Six capabilities across terminals
On this page
Does any terminal today have all of these: a structured record of commands, a standard interface for agents, a control layer built on those two, sessions that persist, local-first operation, and rich output that falls back to plain text? No. Each capability exists somewhere, but no product has all six, and every control layer found governs only the product’s own agent.
The survey was made in another working session on 2026-11-22. For this note, the claims about Zellij, VS Code, Warp, iTerm2, Wave, WezTerm, kitty and Ghostty were read again on the official pages the same day, and several were corrected. Each correction is marked. The claims about the other products carry the labels the survey gave them. Nothing was installed or run.
Reliability labels:
- [official]: from the project’s own site or repository.
- [third party]: from a blog, a review or a directory. Not verified.
Versions and dates change fast. Check the official source before relying on them.
The six capabilities
- Record. For each command: its boundaries, exit code, directory and output.
- Interface. An outside agent can read the record, send input, and wait for a command to end.
- Control. On top of 1 and 2: approval policy, output redaction, snapshot and undo, audit.
- Persistence. The processes stay alive when the client disconnects or restarts.
- Local first. No account is needed, data stays on the machine, and the model can be changed.
- Rich output. A program emits rich content, and a terminal without support shows plain text.
Comparison
| Product | Record | Interface | Control | Persistence | Local first | Rich output |
|---|---|---|---|---|---|---|
| Warp | yes | partly | partly | no | partly | partly |
| VS Code terminal | yes | partly | yes | partly | partly | partly |
| iTerm2 | yes | partly | partly | partly | yes | yes |
| Wave | partly | partly | partly | partly | yes | partly |
| Zellij 0.46 | partly | partly | no | yes | yes | partly |
| WezTerm | partly | partly | no | yes | yes | partly |
| kitty | partly | partly | partly | no | yes | yes |
| herdr | no | yes | no | yes | yes | no |
| cmux | no | yes | partly | partly | partly | no |
| DomTerm | yes | no | no | yes | yes | yes |
| Hyperia | no | yes | partly | no | yes | no |
| ht and terminal MCP servers | partly | yes | partly | no | yes | no |
“No” means that nothing was found. It does not mean that absence was proved. A rating is a judgment, and the evidence for it is in the next section. A rating in the Control column says that the capability exists. It does not say which agents it governs. That is the second finding below.
Two ratings differ from the survey. For Wave, the survey had “no” for Record and “yes” for Persistence.
Evidence
Warp
[official: docs.warp.dev, warp.dev, the Warp repository]
- Record. Each command and its output form a block. A block can be attached to the Warp agent as context.
- Control: permissions.
- An agent profile sets a level for each kind of action: agent decides, always ask, always allow, or never.
- A command allowlist, written as regular expressions, runs matching commands without asking. A denylist “takes precedence over both the allowlist and
Agent decides”, and a match asks the person. - The denylist has a hole by default. In “Run until completion” mode, commands on the denylist run without asking unless a setting is turned off. A denylist set by a team administrator is never bypassed.
- MCP servers have an allowlist and a denylist of their own.
- Control: redaction.
- Secret redaction matches a list of regular expressions. A match is “redacted instead of being sent to our servers or any LLM provider”.
- The secret redaction page says it is disabled by default. The survey found a privacy page that describes it as always on in AI interactions. The two pages disagree.
- It is not applied in session sharing.
- Agents the user brings.
- Warp recognizes sixteen third-party CLI agents, among them Claude Code, Codex, OpenCode and Gemini CLI. It gives them a rich input editor, notifications, inline code review and remote control from another device.
- Notifications work for Claude Code and OpenCode through a Warp plugin, and for Codex through a configuration change.
- The page says nothing about permissions, redaction or blocks for these agents. Their sessions do not appear in the panel that lists Warp agent conversations.
- Open source. The client has been open source since 2026-04-28. The repository is AGPL v3, and the two UI framework crates are MIT. The survey reports that the server is closed and that AI requests go through Warp’s servers. That was not read again.
- Persistence. Nothing was found that keeps processes alive after the client exits.
