Since 0.1.100 the app downloads new releases on its own and asks whether to update now or on the next start. On 0.1.94 or older, run the new installer once over it — updates are automatic after that.
Notes use the full width of a wide card, Claude Code cards wear a new “CC” tile, and two Claude Code fixes: a late “finished” no longer ends your next turn, and Claude Code on non-Anthropic OpenRouter models works. Windows only in this release.
Improved2
Wide notes fill the card: a wide note card’s text no longer stays in a narrow column in the middle — resize the card and the note follows. Suggested on the ideas board.
Claude Code cards show an orange “CC” tile everywhere the agent CLI’s icon appears — the sidebar, card headers and the add-card menu.
Fixed2
A late “finished” no longer ends your next turn: when Claude Code reports a turn as finished late, the next turn you already started stays working.
Claude Code on OpenRouter models that are not Anthropic’s gets plain Messages requests, so those models answer instead of failing on Claude-only features.
Antigravity CLI, Google’s successor to Gemini CLI, joins as a full agent CLI in Gemini CLI’s place, and cards now run OpenCode 2 next to 1.x — with dangerous mode that switches live. Gemini CLI cards you already have keep working and move to Antigravity in one press. Windows only in this release.
New2
Antigravity CLI: Google’s `agy` as a full agent CLI. Working, needs-input with the command or question it waits on, and finished from its own hooks; conversations that resume after a restart; the app’s MCP tools by arrow, refreshed live, with connected MCP servers asked on every call; canvas mode; Escape to interrupt; Usage, Run Stats and Agent Pulse with subagents; a context chip and a quota chip; model and CLI switch, templates, OpenRouter and your own providers, and Caveman, Memory, Context7 and Graphify through its hooks. Each card has its own Antigravity folder next to your ~/.gemini, which is never written. Dangerous mode applies when the card restarts; Compact, RTK and WSL/SSH workspaces are not available yet. Install it from the setup wizard.
OpenCode 2: cards run OpenCode 2 (`@opencode/cli`) as well as 1.x — the version is read when the card starts, and 1.x runs exactly as before. On 2.x dangerous mode switches live, without a restart; the app picks the session id before the first turn; the app’s tools appear live as you draw arrows; a model change applies to the same session; and OpenRouter requests carry NeuroSquad’s attribution also on your own OpenRouter login. Usage, Run Stats and Agent Pulse read 2.x sessions, subagents and compaction included.
Improved1
Gemini CLI cards: new cards use Antigravity CLI. Cards you already have keep running, labelled “Gemini CLI (legacy)”, with Switch to Antigravity — one press moves the card over with its conversation’s context, after the running turn if there is one. Templates with Gemini CLI cards import them as Antigravity CLI cards.
Three new agent CLIs — MiMo Code, PenguinHarness and ZCode — with the same depth as the rest: statuses from their own events, sessions that resume, the app’s tools by arrow, usage in Usage, Run Stats and Agent Pulse, OpenRouter and your own providers, and a switch to and from any other CLI. NeuroSquad now runs 22 agent CLIs. Windows only in this release.
New3
MiMo Code (Xiaomi’s terminal coding agent, a fork of OpenCode) as a full agent CLI: working, needs-input and finished from its own events, sessions that resume after a restart, the app’s MCP tools and connected MCP servers by arrow, canvas mode, interrupt with Escape, and dangerous mode that switches live, without a restart. Its in-session subagents (“actors”) keep the card working and are counted once in Usage, Run Stats and Agent Pulse. The card starts without MiMo’s workspace-trust question, and a session MiMo no longer has opens a fresh one instead of a blank screen. Install it from the setup wizard.
PenguinHarness (penguin.ooo, open source) as a full agent CLI: NeuroSquad starts its own Penguin server for its own data folder and gives every card its own Penguin agent, so your ~/.penguin is only read — the models and keys you set up there are used as they are. Statuses with the text of what it asks, approval from the card, dangerous mode that switches live, canvas mode, interrupt that stops the task on the server, sessions that resume, usage with subagents marked. Caveman, Memory, Context7 and Graphify reach Penguin on releases after 0.2.13 (their tools by arrow work on 0.2.13 too); on 0.2.13 a model change starts a new session that points to the old one.
ZCode (Z.ai’s agent for its GLM models, the `zcode` command) as a full agent CLI: statuses, questions and plans from its hooks, sessions that resume, the app’s MCP tools by arrow, canvas mode, Escape to interrupt, usage from its own ledger with subagents marked, OpenRouter and your own providers. Each card has its own settings, session database and provider file next to your ~/.zcode, which is only read; tokens and keys never land on disk and are hidden from the agent’s shell commands. Dangerous mode restarts the card. ZCode’s automatic project memory stays off in cards unless you turned it on yourself, because its model calls are not recorded anywhere Usage could count them.
A fixes release. A prompt sent to an agent that is waiting at a question now waits in its queue instead of becoming the answer, agents from an imported template start with dangerous mode off, Memory’s smart search no longer holds up the app, and SSH, WSL, Graphify, Code Graph, recording and the phone grid get a batch of fixes. Windows only in this release.
Improved1
Templates: agents from an imported template start with dangerous mode off — tick it per card in the import dialog. An arrow from a dangerous-mode agent to an MCP server you already have is held back and named instead of drawn silently. The dialog now lists card settings that reach beyond the template (Memory shared by every workspace becomes this workspace’s own, auto-save and auto-recall, the fact model, Graphify mapping files your .gitignore leaves out) and the workspace scripts it brings; exporting and publishing remind you that scripts go out exactly as written.
