files next to the conversation: search, tabs, ⌘P to open any file and, for Markdown or HTML, the view alongside.
• What an agent changed. After a #3 EDIT turn, its lines stay in its colour with Keep, Undo and Ask why.
• Gemini and OpenCode without a terminal. Gemini connects by pasting your Google AI Studio key on its card, and OpenCode with GitHub Copilot, with one button. The terminal is still there, under "Prefer your terminal?".
• Local AI sized to your computer. Modules › Local AI reads your machine and tells you which Qwen2.5 size to use and which one you can train, from 0.5B to 72B.
• Projects that share. A connection can share the Continuity state and receive notes, which only enter the memory if you accept them. The Coordinator learns from the projects you mark and can sit an exam.
• Neural mode. The maps are drawn as the neurons of a brain: only the paths something travels light up, with far less work for your computer.
• Your own wallpaper. In Settings › Appearance you choose a photo, frame it, and it becomes the background of Home and of the projects that use it.
• Turn MADRE off. A button at the end of the bar closes every project in order, and each one saves what is pending.
• A chat that holds still. You can scroll up while an agent works without the conversation pulling you back to the end, and long commands can now run.

See this release in the CHANGELOG

• Copilot. Press the robot next to mode · model · effort, choose how long, and the agent you write to changes files, makes plans and runs safe commands without asking you each time. It asks for your permission only once.
• The output comes back on its own. When you run what an agent asked for, the output goes back to it and it carries on with its work.
• Permission, with its time. An agent that needs more asks for the permission and for how long: 5, 15 or 30 minutes, or just this message.
• Outside tools (MCP). In each project's Settings, agents can use other apps and services: a program on your computer or one on the internet. If you already use some in Claude Code, Codex, Gemini or Cursor, you can bring them over.
• A chat that is easier to read. Answers in a bubble with the agent in its pill, attachments as cards, the plan as a task list, numbered choices, commands as steps and each command that ran as a pill.
• Notices as dots. Notices shrink into a small dot after 10 seconds; a click opens them as a tidy list. The ones that need something from you stay open.
• Your files up to date. The file list updates by itself after every answer and every command, and a dot on a folder tells you which agent changed something inside.
• OpenCode with effort. Each OpenCode model offers its own reasoning levels, and adding a provider shows every one a single API key unlocks.
• In your language. The first time, MADRE opens in your browser's language, and the README and the documentation are in English too.

See this release in the CHANGELOG

• Home your way. Right-click any card to move it, change its width and height or take it off; they drag too. The projects drawer shows three rows and every page has its own wallpaper.
• Widgets. System resources, Servers, Memory quality, Coordinator training and Local models, each with real data and only when something feeds it.
• Clearer words. The whole interface was reviewed so it reads without knowing about AI, in Spanish first: Diagnostics speaks of problem and solution, and a memory's states say what they are.
• Local AI by project. In Home › Settings › Local AI you choose Automatic, One by one or Manual. Changing the mode leaves existing projects as they are.
• Connections, a sky. Each project is a star in its own tone and each connection a flow of light; the full map opens in its own window.
• Claude with your API key. When signing in you choose your Claude account or the Anthropic Console, or paste your key; a browser sign-in is finished inside MADRE.
• OpenCode by provider. The picker groups OpenCode's models under their provider, and in Settings › Agents you paste or remove each provider's key.
• A tour for whoever had 1.0. A greeting of its own and only what changed, with guides to the widgets and a card the first time you open each key of the bar.
• Restart all. After updating, Home restarts at once the projects still on the older version and leaves the busy ones for later.
• A daily usage notice. Once a day each copy says it is still in use: a random identifier, the version, the system and the Node version. Nothing about you or your projects, and no IP kept.