VS Code terminal
[official: code.visualstudio.com]
- Record. Shell integration reads OSC 633, or OSC 133 with fewer features. It gives the command line, the working directory and the exit code of each command.
- Interface. It is for extensions only. The published API has
window.onDidStartTerminalShellExecution,window.onDidEndTerminalShellExecutionwith anexitCode, andTerminalShellExecution.read()for the output. No interface was found for a program outside VS Code. - Control: approval.
- The setting
chat.tools.terminal.autoApproveholds a rule per command pattern. A pattern can be a regular expression. - By default each subcommand of a compound command is matched separately.
- The page says “Command matching is best effort”, and recommends the sandbox to limit file and network access.
- The setting
- Control: sandbox.
- Platforms: macOS 15 or later. glibc Linux and WSL2, with bubblewrap 0.5.0 or later, slirp4netns, util-linux 2.35 or later, and iptables. Windows 11 is marked experimental.
- Network: a list of allowed domains and a list of denied domains. “Denied domains take precedence.”
- A command can ask to run outside the sandbox. Setting
chat.agent.sandbox.allowUnsandboxedCommandstofalsemakes such a request fail.
- Scope. Both controls belong to the Copilot harness. The sandbox page says: “Turning it on doesn’t configure sandboxing for Claude or Codex sessions.” The approvals page says: “The available controls depend on the agent harness selected for your session, not the model.”
- Persistence. A window reload reconnects to the running process. After a restart of VS Code, the content is restored and the process is launched again.
- Rich output. Images work with the Sixel or the iTerm inline image protocol. This is off by default.
- Corrected from the survey. socat is not a requirement on Linux. Automatic approval inside the sandbox is described only for the deprecated Local harness.
iTerm2
[official: iterm2.com]
- Platform. macOS only.
- Record. With shell integration, a past command shows its “return status code, working directory, running time”. The output of the last command can be selected.
- Control.
- An AI chat can be linked to a session. Eight permissions each take one of three levels: never, ask, always. The permissions are: check terminal state, run commands, view contents, control terminal, view manpages, write to the clipboard, write to the file system, and act in the web browser.
- An orchestration chat sees every session. By default it must ask before it types into one.
- Local first. The person can define models. There are presets for OpenAI-compatible providers such as Ollama, and a model on the local network can be reached over plain HTTP.
- Agents the user brings.
- A Claude Code integration installs hooks into
~/.claude/settings.json. It shows the state of each Claude session: working, waiting or idle. - It also arranges a chat, a diff viewer and a code review session side by side.
- The page describes no approval, policy, redaction or audit over what Claude Code runs. Other agents are not mentioned.
- The release notes date it to 3.7.0 beta 2 (2026-05-05). Version 3.7.7 is dated 2026-11-03.
- A Claude Code integration installs hooks into
- Interface. A Python API. The Claude Code integration needs it.
- Persistence. Through tmux integration (
tmux -CC): “when iTerm2 quits or the ssh session is lost, tmux keeps running”. - Rich output. Inline images through OSC 1337, and the kitty graphics protocol.
- Out of date in the survey. The survey mentions a tool-call bug in 3.6.10 that waited for 3.7.0. 3.7 has shipped.
Wave
[official: docs.waveterm.dev, release notes. The newest version is 0.14.5, 2026-04-16]
- Record. Shell integration for bash, zsh and fish came in 0.12.1 (2025-10-20). Wave “tracks when your shell is ready to receive commands, the last command executed, and exit codes”. A Wave “block” is a pane, such as a terminal, a preview or a web page. It is not a command.
- Interface. The
wshcommand, run inside Wave:wsh run,wsh blocks list, saving scrollback, andwsh aito send files and command output to Wave AI.wsh badge(0.14.2) is meant for use with Claude Code hooks. No API or MCP server for outside agents is documented. - Control. Wave AI creates and changes files with the person’s approval, after a diff preview (0.12.2).
- Persistence. Durable sessions came in 0.14.0 (2026-02-10), for remote SSH only. They survive “connection interruptions, network changes, and Wave restarts”.