Fixed6
Prompts: a prompt from another agent, a note, Telegram or the queue is never typed into an agent that is waiting at a question — the text would have become its answer; it waits in the queue until the turn ends, and Telegram tells you so. Two prompts sent at the same moment arrive as two messages, not glued into one. A multi-line first prompt is pasted whole, so no line break sends part of it.
Agents and MCP: an agent can remove only cards it created (or a helper it started did) — never your notes, to-do lists or boards. A model or CLI change an agent asked for is checked again when its turn comes and never switches dangerous mode on. An agent has at most one install or permission question open at a time, and Allow applies only to the MCP server you saw. A deleted card’s MCP sessions are closed, and oversized requests to the app’s local MCP server are refused.
Plugins: Memory’s smart search runs its model off the app’s main thread and keeps memories ready in memory, so auto-recall on every prompt no longer stalls the app; nothing is saved for an agent deleted mid-turn. Graphify replaces a process that stopped answering, keeps a new card’s graph when a deleted card on the same folder is cleaned up, and a late search result no longer lights up the map for an older query. Code Graph closes its engine when you cancel while it starts and never deletes a download in progress when it cleans up old versions. RTK looks for its binary along the PATH once a minute, not before every command. Web watch stops a pattern that takes too long instead of hanging.
SSH and WSL: when a server refuses another session (sshd’s MaxSessions), the card says so instead of “not installed”; short commands share a few channels, and opening a channel times out instead of hanging. HostKeyAlias and ProxyCommand tokens work as in OpenSSH, editing a host’s label or key file no longer cuts its cards, and after a disconnect or a cancelled password nothing reconnects in the background. Transcripts of long sessions catch up in one go, and multi-line scripts work with csh/tcsh. In WSL, files carrying a card’s token are copied with private permissions and removed with the card.
Grid and phone: keys typed in a grid tile no longer move or select the hidden canvas card, a grid layout change that changed nothing is no longer an undo step, and on a phone the full-screen agent you opened from “Needs input” stays open after you answer it.
Other fixes: workspace recording loses at most one frame when a disk write fails and no longer piles up writes on a slow disk; a scheduled prompt held by the budget brake goes out once the brake is lifted, a refused one is retried, and a time that passed while the app was closed doesn’t fire late; a data file that can’t be read is kept as a .corrupt copy instead of being emptied, and an interrupted save is recovered from its backup; Codex on your own servers reads streamed answers split across lines; a worktree setup script that never exits no longer blocks the agent’s start; an unusable dictation hotkey falls back to the default.
Remote access now adapts to the device: phones and tablets get a grid of your agents made for touch, with a prompt box, a full-screen terminal with the keys a phone keyboard lacks, and status filters; a computer’s browser gets Canvas and Grid like the desktop. Agents in dangerous mode can ask for an MCP server, and you allow it in one dialog. Run Stats and Agent Pulse show what subagents spent. And the app writes to disk less and starts many cards more smoothly. Windows only in this release.
New2
Remote access on any device: the app tells a phone, a tablet and a computer’s browser apart, and you can pin the layout yourself (Automatic, Mobile layout, Desktop layout). On a phone a workspace opens as a grid — one agent per screen, two or three side by side on a landscape phone or a tablet — with a workspace switcher, filters by status (needs input, working, done) and a switch to the canvas. Tap an agent and type in the box at the bottom (autocorrect and the voice keyboard work); Stop interrupts the turn the agent’s own way, and 1 / 2 / 3 answer its question. Full screen adds a row of keys — Esc, Tab, Shift+Tab, arrows, Enter, and Ctrl+C where it is safe — and a swipe moves to the next agent. Only the agents on screen hold a live terminal, so the phone stays light; the rest keep their last lines. A computer’s browser gets the app as on the desktop, now with Canvas/Grid. Grid on a phone or in a browser never changes what the desktop shows.
Dangerous mode and MCP servers: an agent in dangerous mode can now ask for an MCP server instead of stopping at “only you can draw that arrow”. You get one dialog with the agent, the server, its version, how it starts and the tools it would get; Allow connects it as if you had drawn the arrow, Deny is remembered until the app restarts. A server that isn’t installed yet asks once: Install and allow.
Improved2
Run Stats and Agent Pulse: Run Stats 1.1 in the verified catalog splits a run into the main agent and its subagents (requests, input, output, cache, total, with a share bar, also in the MD and JSON exports); Agent Pulse 1.2 draws subagent requests on their own lane with the subagent’s name in the tooltip. Both need NeuroSquad 0.1.270 or later.
Performance: settings, agents and layouts are written to disk only when something changed; terminal output is drawn in step with the screen; when a workspace with many cards opens, their terminals start one after another instead of all at once; dialogs load when first opened; and cards poll only while the window is visible.
Fixed1
Agent CLIs: more reliable ready detection for Cline, Droid, Kilo Code, Kimi Code and Qwen Code, so the first prompt arrives at the right moment; session reading fixes for Crush, Cursor and Gemini CLI; and Usage reads only the new part of the logs instead of scanning them again.
Agents can now equip themselves: an agent looks up the plugins, skills and MCP servers it could have and connects the one its task needs to its own card — anything new is installed only after you approve it. Cards stay “working” while their subagents run, and subagent requests are counted once in Usage and Run Stats, with their own share shown. A two-column add-card menu with everything in view, and a large batch of fixes. Windows only in this release.
New2
Agents equip themselves: two new tools, neurosquad_catalog and neurosquad_equip, let an agent see which plugins, skills and MCP servers it could use (with what each one is for and whether it works for that agent CLI) and add one to its own card — the card appears beside the agent with an arrow, or an arrow goes to the matching card already on the canvas. Plugins and already-installed skills and servers connect at once; a skill or server that isn’t installed yet is installed only after you press Install and connect in a dialog that shows its package, source, version, pinned hash and security scans — asked even in dangerous mode. A notice says what an agent connected, with Undo. At the start of a session agents get a short list of what is available (Claude Code, Codex, Qwen Code, OpenCode, Kilo Code, pi, omp, Gemini CLI). The workspace setting “Agents can add plugins, skills and MCP to themselves” turns it off.