See this release in the CHANGELOG

• Home: all your projects in one place. madre opens Home. You add your folders (adding does not open them), open the ones you will use and see what happened in each one, who is waiting and what was decided. The first time, three steps: agents, projects and open.
• The Coordinator. An agent that lives in Home: it reads across your projects, signs connections and carries orders from one to another, up to #2. From any project you ask it with /bishop.
• Connections between projects. One project looks up what another already learned: you sign which kinds of note cross and for how long. Never your files or your conversations, and you revoke them whenever you want.
• Several conversations at once. A second conversation fits beside the first, every panel opens beside you or in a window you move, and the viewer edits code.
• Permissions your way. Modes are called Incognito, Chat only, Create, Edit and Full access. You choose which key unlocks #3 and #4, an accepted key can last a few minutes, and the project itself checks it.
• The maps. NOSTROMO is now the Memory map, with its key on the bar; the Project map, in Home, shows your projects around the Coordinator.
• Settings and Diagnostics. Everything you configure lives in one window with a search (⌘ ,). MU/TH/UR is now Diagnostics: you say what happens and it gives you the remedy.
• Modules as apps. A board of tiles with Token saver, Intents, Local AI, GitHub (/gh) and Continuity, among others.
• A tour at your pace. Newcomers see “Getting to know MADRE”; people coming from 0.7.2 get a tour that starts with what moved, and their projects show up in Home.
• Safer and sturdier. Before closing 1.0 the whole codebase was reviewed with security tests: the project checks the #3 and #4 key, MADRE’s git does not run what an agent leaves configured, what crosses between projects is always cleaned, and nothing is overwritten or corrupted when two changes arrive at once.

See this release in the CHANGELOG

• An agent can ask you to run a command. With the Runs module on, it leaves one button per command, with the reason on top. Nothing runs until you press, a checkpoint is taken first and UNDO reverts whatever changed.
• More permission in one click. If an agent says it needs a higher mode, GRANT AND CONTINUE gives it your permission with the usual designation and the agent carries on.
• Whole options, the recommended one first. Every option reads in full, in numbered rows, and the one the agent recommends comes first.
• Effort beside the model. You choose how hard each agent thinks before answering, from Low to Max in Claude and from Minimal to Very high in Codex. It is remembered per agent.
• Files that behave as you expect. Copy, move and create no longer collapse the tree, pasting next to a file with the same name makes a copy, and the panel is wider.
• Fixes. Updating Ollama now changes the version that answers, and reopening the room no longer shows an old sign-in.

See this release in the CHANGELOG

• Codex in a project without git. If the project folder was not a git repository, Codex failed on every message even when signed in. It now answers anyway, limited to the mode the room chose.
• The error says its cause. When an agent ends with an error, the message carries the last line it printed.

See this release in the CHANGELOG

• NOSTROMO in three dimensions. The archive is shaped like a brain and you turn it by dragging; Shift + drag moves the camera and RECENTER returns to the side view. What is near looks bigger, and what sits behind MOTHER is drawn behind her.
• Every memory is a star. Its brightness says how much the room relies on it; its surface, whether you confirmed it, nobody has judged it, it turned out false, or it is cooling. At the centre, MOTHER is magma, with the odd explosion.
• Choosing a memory brings it forward. The archive turns until it faces you, everything else darkens and a green sight closes on it: TARGET LOCKED. Forgetting a memory is an implosion, and its links crumble into dust.
• MEMORY is a dashboard. It has its own button in MU/TH/UR. It shows the archive’s figures against the week before, 30 days of activity, what it is made of and the three tests. It reads without the project designation: these are counts.
• What’s new after an update, and a tour over the room. After an update, a sheet tells you the main points and where they live; ABOUT THIS VERSION opens it again. The first time, a four-stop tour lights up the crew, the modes, MU/TH/UR and the modules.
• Ideas for a module of your own. /module offers eight ideas ready to edit, each with its icon, and typing /modules or /modulo now finds it. Connectors keep their key in a vault for your user only, and the SDK documents its 36 fields with an example that works as is.
• Stop one reply, not the room. STOP, next to the turn’s seconds, halts a single agent. The room, the other agents and other plans carry on; STOP ALL is still the master brake.
• An outside page can no longer get in. The room only answers when it is called by this computer’s name, and anything that changes something has to come from the room itself. The terminal, your agents and MADRE’s CLI get in as before.
• Your thumb moves the memory. Rating a reply changes where what the room remembered from it stands, without rewriting a word. There are four ways to judge: besides thumbs up and down, “this is how I want it” and “never again”.
• And more. The text box grows up to fourteen lines, the light theme reads again, the economy speaks without abbreviations, and Gemini no longer leaves a turn hanging for three minutes or claims files it never wrote.

