Agent tools: multiplexers, desktop environments, sandboxes, gateways and more

Agent tools are third-party software around Claude Code, Codex and Cursor: terminal multiplexers, desktop agent environments, remote clients, sandboxes, review bots and model gateways. The agents now ship worktrees, agent views and remote control themselves, so a third-party tool earns its place only with cross-agent state, a review GUI, or isolation beyond files.

Use this page when someone proposes a fourth app to manage your three agents: check the built-ins, place the tool on a shelf, and decide within a week.

What the agents already do before you install anything

Checked against Claude Code 2.1.283 and codex-cli 0.157.1 on 2026-09-26; the Cursor column comes from Cursor’s documentation as read on 2026-08-28.

Job	Claude Code	Codex	Cursor
A git worktree per task	claude --worktree NAME; .worktreeinclude copies gitignored files	codex --worktree or /worktree	Worktrees in Agent
A terminal session per worktree	--tmux (--tmux=classic for plain tmux)	—	—
One screen for all agents	claude agents (agent view, research preview)	codex agents	—
Agents cooperating	Agent teams (experimental, off by default)	Subagents	Subagents
Steering from another device	claude --remote-control	codex remote-control (experimental)	—
Diff review	/code-review	codex review	Bugbot (on pull requests)

Cursor: agent view and remote control not documented as of 2026-08-28 (host blocked since).

Try the native worktree workflow first

Start two agents on two tasks, each in its own checkout.

Claude Code

In two terminals, from the repository root:

claude --worktree feature-auth --tmux
claude --worktree fix-invoice-rounding --tmux

Each session gets its own worktree and branch; --tmux opens it in its own tmux session (iTerm2 native panes when available), so it survives a closed terminal. List .env in .worktreeinclude to copy it into new worktrees; watch both with claude agents.

Codex

In each terminal, from the repository root:

codex --worktree

Codex starts the session in a new managed worktree; /worktree does the same inside a session. codex agents lists the sessions on the local app-server daemon.

Cursor

Start each agent in a worktree from the Agent interface; the control moves between releases, so follow Cursor’s worktrees documentation.

Then prove the isolation:

git worktree list

Expect the main checkout plus one line per agent, each on its own branch. A single line means an agent is editing your main checkout: stop it.

Copy-paste prompt: a worktree task that proves itself

You are working in a dedicated git worktree on one task: fix the rounding
error in invoice totals (round half-even to 2 decimals).
1. Before changing code, write a failing test that reproduces the bug and
   show me the failing output.
2. Make the smallest fix that turns it green. Touch only src/billing/ and
   tests/billing/.
3. Run the full test suite, the type checker and the linter. Paste the
   summary line of each.
4. If a server or test needs a port, do not assume the default; tell me
   which port you used.
5. End with: branch name, files changed, commands run, and anything you
   could not verify.

The agent hands back evidence, not a claim: a red-then-green test, three gate results and a scope you confirm with git diff --stat main.

The six shelves of agent tools

Shelf	What it adds over the built-ins	Go deeper
Agent-aware multiplexers	Working, blocked or idle state across different agents; sessions that survive SSH drops; worktree managers such as Worktrunk	herdr, tmux for agent fleets
Desktop agent environments	A GUI diff and review board over a worktree per task	desktop environments compared
Remote and mobile clients	Approvals and steering from a phone, for mixed agents	agents from your phone
Isolation runtimes	A container, VM or cloud sandbox per agent: ports, databases and credentials isolated, not only files	sandboxes compared
Review bots	A second model on every pull request	AI code review bots
Gateways and local models	Provider switching or local inference; you lose Remote Control and /voice	gateways and local models

Which agent tools are worth knowing in September 2026?

One or two picks per shelf; the shelf pages hold the rest and the install lines. Stars are GitHub stars read on 2026-09-26: attention, not use.