Subagents: a card stays “working” while the subagents its agent started are still running, and finishes only when the whole turn is over — no early “done” sound while background subagents work. A chip in the card’s header shows how many run, with their types in the tooltip. Claude Code, OpenCode, Kilo Code and Qwen Code report them; a subagent’s permission question is shown on its parent card with the subagent’s name. Their model requests are counted once — in Usage, Run Stats and Agent Pulse — and Usage shows “incl. N subagents” on the agent’s row. For card authors, Card SDK 1.3 adds the subagent share to agents.usage and marks subagent requests in agents.timeline.
Improved3
Add-card menu in two columns: agents on the left (installed CLIs first, then your saved agent templates), cards on the right, Skills, MCP and Plugins on top — everything visible at once, no folded groups. The row under the pointer is explained next to the search field, and the arrow keys move between and within the columns.
Budget: Resume on a workspace that went over its limit now lets its agents spend about another tenth of the limit and then pauses them again; the Budget card says where. To keep going longer, raise the limit.
Browser cards: deleting a browser card now deletes its profile — cookies, logins and history — instead of leaving it on disk. A browser that was still starting no longer stays running after you quit or sign out.
Fixed7
Agent status: an agent’s status no longer jumps when its CLI was restarted and the old process reports late; a Claude Code prompt waiting in the CLI’s own queue no longer ends in a false “finished”; a transcript Windows briefly locks is retried instead of being given up on. Sounds and notices come only from the card itself — a pop-out window or the phone no longer plays a second sound — and all of an agent’s notices close as soon as it works again.
Terminals: your own Claude Code settings in the environment (such as CLAUDE_CODE_GIT_BASH_PATH or CLAUDE_CODE_USE_BEDROCK) now reach Claude Code cards — only the markers of a parent Claude Code session are removed. A line break inside a pasted text is no longer taken for Enter.
Plugins: Memory searches no longer wait behind a memory being saved, and memories of a deleted workspace or agent don’t come back after switching the search model. Memory and Context7 share one short time limit on each prompt, so they can’t hold up the agent. Caveman sends its full rules again if the agent never received them. RTK rewrites a command with exactly the rtk the agent’s shell will run and allows fewer commands on its own outside dangerous mode.
Usage: Kilo Code’s subagent cost is no longer counted twice, a forked Codex subagent no longer repeats its parent’s tokens, Hermes Agent’s child sessions are credited to the card that started them, and pi and omp requests on your own servers show “no price” everywhere (Usage, cards, budget) instead of $0.
SSH and WSL: host keys are checked the way OpenSSH does (a server is asked for the key type already known for it first), each ProxyJump hop uses its own SSH config, cancelling a connection while it connects really stops it, and fish or tcsh login shells work. WSL workspaces pick up a changed network after wsl --shutdown, and two files with the same name no longer overwrite each other.
Dictation: model files are downloaded from a fixed version and checked against a pinned SHA-256, a resumed download included. The hotkey can be a single key such as F9.
Account and remote access: sign-in codes and requests from the app are tied to the session that asked for them and work only once. The token of a named tunnel for phone access is never shown back to a window or the phone, and the phone reconnects by itself after the desktop restarts or updates.
Lead agents can move the agents they lead to another model, provider or agent CLI and keep the conversation, prompts sent between agents are checked until they really go out, and AI cards on Windows run on a newer console so typed and pasted text arrives whole. Plus fixes for Codex on your own servers, workspace recording, Run Stats and Graphify. Windows only in this release.
New1
Agents change agents’ models: a lead agent can move an agent it leads by an arrow — or itself — to another model or provider (the CLI’s own subscription, OpenRouter or one of your servers) or to another agent CLI altogether, with the same session, so the conversation is kept. A change waits for the current turn to end unless the lead asks for it right away; then the turn is interrupted first. A notice in the corner says who changed which card to what (a waiting change has Cancel), and the arrow’s log keeps it. neurosquad_models now lists your own servers with their models and which CLIs can use them — never their keys or addresses. Only along an arrow, nothing while the workspace budget is paused, and the workspace setting “Agents may change models” turns it off.
Improved2
Prompts between agents are checked: agent_send, a note’s Send to connected agents, the harness switch and the prompt queue type the prompt, wait until the CLI has taken it in, press Enter and then check that a turn really started. A prompt left sitting in the input box gets Enter again (never retyped, so it can’t go twice), and if it still isn’t sent the sender is told so instead of it waiting there silently. An agent busy with a turn gets the prompt through its queue when the turn ends.
Newer console on Windows: AI agent cards now run on a newer console host than the one built into Windows 10. With the old one, Codex CLI lost characters missing from the active keyboard layout (such as [ and ] under a Russian one) and ’ or —, and a long paste into Claude Code could arrive in pieces and never be sent; both now arrive whole. Shell cards keep the system console.
Fixed6
Codex CLI on your own servers: a reply the server cut off at its length limit or content filter is reported as incomplete, and a stream that broke off as failed, instead of looking finished; several tool calls in one answer no longer merge into one; an oversized request gets a clear error; and a Codex card on your server refuses to start while the local translator is down instead of falling back to your OpenAI login. Refreshing a server’s models no longer overwrites a change you made meanwhile.
