Skip to content

Changelog

Notable changes per release, in the format of Keep a Changelog. This project follows Semantic Versioning. While it is pre-1.0, a backwards-compatible change is a patch bump (0.1.0 → 0.1.1), and a change that breaks a swarmed command, a file in the data directory or swarmed-core’s public API moves to the next minor (0.1.x → 0.2.0).

The release workflow publishes the section matching a tag as that release’s notes, so an entry here is what a user reads on the GitHub release page. Each pull request with a change a user would notice adds its line under Unreleased.

  • Swarmed updates itself. About shows whether a newer version is out, and Update and restart installs it. With Update automatically on, as it is unless turned off under Settings → About, Swarmed looks at launch and every few hours, downloads a new version, and installs it when you quit; a dot on the version in the corner says one is ready. Off, nothing is fetched until About opens, and nothing is installed unless you ask. An update is installed only if it is signed with the release key, for the version it claims to be; one that could not be put in says why at the next launch. An app built from the sources is never updated. This version still has to be installed by hand: the ones after it update from here.
  • swarmed-core: Settings has updates, an UpdateMode: automatic or when_asked.
  • A swarm dictionary, in Settings just before About, in the swarm’s styles. It lists every swarm word with the plain word it stands for and what that is, and you can search it. It is made from the words the app uses, so it always matches them.
  • The swarm’s styles reach further. Settings is the Evolution Chamber and the keyboard shortcuts are Reflexes. The version is an evolution: e0.2.1 in the corner, “Evolution 0.2.1” in About. Untitled sessions, empty and loading sessions, forks, changes, files, sub-agents, actions, the inbox, search and the activity charts now take the swarm’s words where they still had the plain ones. True swarm goes all the way. Repositories and tasks are colonies and broods, a worktree is a cell, a branch a strain and the base branch the mother strain. Committing is sealing, pushing is releasing, and a pull request is an offering. Agents are species, models minds, MCP servers symbionts, skills genes, prompts pheromones, and the rules are the hive law. Questions, errors and setting hints stay plain, and so does everything an agent is given. The plain style reads as before.
  • url 2.5.9, for RUSTSEC-2026-0592: in 2.5.8, Url::origin() on a URL nested in many blob: prefixes overflowed the stack and ended the process. Swarmed calls it itself only in a build run from the sources, but Tauri and its web view’s crates use url as well.

  • An OpenCode session left on “Default model” runs on the first model listed for OpenCode in Settings, and the model menu names it. It used to run on OpenCode’s own default, whichever model OpenRouter added last, which changes without notice and may be one the account’s allowed providers don’t serve: a first message could fail with “No allowed providers are available for the selected model”. The start of the session also named a model other than the one the message ran on. A model set in your own opencode.json stays OpenCode’s default.

  • swarmed-core: AgentInfo and AcpBackend have default_model, what a session that leaves the model open runs on where Swarmed chooses it; None leaves it to the agent.