Tool	Shelf	Pick it when	License	★
herdr	Multiplexer	Mixed agents in one terminal	Apache-2.0	40.8k
tmux	Multiplexer	No new dependency allowed	ISC	49.5k
Orca	Desktop	Free, cross-platform, best-of-N	MIT	78.4k
Conductor	Desktop	Mac team, review board (secondary sources)	closed	—
Happy	Remote	End-to-end-encrypted phone client	MIT	23.9k
container-use	Isolation	Container plus branch per agent over MCP (experimental)	Apache-2.0	4.0k
CC Switch	Gateway	Switch providers across agents (site: ccswitch.io)	MIT	136.9k
Ollama	Gateway	Code must stay on the machine	MIT	181.7k

Clear the license with legal before a tool becomes a team standard: AGPL-3.0 (Claude Squad, Coder) and GPL-3.0 (cmux) bite when you embed, host or distribute; ELv2 (Superset) forbids hosting as a service; closed-source Conductor needs a SaaS review.

Which agent tools should you not recommend?

As of 2026-09-26, keep these out of a team standard: Vibe Kanban (sunsetting), Uzi (dormant since 2025-06-04), Crystal (continues as Nimbalyst), Roo Code and coder/agentapi (archived), Continue (read-only), and subscription-resale proxies such as CLIProxyAPI and sub2api (terms compliance unverified).

When should you add a third-party agent tool?

A tech lead runs this before a tool enters the team setup.

1. Name the gap in one sentence, such as “we cannot see which agent waits for approval.” If a built-in covers it, configure that instead.

2. Place it on a shelf. A multiplexer does not isolate ports; a sandbox has no review board.

3. Install from the vendor’s README only. On npm, herdr is a placeholder, claude-squad, conductor and oh-my-claudecode belong to unrelated authors, and coderabbit is a security holding package. Compare npm view TOOL repository with the README.

4. Check health and license: last release, archive flag, LICENSE. Pin and hash Python gateways: LiteLLM 1.82.7 and 1.82.8 on PyPI were malicious (2026-03-24).

5. List what you lose. A gateway (ANTHROPIC_BASE_URL other than api.anthropic.com) or an API key removes Claude Code’s Remote Control, so phone steering needs a third-party client; /voice needs a claude.ai login.

6. Pilot for one week: two people, one backlog slice, unchanged CI gates. Record agent wait time, first-push CI result and review time per task.

7. Keep it only if the gap closes and the gates stay green. Native worktrees stay the fallback: this category churns fast.

Copy-paste prompt: vet a third-party agent tool (replace OWNER/REPO)

Evaluate the agent tool at https://github.com/OWNER/REPO for a team of
eight that uses Claude Code and Codex. Read the repository itself (README,
LICENSE, latest release, archive status, open security issues), not blog
posts. Report as a table:
- License, and whether it allows local use, embedding, hosting for others
- Official install command, and whether the npm/PyPI package lists the
  same repository in its registry metadata
- Last release date; archived or sunsetting?
- What it adds over the Claude Code/Codex built-ins: cross-agent state,
  GUI diff review, isolation of ports/DB/credentials, remote control
- What it switches off (Remote Control, voice, auth mode)
- Where prompts, code and credentials are stored; does data leave the machine?
End with one line: adopt, pilot or reject, and why.

How do you verify work that agent tools multiply?

An agent tool multiplies code; it checks none of it. CI gates decide merges, a review bot reads first and a named human owns the merge (agent pull request review). git worktree list proves file isolation only, so run --dangerously-skip-permissions only in a disposable sandbox (permissions and sandboxing).

What goes wrong with agent tools, and how to recover

Symptom	Cause	Recovery
Two agents test against one dev server	Worktrees isolate files, not ports	A port block per worktree, or a sandbox
An agent in a new worktree lacks secrets	.env is gitignored	List it in .worktreeinclude
The standard tool stops shipping	Category churn	Fall back to native worktrees

Where to go next with agent tools

Running many agents at once Worktrees, ports, merge train.

herdr: the agent multiplexer Cross-agent state in a terminal.

Desktop agent environments compared Review GUIs and their licenses.

Agent sandboxes compared Containers, VMs, cloud sandboxes.

Gateways and local models LiteLLM, CC Switch and Ollama.

Headless agents in CI Agents in your pipeline.

Other coding agents Cline, Goose, Aider and others.

Voice input Dictating prompts to agents.

For skills, MCP servers and plugins, see the ecosystem overview; for built-in supervisors, Claude Code agent view and Codex worktrees.