Workspace recording: a recording keeps only what belongs to its workspace (never the remote-access token or tunnel address) and masks keys and tokens it finds in what it writes; the player and the Recordings list are proper dialogs, so Delete, Ctrl+Z and Space no longer reach the live canvas behind them; dragging the timeline back no longer piles up redraws, and a long recording opens without reading the whole file at once; a failed Stop no longer leaves the Recording chip on; deleting a workspace stops its recording; one unreadable recording no longer empties the whole list.
Run Stats and Agent Pulse: a working stretch cut short by a crash ends where the app last saw it instead of running on; Hermes Agent and Droid usage is no longer counted twice across adjacent time windows; after a harness switch each CLI’s usage is measured on its own.
Graphify: no re-indexing after you quit or delete the card, a failed or cancelled rebuild is reported as an error, and cancelling the install stops it completely. Code Graph also stops re-indexing after its card is deleted.
Templates: an unverified custom card that isn’t installed yet is placed without the template’s arrows, so it gets no access before you install it; a Budget card in a template can only add or tighten a limit — it never unpauses a paused workspace — and Undo puts the previous limit back.
Notes and Grid: a note’s Send to connected agents sends the text as it is in the editor at that moment, to exactly the agents the dialog listed, and “nothing sent” is no longer shown as success. Ctrl+Shift+G no longer switches to Grid while you type in a dialog or text field.
Record a workspace and replay it later on a timeline, exact numbers for every agent run with the official Run Stats and Agent Pulse cards, a note that sends itself to its agents, and Codex CLI on local servers that only speak Chat Completions. Windows only in this release.
New3
Workspace recording: Canvas tools → Start recording keeps everything the workspace shows — every byte each terminal printed, the canvas with its frames and arrows, notes and to-do lists, browser card frames and pictures of other cards — while a red Recording chip in the title bar shows the time. Canvas tools → Recordings… replays it in a read-only copy of the canvas: terminals byte for byte, Play/Pause, 1× to 64×, a timeline with ticks for prompts sent, finishes and “needs your input”, previous/next marker, ←/→ for 5 seconds, and Clean view (no controls, the size of your live canvas) for a screen recorder. Nothing runs and no model is asked during a replay. Recordings stay on this computer, include every key typed into a terminal, and a recording cut off by a crash plays up to where it stopped.
Run Stats and Agent Pulse 1.1: two official verified cards for measuring agents. Run Stats shows one agent’s run in exact numbers — prompts, model requests, input, output and cache tokens, elapsed and working time, cost — freezes them at the finish and copies them as Markdown or JSON. Agent Pulse’s new Requests view draws every model request of the connected agents as a bar on one time axis. A number a harness or server does not report says “not reported”, never 0, and an unknown model has no price, never $0. Both are built on Card SDK 1.2 (agents.usage, agents.timeline), so community cards can show the same data.
Send to connected agents: a note’s new button sends its text as one prompt to every AI agent connected to it by an arrow, at the same moment — the same task for several agents in one click. Agents that aren’t running or are over the workspace budget are skipped and named.
Improved3
Codex CLI on your own servers: it now works on servers that only speak Chat Completions (LM Studio, llama.cpp, vLLM, Ollama and many remote APIs) through a small translator NeuroSquad runs on your computer. The server’s key stays in the app and token counts are the server’s own.
Context window: when your own server’s model list says what context window it serves, Claude Code, Codex CLI and OpenCode are told, so long sessions compact in time instead of ending in an error past the window.
Custom cards: a card whose frame crashed (for example when the computer ran out of memory) reloads itself instead of staying grey; after three crashes in ten minutes it shows Reload.
Fixed1
A workspace just created from a template could lose its layout and all its arrows the first time it was opened; it now opens exactly as the template laid it out.
Switch a card to another agent CLI and keep the conversation, add a template to a workspace you already have, and Graphify, a live map of your code that agents query instead of grepping. Windows only in this release.
New3
Switch harness: the card menu’s Switch harness… moves an agent card from one AI harness to another — Claude Code to OpenCode, or any to any — in the same card and terminal. Its place, arrows, frame, name and notes stay. The conversation goes over as the new harness’s first message: prompts, answers, edited files and the to-do list, taken from the session’s transcript (no model is asked to summarise it, so it costs nothing), sent automatically or only pasted for you to review. Switch back and the earlier harness resumes its own session, with only what happened since it left. A card in the middle of a turn can be interrupted and switched at once or switched when the turn ends; dangerous mode, canvas mode and the provider come back when you return to a harness that supports them, and usage of the earlier sessions stays on the card.
Templates into an existing workspace: the import dialog has New workspace / Add to existing, and Canvas tools → Insert template… or Template… in the add-card menu opens it for the workspace you are in. Cards the workspace already has — Squad status, Budget, Code Graph, RTK, and Memory, Caveman or Context7 with the same settings, an MCP or skill card for the same server or skill — are reused instead of doubled. The new cards arrive together beside yours, in a frame named after the template, the camera flies to them, first prompts are pasted only into the new agents, and nothing already on the canvas is restarted. The toast’s Undo removes exactly what was added; installed components stay installed.
Graphify plugin: an arrow from an agent to a Graphify card gives it graphify_* tools — who calls a function, what a change affects, how two parts connect, what a subsystem is made of — answered from a knowledge graph of the workspace built by the open-source graphify (Apache-2.0). The card draws that graph and plays every query on it live: the starting points ripple, the search spreads hop by hop, the result stays lit, with file:line links that open in your editor and a timeline that replays past queries. Installed on request into the app’s own folder (a private Python, about 250 MB, every download checked against a pinned checksum); the whole project is mapped on first view with progress, then kept up to date as files change. Agents get a short note that the graph is there, and Claude Code a one-line hint when it greps for a symbol — the search still runs. No language model and no API keys involved, nothing written into the project; works in local and WSL workspaces, not over SSH.
