Compare
Forkbench vs Claude Code Agent Teams
Agent Teams is Claude Code's experimental, opt-in feature for coordinating Claude sub-agents inside a single session: one team per session, teammates are always Claude, nested teams are forbidden, and the whole team config directory is deleted when the session ends.
This isn't really an either/or, but it's worth being precise about what Agent Teams actually is. It's an experimental Claude Code feature, off by default behind an environment flag, that lets one session spawn Claude teammates sharing a task list, capped at one team per session, teammates are always Claude, nesting is forbidden, and the whole config directory is deleted the moment the session ends. Forkbench sits a layer below any single vendor: the workspace where Claude Code (with or without Agent Teams), Codex, Gemini CLI, and your own scripts all run together, and what they leave behind outlives the session that created it.
| Forkbench | Claude Code Agent Teams | |
|---|---|---|
| Scope | Cross-vendor workspace + control surface | Claude sub-agents in one session (experimental, opt-in) |
| Agent support | Any terminal CLI, mixed vendors | Claude only, teammates are always Claude |
| Worktree isolation | One-click, manual | Claude Code Desktop: automatic, per session |
| Runs where | Any/all your Macs + web | One session, one machine |
| Secrets handling | Vault (Pro), sealed on your Mac | Inherits the shell env |
| Survives session end | Yes, board + Notebooks persist through pane close, quit, reboot | No, config directory is deleted at session end |
| Persistent memory for agents | Notebooks: end-to-end encrypted, sync across Macs, free | CLAUDE.md + Auto Memory: machine-local plaintext, no sharing |
| Remote session view | Talk: attach-only, any vendor, dies with local session | Channels: attach-only, Claude only, dies with local session |
Where Claude Code Agent Teams is strong
- First-party and native to Claude Code: no extra app, no setup beyond the flag.
- Deep integration with Claude's own model, context windows, and hooks.
- Claude Code Desktop's redesigned app (April 2026) gives each parallel session its own git worktree automatically, a genuine edge over Forkbench's worktrees, which are one click but still a manual step.
- The right tool when your whole workflow is Claude, in one session, on one machine, and nothing needs to outlive that session.
Where Forkbench differs
- A key you put in Vault reaches the command and never the agent; Agent Teams inherits whatever is already in the shell environment. Vault is part of Pro.
- Folders the macOS kernel enforces. A Thread can be locked to the folders the job is about, applied as each shell starts and inherited by every process it spawns, so an agent cannot read or write outside them. Builds still run and the network is untouched.
- Cross-vendor: coordinate Claude Code, Codex, Gemini CLI, Aider, and custom scripts on one board, not Claude teammates alone, with no one-team ceiling and no ban on nesting.
- Durable and sealed: the task board is a per-Thread file with atomic writes, sealed the same way the Vault is, so a task one agent writes today can be claimed by a different vendor's agent tomorrow; Agent Teams' entire config directory is deleted when the session ends.
- Reaches every Mac you own plus the web and your phone; Agent Teams is one team, one session, one machine.
- Notebooks are end-to-end encrypted and sync across your Macs, free with no passphrase. Invite someone into a Thread and they get that Thread's key, never your account Vault or your note library. Claude Code's own memory (CLAUDE.md + Auto Memory) is machine-local, in plaintext, with no sharing model at all.
The bottom line
If your whole workflow is Claude, in one session, on one Mac, Agent Teams is a reasonable fit, more so now that Claude Code Desktop hands each session its own worktree automatically, a step Forkbench still makes you click through. If you mix vendors, need what an agent produces to survive the session that made it, a plan, a claimed task, a resolved discussion, still readable tomorrow by a different agent or a different person, or want to reach your agents from another Mac or your phone, that is what Forkbench is built for. One honest caveat: Anthropic isn't the only place a cross-vendor coordinator could come from, GitHub's Agent HQ already runs Claude, Codex, and Copilot side by side, so 'no model vendor builds this' is true of Anthropic specifically, not of the whole market.
Keep reading
- How keys, notes and boards are actually protected, including where each defence stops.
- Keeping API keys away from AI coding agents
- Every comparison
Comparison based on publicly available information as of mid-2026. Spot something out of date? Tell us.