- Local first. OpenAI-compatible endpoints since 0.13.0: Ollama, LM Studio, vLLM and OpenRouter.
- Corrected from the survey. Record is “partly”, because of the exit codes in 0.12.1. Persistence is “partly”, because local sessions are not durable.
Zellij
[official: zellij.dev, the Zellij repository. MIT]
- Earlier versions. Session resurrection in 0.39, a web client in 0.43, and remote sessions, Windows support and CLI automation in 0.44.
- 0.45 (2026-08-20). Nested sessions, the kitty graphics protocol, and OSC 133.
- With OSC 133, Zellij knows where each prompt starts, where the command and its output begin and end, and the exit status.
- It uses this to select a command with its output, to jump between prompts, to copy the output of the last command, and to mark failed commands.
- 0.46 (2026-11-11).
zellij terminalruns Zellij as a desktop terminal on Linux, macOS and Windows.- The session manager has an inbox. It shows which pane, in which tab, in which session needs the person.
- The inbox is fed by OSC 7501. The states are
idle,working,done,blockedanderror. Ablockedstate can say what is needed:permission,questionorauth. - The inbox also takes desktop notifications (OSC 9, 99 and 777), the bell, and the end of a command from OSC 133 or OSC 633.
- Programs listed as reporting OSC 7501 themselves: Claude Code, Codex CLI, Pi, Crush, OpenCode 2.1.0 or later, GitHub Copilot CLI, Qwen Code, goose and Factory Droid CLI, among others. Gemini CLI, Cursor CLI and Aider do not, and fall back to notifications.
- Session previews, a settings screen and a right-click menu.
- Zellij can run as a library inside another Rust application.
- Interface. No CLI or plugin API is documented that reads the inbox, command boundaries or exit codes. A program posts to the inbox by printing OSC 7501.
- Corrected from the survey.
- A third-party report said the signal can only mean “waiting for a person”. That holds only for programs that do not report OSC 7501.
- The survey could not confirm the library. The official announcement states it.
WezTerm
[official: wezterm.org]
- Record. The Lua method
pane:get_semantic_zones()returns the OSC 133 zones of a pane: Prompt, Input and Output. It is available to the Lua configuration inside WezTerm, not to the command line. The page mentions neither exit codes nor command text. - Interface.
wezterm cli get-textselects by line range only. The other subcommands list, spawn, split and move panes and send text. None is documented to wait for a command or to return an exit code. - Persistence. The mux server keeps sessions alive. This comes from the survey and was not read again.
kitty
[official: sw.kovidgoyal.net/kitty]
- Interface. The remote control protocol,
kitten @.get-text --extentacceptslast_cmd_output,first_cmd_output_on_screen,last_visited_cmd_outputandlast_non_empty_output. These need shell integration.launch --wait-for-child-to-exitwaits for the program it launched and prints its exit code.- This is the closest any outside interface comes to reading by command. It gives no list of commands, and no exit code for a command typed into an existing shell.
- Control.
remote_control_passwordcan be set several times, each with its own allowed actions, and glob patterns are accepted. When a password is used that is not in the configuration, kitty asks the person. This is access control on the interface. It is not a policy on commands. - Rich output. The kitty graphics protocol.
Ghostty
[official: ghostty.org, the Ghostty repository]
- AppleScript since 1.3.0, on macOS only. It creates windows, tabs and splits, and sends text, keys and mouse events. A terminal object exposes its id, name and working directory. No command returns the screen content.
- Issue 12556, which asked for cross-platform command line actions for splits, was closed as not planned on 2026-05-02.
Others
These entries come from the survey and were not read again.
- DomTerm [third party].
- It stores prompts, input, standard output and error output as structured nodes.
- Output can be HTML, and a command can be folded.
- A session can be detached and attached, and can have several windows.
- It has no agent interface.
- Arcan cat9 [official repository, third party]. Commands behave like notebook cells, and later commands can refer to earlier output. It depends on Arcan and is not compatible with the traditional terminal protocols.
- Hyperia [third party, README].