Improved2
Marketplace: templates made by the NeuroSquad team carry an Official badge, and like and download counters are hidden while they are still zero.
Canvas mode’s tooltip on a card now says what actually happens for that agent CLI: its own shell and web tools are off from the next launch, and for Cursor CLI they are blocked right away.
Templates: save a workspace as a template, share it as a file or through the new marketplace, and import someone else’s with one consent that installs what is missing. Also: Arrange by connections, and Node.js installed for you when a machine has none. Windows only in this release.
New4
Templates: Canvas tools → Save as template… (or the canvas menu for selected cards) keeps a workspace’s shape and settings — cards, frames, arrows, agents with their models and modes, built-in card settings, plugins, skills and MCP servers — as one JSON file or in your library. Never included: API keys and tokens, folder paths, chats and transcripts, memories, browser data, Telegram chats; card content only where you tick “Include content”. Anything that looks like a key, token or a path on your computer is flagged in the dialog before you can save — nothing is removed silently. Import a file, a marketplace template or a neurosquad://template link: one dialog lists exactly what will be installed (agent CLIs with their install command, MCP servers from the official registry with their launch line, skills from skills.sh pinned to the version in the template, verified custom cards), which agents run in dangerous mode (switch it off per card), the workspace scripts and the first prompts — then installs it all after one confirmation and opens the new workspace on this computer, in WSL or over SSH.
Templates marketplace on neurosquad.ai/templates and in the app (Templates → Marketplace): publish a template from your library, a moderator checks every version — prompts, scripts, dangerous mode, the servers it installs — and it appears for everyone. Search, sort by popular, new or trending, filter by category, like with your NeuroSquad account. Only the author can update a template: save a new version over it and publish the update; the public version stays until the update is approved. “Open in NeuroSquad” on the website opens the import in the app.
Arrange by connections (canvas controls, bottom left): every group of connected cards becomes a cluster around its lead agent, with the cards of each sub-agent fanned out behind it, real card sizes and no overlaps; cards without arrows go in a grid below. Frames are arranged as units. One Undo — the toast’s button or Ctrl+Z — puts everything back. Cards keep their size and the camera never zooms in.
Node.js for machines without it: the agent CLIs install through npm, so when npm is missing the setup wizard installs the current Node.js LTS from nodejs.org itself — the official archive, checked against its published SHA-256 (no checksum, no install), for your account only, without administrator rights. The Windows installer does the same on a fresh install (“Skip Node.js” skips just that step; --no-node turns it off). A Node.js already on PATH is never touched, and the wizard says how to undo it.
Improved2
Workspace templates saved with earlier versions appear in the new library, marked Legacy, and create workspaces as before.
The old grid layout stays as Tidy grid next to Arrange by connections, now also one undo step with an Undo toast.
Two new plugins for agents: Code Graph, a map of your code they query instead of grepping, and Context7, current library docs instead of what the model remembers. Both — and Memory — also work in WSL and SSH workspaces.
New2
Code Graph plugin: draw an arrow from an agent to a Code Graph card and it gets 12 codegraph_* tools — find a function, trace who calls it, read a snippet, see the architecture — answered from a knowledge graph of the workspace’s code instead of grep-and-read loops. Built on the open-source codebase-memory-mcp engine: a one-time download of about 40 MB (about 300 MB unpacked), checked against a pinned SHA-256; it runs on your computer with no account and no telemetry, and your code never leaves the machine. The graph updates itself a few seconds after files change. The card shows symbols, files and calls, what the code is made of, a name search and which agent asked what.
Context7 plugin: an arrow gives an agent two tools that fetch current, version-specific documentation and code examples for a library from Context7, so it stops writing against APIs that have changed. No key needed for 200 requests a month; a free Context7 key raises that and is stored encrypted. Answers are cached for 24 hours. The card shows the requests left, every lookup and which agent made it, and lets you try a lookup yourself. Optional auto-docs points an agent to the docs when your prompt names a dependency of the project. Library names and questions are sent to Context7 (Upstash), and the card says so.
Improved2
Code Graph, Memory and Context7 work in WSL and SSH workspaces. Code Graph runs its engine where the code is — inside the distro or on the host — after copying the verified archive there once; its data stays in the app’s own folder on that machine. Remote hosts need Linux on x86-64 or ARM64. Memory and Context7 stay on your computer and reach the agent through its arrow, auto-docs included.
Plugin cards polished: overview tiles no longer clip the unit under the figure, the Code Graph tile shows downloading and failed states correctly, the Context7 card no longer says “No lookups yet” above a list of lookups after a restart, the Memory card’s Smart search offer takes one line instead of pushing memories out of a small card, monospace text in Chinese no longer falls back to a serif font on Windows, and the status bar says “waiting for you”.
Fixed2
After signing out and back in, new agents and terminals start again — before, every one failed with “NeuroSquad is quitting” until the app was restarted.
The Memory card’s switches — automatic recall, automatic saving, agents may edit any memory — toggle again. The scope picker wraps long labels instead of scrolling sideways, and an empty card no longer covers the credit line.
Memory for agents: a plugin that keeps what your agents learn on your computer and recalls it in later sessions. And Claude Code cards that no longer get stuck on Working.
New5
Memory (mem0) plugin: draw an arrow from an agent to a Memory card and it gets six memory tools — search, add, list, get, update, delete. Remove the arrow and they are gone. Memories are kept in the app’s data folder — no account, no server — per workspace, across all workspaces, or privately per agent. An agent may change only what it saved itself unless you allow more. Built on mem0’s open-source engine, with its telemetry switched off.
Smart search for Memory: a multilingual model that runs on your computer finds memories by meaning, in any language — a question in Russian finds a note written in English. A one-time download of about 150 MB, every file checked; keyword search keeps working offline with nothing downloaded. An embedding model on OpenRouter or your own server is another option.
