Security
Supported versions
Section titled “Supported versions”Only the latest release gets security fixes. Swarmed is pre-1.0, so a fix ships as a new release
rather than a backport; update to pick it up (run the install script again, or
brew upgrade --cask swarmed).
Reporting a vulnerability
Section titled “Reporting a vulnerability”Report it privately through GitHub: Report a vulnerability. Please don’t open a public issue, pull request or discussion for it.
A useful report says which version and install source you used (the install script, Homebrew, Cargo or a build from source), which agent the session ran on, what you did, and what happened.
The report stays private while it is investigated and fixed. Once a fixed release is out, the advisory is published with credit to you, unless you’d rather not be named.
What Swarmed promises
Section titled “What Swarmed promises”- It never signs you in and never reads the logins agents keep. Each agent uses its own.
GitHub is reached through your own
gh; Swarmed holds no token. - API keys, and the secrets of MCP servers set up in Swarmed, live in the macOS Keychain. Swarmed never shows them again and never logs them. API keys go to agents other than Claude Code, and a server’s secrets to that server.
- MCP servers are handed on as they are configured. To share servers between agents, Swarmed reads Claude Code’s, Codex’s and OpenCode’s MCP configuration, secrets in it included, and passes each server to the agents of your sessions. Claude Code gets its servers in a file only you can read, removed when the agent ends (after a crash, when its session next starts), never on a command line other processes can see.
- Notes stay out of agents’ hands. Notes, and the environment files Swarmed creates, are
readable by you alone, and notes are never handed to an agent. An environment’s variables are
set in its tasks’ processes, agents included, but Swarmed’s own tools tell an agent only their
names, so their values stay out of conversations. Snapshots, which may hold
.envfiles, sit in folders only you can open. - An agent’s HTML cannot reach the app. Views are shown in a sandboxed frame on an origin of their own, under a policy that lets them make no request. A link in a view opens only when you press Open.
- What an agent may change in Swarmed is limited. Its
configuretool cannot reach API keys, MCP servers, sign-ins, supervision or notes. A rule, an environment, an action that runs a command and a status line script are added only when you accept its proposal. - A value in a hook is one word. A branch or task name cannot run as a command.
- A notification never carries the input of a Claude Code tool. It names the tool with the agent’s own summary, or gives the first line of a finished answer. Other agents name their tool calls themselves, and a notice shows that name as given.
- Swarmed makes no network requests of its own apart from fetching OpenRouter’s public model
list, with nothing of yours in it, running an agent’s installer when you ask for one, and
checking whether a stack answers on this machine. There is no telemetry and no update check.
The programs it runs for you, such as the agents,
gitandgh, make their own.
Swarmed does not sandbox the agents. An agent runs as you, in its task’s worktree, and what it may do is decided by the session’s supervision and the agent itself.
What counts
Section titled “What counts”Anything that breaks one of the promises above is in scope, for example:
- An agent’s output, a repository’s content or a branch name causing code to run, or a secret reaching a log, a file, a notification or an agent the promises above keep it from.
- An HTML view reaching the app, the network or your files.
- A problem in a release artefact or the way it is built and published: the app, the install script, the Homebrew cask or the crates.
An agent doing what its supervision allowed is not a Swarmed vulnerability. A vulnerability in a dependency that Swarmed doesn’t actually reach is better raised as a normal issue.