See this release in the CHANGELOG

• RIPLEY, a browser next to the chat. It opens the servers answering on your machine in tabs. It only accepts local addresses and runs nothing to find them. If a site refuses to be framed, it offers to open it in another tab.
• The economy, in view. Every message says at its foot how many tokens were not charged. The ECONOMY panel shows input, output, cache and a table per agent, without promising a savings percentage.
• Ash tells you what it costs. Asking for compact replies has a cost per turn, and Ash shows it before the savings. The choice you make when you send is kept for the whole turn.
• What the room knows, where you decide. Current decisions appear above the chat and open questions under the text box. If you ask about something already discussed, @madre answers from your machine, with no tokens.
• Choices: options as buttons. An agent with three or more real options gives them to you as buttons. Pressing one types the option and stops there: nothing runs until you send.
• Permission, next to whoever asks. If a reply says it needs more permission, the button to raise it appears under it. For #3 and #4 it still asks for the project designation.
• Files with a right click. Copy, move, rename, new file and new folder from the panel, plus a search by name. What is protected, such as .git or a .env, asks for the project designation.
• A more compact bar and a hidden passenger. The bar makes room when the browser is open and the footer shows the version. And if you type madre alone in the box, a ship arrives that is not an agent.
• The fixes. Memory chips read in the light theme, the / palette shows every description, the page can no longer fail on its first load, and PLAYWRIGHT opens the room again.

See this release in the CHANGELOG

• Module commands from the start. /git, /ahp and the other module commands answered “unknown command” when the room opened, even with the module on. Now they work as soon as it loads, and update when you install, remove, turn on or turn off a module.
• The history shows everything. After a reload, the CREATE strips of earlier turns didn’t appear, and neither did module installs and removals. Now they do.
• Sentences written properly. The Spanish room no longer says “@CODEX PUEDE CREA”: it builds the full sentence, says “por ti” instead of “por you”, and the /module guide explains what to put as the author.
• Tests that check everything. The test that draws the history covered 8 of the 61 event types the room can show. Now it covers all of them, and on its first run it found a fifth bug.
• Releases that aren’t rewritten. Each published number maps to exactly one code: fixes that arrive later go in the next release, never inside one already published.

See this release in the CHANGELOG

• Coloured code, in the viewer. JavaScript, JSON, CSS, HTML, Markdown, Python, Shell, config and SQL are highlighted, with nothing to install. What you see is exactly the same characters that are on disk.
• You know who touched what. In the file tree, a dot in each agent’s colour marks the files it wrote in this room.
• Select and ask. Click a line number (Shift for a range) and REVIEW WITH sends it to the agent you pick. Only the reference goes in: you write the question.
• It warns you if your messages go through another server. If your machine has an intermediate server configured for OpenAI or Anthropic, MADRE shows it and holds those turns until you allow it or remove it.
• /module to build modules. Type /module and what you need; the agent writes the module and MADRE shows you the card to install it. It needs the CREATE permission, and typing the command never turns it on for you.
• Modules are easier to write. MADRE warns you when a module has a misspelled field, each module of yours uninstalls from its own card, and the guide shows the right way to write one.
• Interface details. Switching conversations no longer reloads the page, the project name never gets cut off in the bar, and the cursor no longer lands inside references.
• The 0.5.1 fixes. The viewer no longer mistakes regular expressions for comments, the Spanish room mixes in less English and the sentinel no longer sends warnings MU/TH/UR already explains.

See 0.5.1 in the CHANGELOG See 0.5.0 in the CHANGELOG

• Report without an account. A form on madre.run that asks for no account, and a link to GitHub for those who prefer it.
• A code to ask for help. Every known problem has its code (MU-001 to MU-054), and the log downloads as a file any spreadsheet opens.
• Fixes with a button. The instructions are in Spanish and eight problems are solved with one click in CONNECTIONS, with no commands to type.
• Errors that explain themselves. Instead of “code 143”, MADRE tells you something stopped the agent, and keeps the cause apart from what the program printed.
• Alerts you can close. Dismiss each alert, or all at once, without losing the history.
• Fixes. Image Studio finds your Gemini key, PLAYWRIGHT connects again and the Spanish interface is complete.