Automatic memory: with recall on, the memories that fit your prompt are added to the agent’s turn in Claude Code, Codex, Qwen Code, OpenCode, Kilo Code, Gemini CLI, pi and omp. With a fact-extraction model chosen on the card, your prompt and the agent’s answer are saved after each turn — the model keeps only lasting facts.
Memory history, export and import: every memory shows when it was added and how it was edited; an export can be imported into another workspace with a preview (new, already there, skipped) and undone in one press.
“Copy status trace” in an agent card’s menu: the card’s recent status events — event types and transitions only, never prompts, replies or commands — to attach to a bug report.
Improved2
Claude Code cards now check their status against the session’s own transcript, so a turn that ended without the usual signal is still seen as ended.
Deleting a workspace deletes its memories, and deleting an agent deletes its private ones; the delete dialog says how many first.
Fixed5
A Claude Code card no longer stays on Working after you interrupt it with Escape or Ctrl+C — it goes idle at once.
When Claude stops on an API error (a rate limit, an overloaded server), the card says “Claude stopped: API error” instead of working forever, and queued prompts are not sent into it.
“Needs your answer” appears only for a real permission prompt or question — not for Claude’s other notifications after a turn that simply finished.
A prompt typed while Claude is still working no longer rings “finished” between the two turns, and a lost prompt hook no longer leaves the card idle while Claude works.
A reloaded window or a popped-out card shows the agent’s current status right away.
The Mac app is now signed with Apple Developer ID and notarized by Apple.
Improved1
The Mac app is signed with an Apple Developer ID and notarized by Apple: a .dmg downloaded in the browser opens with macOS’s usual “downloaded from the Internet — Open?” question, without a trip to Privacy & Security. Installs from an earlier Mac build ask once more for microphone and Accessibility access after this update.
NeuroSquad for macOS, a Grid mode next to the canvas, workspaces in WSL and over SSH, and a “needs attention” signal you can trust.
New4
NeuroSquad for macOS 12 or newer, on Apple silicon and Intel: a native window with traffic lights, the macOS menu, Cmd shortcuts, and microphone and Accessibility prompts for dictation. Install it with one line in Terminal (`curl -fsSL https://neurosquad.ai/install.sh | bash`), with Homebrew (`brew install --cask glmn-ai/neurosquad/neurosquad`) or from a .dmg. It updates itself, and a stable signature keeps the permissions you granted across updates.
Grid mode next to the canvas (Ctrl+Shift+G, Cmd+Shift+G on a Mac): agents as tiles in layout presets — resize, swap, maximize, move between tiles from the keyboard. Chips under each tile show what the agent is connected to; “Only agents” and the cards strip filter the view, the Squad panel is one click away, and the grid has its own undo. The same terminals move between canvas and grid — nothing restarts.
Workspaces in WSL on Windows: choose a distro and a Linux folder in the workspace dialog. Shell cards and Claude Code run inside WSL with statuses, MCP tools by arrow and transcripts. Other agent CLIs don’t run in WSL yet.
Workspaces over SSH: a folder on a remote host, with shells and Claude Code running there — statuses, MCP tools by arrow and transcripts included. Host keys are verified, your keys and SSH agent are used, one connection is shared per host and it reconnects on its own. Other agent CLIs don’t run over SSH yet.
Improved4
“Needs attention” clears the moment you answer the agent in its terminal — Enter, a menu number or y/n — and Escape or Ctrl+C sets it back to idle without a false “finished”. Repeated and stale alerts are dropped, and a restarted agent never keeps an old one.
The context meter shows the window the CLI really uses — 1M for a second Claude account, where it used to read 200K.
Undo and redo for moving and resizing cards on the canvas: Ctrl+Z, Ctrl+Shift+Z or Ctrl+Y (Cmd on a Mac).
Workspaces, cards and settings are saved atomically: a crash or power loss in the middle of a save can no longer leave a half-written file.
Fixed9
Security hardening from a full code audit: “Open in editor” can no longer be tricked into running commands, a path-traversal hole is closed, the OpenRouter relay is pinned to its host, and channels to secrets that a connected phone could reach are gone.
Usage: costs below a millionth of a dollar no longer show as $0, and OpenCode on your own provider shows “no price” instead of $0.
Claude Code’s “Do you trust this folder?” question is answered automatically on macOS too, as on Windows — found by the end-to-end test on a real Mac.
Quitting is clean: the Windows tray icon disappears at once, a watchdog makes sure no windowless process stays behind, and the macOS crash on quit with agents still running is fixed.
The bell, the taskbar badge and the inbox show one number — agents waiting for you plus unseen finished turns. Stopping or restarting a terminal no longer adds to it.
After a terminal restarted, stray text like `1;2c` could appear in the shell as if typed — the terminal’s automatic answers now only reach the process they were meant for.
A card no longer takes the program’s path (`…\claude.exe`) as its session title.
The browser card shows static pages again instead of staying blank, and on a Mac it no longer opens a visible Chrome window.
The add menu no longer marks installed CLIs (Codex, OpenCode, Crush) as “not detected” when the first check at startup was slow — it checks again when you open it.
Plugins you switch on with an arrow — RTK-AI Token Saver and Caveman — your own model servers, and dangerous mode that applies without a restart.
New5
Plugins: a third catalog next to MCP servers and skills. “Plugin…” in the add menu opens it; connect a plugin to an agent with an arrow and it takes effect on the agent’s next action — no restart, and your CLI’s own settings stay untouched.