  • Removing a task in the app deletes its local branch too, unless the task started from that branch, or commits on it are on neither the base nor a remote, which a forced removal can leave: then the branch is kept, and the removal says so. A branch on a remote is never touched. swarmed task rm keeps the branch as before, unless given --delete-branch.
  • swarmed-core: NewTaskOptions has free_name, which steps past a taken name to name-2 and so on, and RemoveOptions::delete_branch keeps a branch whose commits exist nowhere else. Error has the variants BranchExists, TaskFolderExists, NoFreeTaskName and Migration.
  • A new session no longer shows the conversation of one culled before it. A deleted session’s number could be given to the next new one, which then showed what the window still kept of the old conversation. Numbers are now never given twice, and what is kept of a culled session, or of a removed task’s sessions, is forgotten.
  • A new task’s name no longer collides with a branch. A name made up from the prompt now steps past the names of removed tasks and branches here or on a remote, to name-2 and so on, where creating the task used to fail with “a branch named … already exists”.
  • Removing a task also deletes what its sessions left: messages still waiting to be sent, views nobody read, and the files handed to their agents, which can hold MCP servers’ secrets.
  • A Claude Code command whose screen exists only in its own terminal opens there in one click. The line that says so, under /vim or /privacy-settings, has “Open in Claude Code”. It starts Claude Code on that command in the session’s terminal, in place of the shell until you quit it or go back to the shell. It runs with your own sign-in, in a conversation of its own, and nothing it shows is sent to the agent. The usage screen has the same button for /usage-credits, where Claude Code buys usage credits and sets their monthly cap. An action’s command still goes to the shell, which is shown again for it.
  • The coins warn before a limit is reached. Once a usage window is 75% used, as Claude Code starts to warn, they carry a small yellow triangle, red once a limit is used up, and the bars in the card and the overview turn yellow and red likewise. Extra usage’s bar is grey now, so that yellow always means a warning.
  • The coins at the top right of a Claude Code session show your usage as it is now. They open an overview of the plan’s limits, extra usage and what the session spent, with a bar for each that carries its share, and ask Claude Code for the figures as they open, with nothing sent to the model and nothing added to the conversation. The context card keeps the figures last heard, from a turn, the coins or /usage, now with a thin bar under each limit.
  • Claude Code’s own commands do in Swarmed what they do in a terminal. /usage, /status, /context, /cost, /memory, /skills, /hooks and /mcp show their screen in the conversation, with your plan’s limits as bars, instead of a block of text or “isn’t available in this environment”. /model, /effort and /config ask where the conversation stands and are set by your answer; /model opus sets it at once. None of them is sent to the model or interrupts a turn. The list after / says which commands are shown here, which ask here, and which are not sent, with what to do instead; those cannot wait in the queue either. Any other command that only opens a screen in Claude Code’s terminal, such as /vim, is said to, instead of “isn’t available in this environment” as if the agent had answered. /usage-credits shows your usage and extra usage, and says that credits are bought in the terminal. /clear, /new and /reset are not sent: Claude Code would forget the conversation while the session went on showing it, so they point to a new session. /cd and /add-dir are refused, since a session works in its task’s worktree, and so are /remote-control, /teleport and /desktop, which would hand the session elsewhere.
  • The context window’s card shows your extra usage: what is spent beyond your plan this month, billed at API prices, against the monthly cap you set, and when this session’s turns run on it. When a plan’s limit is reached and Claude Code goes on with extra usage, the conversation says so, once, instead of “Usage limit reached”.
  • The context window’s card shows your usage limits and what the session has spent. For Claude Code, each usage window (5-hour, weekly and the like) with how much is used and when it resets; for any agent that reports a cost, what the session has spent in all, and whether that is billed or covered by a subscription. Agents over the Agent Client Protocol report no limits.
  • swarmed-core: Core::task_remove and Core::project_remove return what was passed over (Vec<String>) instead of nothing; Core::session_send and Core::session_message do Claude Code’s own commands themselves instead of sending them; and AgentEvent, AgentCommand and Error have new variants, which a match over them has to cover.
  • swarmed-core: Core::slash_commands returns each command with what is done with it (SlashCommand: its name, its kind and a note) instead of only its name.
  • Removing a task no longer waits for its worktree to be deleted, and only unsaved work stops it. The worktree is moved aside and deleted in the background; what is not deleted yet, or could not be, is counted in Settings under what Swarmed takes up. A stop or teardown that fails or takes over a minute is passed over, and what was passed over is shown afterwards. The dialog’s “Keep it” button is now “Close”, and closing it does not stop a removal under way.
  • An environment cannot be removed while a task still uses it.
  • The button that ends a session is the last in the session’s header, at the far right, apart from the ones that carry the session on.
  • The task panel opens a task in your terminal or editor from two icons beside the one that hides the panel. Which terminal and editor they open is chosen in Settings, so the panel no longer has a menu of editors of its own.
  • The swarm’s icons and words are called True swarm and Weak swarm in Settings → Appearance: True swarm, which a new installation starts in, renames the tree too, and Weak swarm keeps the plain names there. Nothing else about them changed, and a setting saved before keeps its choice.
  • The plain icons and words are called Boring in Settings → Appearance. Nothing else about them changed, and a setting saved before keeps them.
  • What a program prints as it ends in a session’s terminal is no longer lost. On a Mac, the last lines of one that printed and ended at once could go missing.
  • What sessions cost is no longer counted many times over. Claude Code reports the conversation’s running total after every turn and keeps it when a session is resumed, but the costs in Settings, the turn statistics and the context card’s “spent here” started it over at every turn. So they added each turn’s whole running total again, and a session that had spent $1.09 could show $6.47. A total that goes down, as after /clear, still counts whole.
  • A task whose removal failed halfway can be removed. “Remove anyway” now deletes what is left of its worktree, where it used to fail every time once git had lost track of it.
  • Removing a task ends what still runs in its worktree, such as a stack started by hand in its terminal, before deleting it. Such a stack used to keep running, recreating files as they were deleted, which made the removal fail.
  • The button that takes a cloud task off the list shows on hover and asks first. It used to be invisible, and one click took the task off at once.
  • A usage warning from Claude Code no longer reads as a limit reached. A window filling up is said as a warning, with how much of it is used, once rather than after every turn. When a window resets on another day, the day is given as well as the time.
  • A stack that would not end when Swarmed quits is found again by the next one, as it already was after stopping a single stack.

