agentgc v0.3.0: isolated environments for parallel AI coding agents

agentgc gives each AI coding agent task its own environment: a git worktree, free ports, a database name and an env file. It can run the services of a task, queue heavy commands, and clean up afterwards, always with a plan and a question first.

agentgc does not start AI agents. It prepares and cleans the environment around them.

What is new in 0.3.0

• agentgc task review NAME shows what a task changed before you finish it. It only reads.
• Claude Code hooks: agentgc hook install adds a PreToolUse hook that queues heavy commands through agentgc run --, and a SessionStart hook that prints the task status and env. The hook asks for approval by default; --allow is the opt-in to skip the question. Running hook install again replaces the old agentgc entry.
• task done resumes after a refused branch delete: the worktree is gone, the registry entry stays, and the next run retries the branch delete.
• agentgc doctor lists hard-coded ports and database names in the files git tracks. It only reads.
• agentgc run --min-interval waits at least that many seconds between the starts of two queued commands.
• db.disconnect in agentgc.task.toml runs your command before the database drop in task done.
• agentgc worktrees --deps estimates how much a shared dependency store would save.

What is new in 0.2.0

• agentgc task list shows the state of every task: changed files, commits ahead and behind the base branch, and how many services run. --json has the same fields. The command only reads.
• Services can wait for each other. In [services.NAME] set depends_on = ["db"] and ready_timeout = 30. A service with a port is ready when the port accepts a connection. up starts the services in order, and status has a READY column. A cycle or an unknown name is refused at load time.

What is in the tool

Area	Commands
Isolate	task new, task list, task done (dry run by default, refuses unsaved work, resumable)
Supervise	up, down, status, logs, restart (opt-in, per task, no system service, no network port)
Queue	run (heavy commands take turns, named exclusive resources)
Clean	scan, worktrees, clean (one question per item, Recycle Bin for risky items), ps, ps --stop
Teach agents	init (writes marked rules into AGENTS.md, CLAUDE.md and a Claude Code skill, after you confirm each file)

Full command list: README and docs/reference.md.

Download

System	File
Windows (x64)	agentgc-v0.3.0-x86_64-pc-windows-msvc.zip
Linux (x64)	agentgc-v0.3.0-x86_64-unknown-linux-gnu.tar.gz
macOS (Apple Silicon)	agentgc-v0.3.0-aarch64-apple-darwin.tar.gz

Each archive holds the binary, the README and the license. Unpack it and put agentgc on your PATH. Other systems, such as Intel Macs: build from source with cargo install --path ..

Verify a download: compare the output of sha256sum <file> (macOS: shasum -a 256 <file>, Windows: certutil -hashfile <file> SHA256) with the .sha256 file next to it.

The binaries are not code signed. Windows SmartScreen or macOS Gatekeeper can warn on first start. If you do not trust a binary, build it from source.

Platform status

• Tests pass on Windows, Linux and macOS in GitHub Actions.
• The full flow was run end to end by hand on Windows only. Linux and macOS are untested by hand. Please report problems in an issue.
• On Windows, ps --stop needs --force, because agentgc has no graceful stop there.

Before the first clean

Run agentgc clean --dry-run. Nothing is deleted without a question for each item.

Found a safety bug, such as a wrong deletion? Use the Security tab, not a public issue. See SECURITY.md.