RTK-AI Token Saver: an agent’s shell commands (git, ls, grep, test runners…) run through RTK, which compresses their output before the model reads it. Works live with Claude Code (normal and dangerous mode), Cursor CLI, OpenCode and Kilo Code (dangerous mode). RTK is downloaded from GitHub only with a verified checksum.
Caveman: with an arrow, the agent answers in caveman’s terse style — Lite, Full or Ultra — and goes back to normal when you remove it. Live for Claude Code, Codex and Qwen Code.
Your own model providers: LM Studio, Ollama, llama.cpp, vLLM, Unsloth or any compatible server. “Test connection” checks which APIs it speaks, and every CLI that can use it offers it — Claude Code, Codex, OpenCode, Qwen Code and eleven more.
“Suggest an idea” and “Report a problem” in the account menu open the new Ideas and Issues boards at app.neurosquad.ai, with your app version and OS filled in.
Improved2
Dangerous mode now switches on and off without restarting a Claude Code card — the session goes on. Other CLIs restart with their session kept.
The add menu shows skills, MCP servers and plugins first, and “New agent…” leads the list of agents.
See how much of your Claude and Codex plan is left, and keep several accounts per CLI.
New2
Plan-limit meters for Claude Code and Codex: the 5-hour and weekly windows with reset times, on each agent card, in the Squad board and in Usage, with a heads-up at 80% and 95%. The numbers come from what the CLIs themselves report — NeuroSquad never reads your login.
Account profiles for Claude Code and Codex — for example Personal and Work. Pick one per card, sign in inside the card with the CLI’s own login, and when a card hits its limit, “Continue on another account” hands the session over with one click.
Lead agents connect cards to their helpers; “Open NeuroSquad” after signing in works.
New1
Lead agents can draw and remove arrows between cards — for example, give a helper the note it should write to — and connect a new helper to cards the moment they create it. The same rules as your own arrows apply: a helper only sees what it’s connected to, and a terminal you opened is reachable only through an arrow you drew.
Fixed2
“Open NeuroSquad” on the sign-in page now brings the app forward and finishes signing in.
Launching NeuroSquad again focuses the window that’s already open instead of starting a second copy.
NeuroSquad accounts with a free plan, lead agents that pick models, and the Squad board in a sidebar.
New3
NeuroSquad accounts: on first launch you sign in with an email code or Google at app.neurosquad.ai — the app opens the browser, you press Connect, and you’re in. Everyone gets the free plan (unlimited). Your agents, terminals, code and files stay on your computer; the app sends usage counts only — never code, files, prompts, names or paths. Account menu in the sidebar; works offline for 7 days.
Lead agents can choose models for the agents they create — from each CLI’s own models or from OpenRouter.
The Squad board can live in a resizable right sidebar, opened from the top bar.
Fixed3
Starting a card for a CLI that isn’t installed no longer fails with “File not found” — the card offers to install it or pick another.
An agent CLI’s own update prompt no longer blocks its card: the update is skipped and you get a notification.
When Windows blocks notifications, they appear inside the app window; Settings has a switch per kind.
19 agent CLIs, all integrated at the same depth: Gemini CLI, GitHub Copilot CLI, Amp, Factory Droid, Aider, Goose, Kimi Code, Cline and Auggie join.
New1
New agent cards for Gemini CLI, GitHub Copilot CLI, Amp, Factory Droid, Aider, Goose, Kimi Code, Cline CLI and Auggie. Each gets session resume, precise statuses with what the agent is asking, MCP tools over arrows plus connected MCP servers and skills, canvas mode, safe interruption, usage where the tool records it, and NeuroSquad’s OpenRouter attribution wherever the tool talks to OpenRouter. Where a tool can’t do something, the docs say so and why.
Improved1
Cursor CLI and Crush are brought up to the same level: session resume, statuses from their own events, MCP, and canvas mode.
Codex CLI, pi, omp, Kilo Code and Qwen Code now do everything Claude Code does in NeuroSquad.
New1
Codex CLI, pi, omp, Kilo Code and Qwen Code are integrated as fully as Claude Code: session resume, precise working / needs you / finished statuses with what the agent is asking, MCP tools over arrows plus connected MCP servers and skills, canvas mode (also in dangerous mode), safe interruption by the budget brake, usage per card including subagents, the model in the Squad card, the context meter and compaction, auto-continue after rate limits, hand-off, journal and helpers.
Improved1
Every harness sends NeuroSquad’s OpenRouter attribution, including on the harness’s own OpenRouter login.
Fixed3
codex.exe no longer survives stopping its card.
Kilo Code cards no longer attach to a shared Kilo daemon and lose their settings.
Qwen Code’s connection token for the app’s MCP server is no longer written to disk.
Hermes Agent and OpenCode now do everything Claude Code does in NeuroSquad; double-click the top bar to maximize.
New3
Hermes Agent is integrated as fully as Claude Code: it resumes its session after a restart, shows working / needs you / finished, uses MCP tools over arrows and connected MCP servers and skills, supports dangerous and canvas modes, usage and model, the context meter and compaction, hand-off, journal and reviewer, and OpenRouter with NeuroSquad attribution. Third-party MCP tools ask for your approval in an app dialog.
OpenCode reaches the same level: session resume, precise statuses through a NeuroSquad plugin, MCP and connected servers and skills, canvas mode and per-card usage. The old crash on Windows no longer reproduces, and the budget brake interrupts OpenCode without closing it.
Double-click the top bar or the sidebar’s logo row to maximize or restore the window; it also leaves fullscreen.
Verified cards: a catalog of community cards reviewed by the NeuroSquad team, with search and a badge.
New1
Verified cards: the app pulls the list of cards reviewed by the NeuroSquad team from glmn-ai/neurosquad-cards. Search it by name in “Custom card…” or Settings → Custom cards and install exactly the reviewed version in one click. Verified cards carry a green shield badge, and new reviewed versions are offered with the usual permission check. Authors submit their cards by pull request.