See this release in the CHANGELOG

• Read what is sent. From NOSTROMO you see the full document each agent receives and the exact command it is launched with. Your keys only appear by name.
• Every connection, in view. The eight addresses MADRE can connect to, what each one carries and how to switch it off.
• Many conversations, one memory. Open separate threads on the same project: they all feed the same memory.
• The memory explains itself. Every answer says which memory notes it used.
• EYECAT looks after the memory. It spots notes that contradict each other; what you mark as false stops being used.
• Your data, hidden. Keys, tokens, emails and personal paths are hidden before they are stored.
• Agents ready from the room. Install agents and sign them in without opening the terminal, and update modules with a click.
• In Spanish. More than a thousand of the room’s texts in Spanish, with English one button away.

See this release in the CHANGELOG

• AIRLOCK (#4). Everything in CONTROL plus commands: tests, builds, git push and deploys. It turns on with two keys: the project name and the word AIRLOCK.
• Extra tools per turn. A module can give an agent additional tools for its turn only. The first is PLAYWRIGHT, a browser only your MADRE can use.
• Git from the room. With Git Pulse you commit and push from the room; no agent can do it for you.
• Update with one click. MADRE installs the new version and reopens the room at the same address, never while an agent is working.

See this release in the CHANGELOG

• New files in the right place. With CREATE, agents create files where the project’s structure says; if they touch something that already existed, MADRE restores it and tells you.
• One permission per step. In a plan, each step can ask for its own level, always within the limit you set.
• UNDO without git. If your folder doesn’t use git, MADRE keeps its copies separately, outside the project.
• Fixes. OpenCode now writes in CREATE and CONTROL, and UNDO also reaches repositories that live inside another one.

See this release in the CHANGELOG

• New-version alert. Once a day it asks npm whether there is an update; only the package name travels. It switches off with a toggle.
• Private names. A list of terms, such as your company or a client, that MADRE replaces before storing or sending them.
• The round table. You ask @madre to consult the crew; @madre asks each agent, one after another, and closes with a cited summary.
• Rate answers. Mark each answer as good or bad; the bad ones stay out of the examples you can train your own model on.

See this release in the CHANGELOG

• All local with Ollama. Search by meaning and the memory summary run on your machine, free and with nothing leaving. Without Ollama, everything stays the same.
• @madre. It answers from the whole memory, with citations; it tells you when something was never discussed and never changes files.
• Modules. The SDK arrives: a module is one file with its settings, its switch and its commands.
• Protected zones. While an agent edits in CONTROL, your .env files and MADRE’s folders stay read-only.
• Your own model. Export the room’s conversations as training examples, with the recipe included.

See this release in the CHANGELOG

• Pages that work. Scripts run and styles, images and fonts load, with no internet and no access to MADRE.
• It reloads itself. When an agent changes the open page, the viewer updates.
• Errors in view. Page failures show up in a strip with a button to ask the room.

See this release in the CHANGELOG

See this release in the CHANGELOG

• Error reports, with your permission. Failures are kept without paths, names, emails or keys, and are only sent if you decide so.
• Copy and reply. Copy any answer or pick which agent replies, with the quote up front.
• Tested on every change. Every change is tested automatically on macOS and Linux, with Node 22 and 24.
• More accurate wording. The room used to say nothing left your machine. Agents do talk to their providers, and now the room says so.

See this release in the CHANGELOG

• Four permission levels. GHOST, EXCHANGE, CREATE and CONTROL, with a copy saved first, a change list and UNDO. CONTROL asks for the project name.
• The room’s memory. What is said is kept and each turn receives what relates to the question; an archivist sums up notes that last.
• NOSTROMO. The map of the memory, protected by the project name. The only thing you can do there is forget.
• MU/TH/UR. 36 known problems, each with its fix for your system.

See this release in the CHANGELOG

See this release in the CHANGELOG