The first release: a macOS app and a swarmed command over one Rust core. Swarmed is not signed with an Apple Developer ID yet, so macOS asks before opening it the first time when it comes through Homebrew or a browser.

  • A task is a branch in a worktree of its own. Add a git repository, say what you want done, and Swarmed creates the branch and a worktree under ~/swarmed/<project>/<task>, runs the repository’s setup hook, and starts an agent there. Nothing is written into the repository itself.
  • Any agent. Claude Code runs with its own login and configuration; Codex, OpenCode, Antigravity and any agent that speaks the Agent Client Protocol are set up from Settings → Agents or in agents.toml. Swarmed never signs you in to an agent and keeps no GitHub token: it uses each agent’s own login and your gh. A new session’s agent can be picked until its first message, and a conversation can be cloned onto another agent.
  • The conversation. Replies stream as they are written, with syntax highlighting, and HTML an agent hands over is shown in a sandbox on an origin of its own. The transcript says what the agent is doing and how long each turn took. Messages queue while the agent works; files can be picked or dropped on the window; commands, skills and saved prompts are offered after a slash.
  • Supervision. Permission requests are answered in the conversation, and an agent’s questions by arrow keys, number or click. Auto mode, plan mode, the model and its effort are set per session and can change mid-session.
  • Sessions. Rename them by hand or by a model, fork one from any message into a new session that holds that message and what followed, close, archive, reopen or delete them, and search what was said in any of them. Unread answers are marked, and notifications and a count on the Dock icon say when a session needs you, finishes or fails; ⌘I lists everything that waits.
  • Sessions working together. Agents get tools to see, read, message, start and close other sessions, and spawned sessions are listed under the one that started them. Claude Code’s sub-agents appear under their session while they work. One prompt can go to several agents at once, and work can be handed to a Codex cloud agent or a Claude Code cloud session: from a session too, as a brief its agent writes or with the whole conversation.
  • The task panel. Changed files with their diffs, comments on lines sent to a session, a Files tab, commit with a message the session’s agent writes, push, and a pull request whose title and description are written for you, with its checks summed up in a line that unfolds into the list, failures first. Two tasks’ changes can be compared, and a worktree opened in the editor, the browser or the terminal application.
  • A terminal in each session, whose output the agent sees, configured in JSON; processes a task’s agents left running are listed and can be stopped.
  • Environments. Named environments per repository with their ports and variables, a start command that keeps running beside the setup hook, and snapshots of the files git ignores, so a new task starts set up like an earlier one.
  • One pool of MCP servers and skills. The servers configured for Claude Code (its plugins’ too), Codex and OpenCode, and Swarmed’s own, go to every agent that lacks one by that name, read when a session starts (a repository’s own .mcp.json only once approved in Claude Code); skills from every agent and from the task’s repository likewise.
  • Your own setup. Rules every agent is given, notes on a repository and on a task, actions of your own, a status line from built-in pieces or a script, keyboard shortcuts, and API keys kept in the macOS Keychain for agents that need them. Agents can propose configuration changes; rules, environments and commands change only after your yes.
  • Looks. Six colour palettes, Swarm among them with its own words and marks, the classic, senior and custom layouts, and the text size.
  • What it costs. What the agents report their work cost, by day and by task; turns by day and model in charts; and what Swarmed takes up on disk and is using right now.
  • About, and a menu bar of its own. Swarmed’s menu has About, Settings and Keyboard Shortcuts. The About view, also in Settings and behind the version in the sidebar’s top corner, says what Swarmed is and which version is installed, and links to its source.
  • The swarmed command for repositories, tasks, sessions, agents, keys and snapshots from a terminal. SWARMED_FAKE_AGENT=1 swaps in a scripted agent that spends no tokens.
  • Installing: an install script, a Homebrew cask that also puts swarmed on the PATH, and cargo install swarmed for the command alone.