Improved1
“Check for update” is hidden for cards pinned to a commit; verified updates have their own button.
Community cards from GitHub, MCP servers and Skills you connect to agents with an arrow, and tighter security.
New2
Community cards: install cards made by anyone from a GitHub repository. Each runs in a sandbox on the canvas, asks only for the permissions it needs, and updates only when you click. Cards can follow agent events, give agents tools, exchange data with other cards over arrows, and look however their author likes. Settings → Custom cards.
MCP servers and Skills: search the official MCP Registry and skills.sh right in the app, install with one consent screen, and connect a server or a skill to a specific agent with an arrow. An agent sees only what is connected to it.
Improved3
Each agent has its own token for the app’s MCP server, so one agent can no longer act as another.
Tools from third-party MCP servers ask before they run.
Microphone access is granted only to the app itself.
Quitting no longer ends with a “Fail Fast exception” error.
Fixed1
Quitting NeuroSquad — from the tray or by closing the window — no longer sometimes ends with “neurosquad.exe - Fail Fast exception”: the dictation hotkey’s keyboard and mouse hook is now stopped before the app shuts down.
Squad status card on every canvas, an add-card menu that starts with agents, and a much faster canvas with many cards.
New2
Squad status card: every AI agent of the workspace on one board — who needs you, who is working, finished or idle, with its CLI, model and how long the turn has run. It is added to every workspace once; click a row to fly to that agent.
The add-card menu starts with agents — the ones installed on this computer are marked — and has a search box; tools and other cards sit in groups you can collapse. An empty canvas opens with big agent tiles.
Improved1
The canvas is much lighter with many cards: while nothing happens it uses a fraction of the CPU and GPU it did, and dragging, zooming and the overview no longer stutter.
Fixed1
Opening an existing workspace for the first time could pile all its cards up in one corner.
Updating from an older version no longer gets stuck on a locked folder; tidier sidebar badges and wizard steps.
Improved2
Count badges on the sidebar’s section buttons sit in each button’s top-right corner instead of floating mid-button.
The setup wizard’s step chips are a size larger and no longer crowd the title.
Fixed3
Updating over 0.1.94 or older no longer fails with “the previous version’s folder is still locked”: terminals of cards left behind by the old version are now listed with the other processes and ended by the same button.
If something the installer can’t identify still holds the folder (an Explorer window, for example), it replaces the files one by one instead of giving up — and puts everything back if any step fails. Files you added to the folder stay.
Windows’ own Russian and Chinese error messages read correctly in the installer log.
A new installer and automatic updates: from this version on, NeuroSquad keeps itself up to date.
New6
Automatic updates. NeuroSquad checks for a new release at every start and every hour, downloads it in the background and verifies its SHA-256 checksum before installing.
You choose when: “Update now” or “On next start”. A ring in the title bar shows the download, then an icon with a dot lets you update at any moment; the next start installs it in any case.
A new installer of our own: choose the install folder (the default one needs no administrator rights), add a desktop shortcut or not, and follow a live log of every step. It speaks English, Russian and Chinese.
If NeuroSquad is still running, the installer lists what is in the way by name and ends it only when you press the button.
Updates never leave a half-installed app: the new version is unpacked next to the old one and swapped in only when complete, so a failure or a cancel leaves the old version untouched.
The uninstaller removes exactly what was installed, its shortcuts and the Apps entry; your app data is deleted only if you ask.
Fixed2
Terminals of cards no longer keep the install folder busy after the app closes, which used to make updates fail.
The setup wizard no longer says the dictation model ships with the app — it is downloaded when you ask for it.
The installer is down to about 96 MB: only what the app needs to run is packed.
Improved1
A much lighter download: the installer now carries only what the app loads at runtime, with maximum compression — about 96 MB, down from about 110 MB in 0.1.92.
The dictation model is no longer bundled, so the installer is about five times smaller.
New1
Dictation turns on by itself as soon as the model finishes downloading — no restart needed.
Improved2
The ~640 MB dictation model is no longer inside the installer. Download it only if you want voice input: at the Dictation step of the setup wizard or later in Settings → Dictation.
Until the model is there, the status bar says “Dictation: download the model” instead of showing an error.
The first public release of NeuroSquad for Windows 10 and 11.
New11
A canvas of agent cards, each with a live terminal, for ten agent CLIs — Claude Code, Codex CLI, OpenCode, Qwen Code, Cursor CLI, Kilo Code, Hermes Agent, pi, omp and Crush — plus plain PowerShell, cmd and bash terminals.
Agents tell you when they finish a turn or need your answer: a sound, a notification and a highlighted card. A click on the notification flies the canvas to that card.
Cards to work next to your agents: notes in Markdown, to-do lists, kanban, a real Chrome browser, a budget, Telegram, reference images, ports, a web watch, a flow timer, sticky notes and a stand-up.
Arrows between cards are tools: an agent connected to a browser, a terminal, a note or another agent can use it through the built-in MCP server.
Remote access: the whole app in a phone’s or another laptop’s browser, over your local network or through a Cloudflare tunnel, paired with a QR code.
Voice dictation with a global hotkey, recognized on your computer (Parakeet or Whisper) and typed into the active agent.
The Usage section: tokens and cost per card and per model, read from the agents’ own logs.
Agents can run on OpenRouter models: pick the CLI, the provider and the model when you create an agent.
Each agent can work in its own isolated git worktree; a budget limit per workspace pauses the agents when it is reached.
A first-launch setup wizard: agent CLI, terminal and Git, dictation model — every step can be skipped.
The interface in English, Russian and Chinese, switched live in Settings → General.