- A fork of Hyper with a Rust side process. The side process holds the pty and serves MCP on
localhost:9800/mcp. It handles consent and identity. - It can use a local Ollama.
- Reports give the number of tools as “30+” and as 63. The consent mechanism is not documented.
- A fork of Hyper with a Rust side process. The side process holds the pty and serves MCP on
- ht [official: andyk/ht, Apache-2.0]. A headless terminal. Input and screen content go over JSON on stdio or over WebSocket.
- Terminal MCP servers [third party]: npcterm, ryamgoh/terminal-mcp, terminai, Forge and mcp-interactive-terminal.
terminal_execin terminal-mcp waits for the command and returns the output, theexit_codeand the new working directory.- terminai filters private content by rule before the screen is read.
- None has a frontend for people.
- herdr and cmux. See Agent terminals.
Findings
- No product has all six.
- Every control layer governs only the product’s own agent. This holds for Warp, VS Code, iTerm2 and Wave. VS Code says so in its documentation. For Claude Code or Codex that the user brings, these terminals add status and notifications. They add no approval, redaction or audit.
- Record and interface are not joined.
- Products that have a record keep it for their own screen, their own agent, or extensions and scripts inside the product: Warp, VS Code, iTerm2, Zellij, DomTerm.
- Products that have an interface for outside agents work on panes and screen text: herdr, cmux, WezTerm, ht, the MCP servers.
- kitty is the nearest exception. It returns the output of the last command.
- Rich output stops at images and HTML. That is the kitty graphics protocol, OSC 1337 in iTerm2, and DomTerm. No protocol was found for semantic blocks, such as markdown or a diff, that an agent CLI could emit.
- Zellij has most of the base these capabilities need.
- It is MIT, has persistent sessions, a web client, exit codes from OSC 133, its own terminal window, and an inbox built on OSC 7501.
- It lacks a control layer, and an agent interface built on the command record.
- It has a WebAssembly plugin system, through which these could be added.
- OSC 7501 is becoming the common status protocol. herdr started it. Zellij builds its inbox on it and lists about ten agent CLIs that report it themselves.
- Knowing who waits is becoming common. Knowing what is asked is not.
- Status from agents the user brings now reaches Warp through plugins, iTerm2 through hooks, Zellij through OSC 7501, and Wave through a badge command.
- OSC 7501 can say that a program is blocked on a permission. It does not carry the command, its impact, or a way to answer.
Corrections to the earlier note
Agent terminals said that nobody has local policy and output redaction, and that approval and secrets in Warp are in the cloud. That is wrong for the agents built into a terminal:
- Warp has permission profiles and secret redaction in the client.
- VS Code has approval rules and a sandbox.
- iTerm2 has permissions per session.
It still holds for the agents users bring.
What this means for cotty
- The gap. A control layer that does not depend on the agent, joined to an interface built on the command record, and semantic rich output.
- Position alone does not give control. An agent the user brings runs commands in its own child processes, so the terminal does not see them. The routes to such an agent are in Capabilities. The first step of one route was tried, in Redirecting an agent’s model endpoint to a local port.
- Where the work is. Persistent sessions, a web client, exit codes and a status inbox already exist in open source. A control layer that can be trusted, and published protocols for the interface and for semantic output, do not.
- OSC 7501. Supporting it as it is, as Positioning says, is confirmed.
- Attention cards. The signal that someone waits is now common. The content of the request, and a way to answer it, are not.
Sources
- Warp: Profiles and permissions, Secret redaction, Third-party CLI agents, Blocks as context, Warp is now open source, GitHub warpdotdev/warp
- VS Code: Approvals, Agent sandboxing, Shell integration, Advanced terminal features, API reference
- iTerm2: AI chat, Claude Code integration, Shell integration, tmux integration, Inline images, Downloads and release notes
- Wave: Release notes
- Zellij: Zellij 0.46, Inbox integrations, Shell integration, Distributions, News, GitHub releases
- Zellij 0.46, third party: AlternativeTo, tproger
- WezTerm: get_semantic_zones, wezterm cli get-text
- kitty: Remote control
- Ghostty: AppleScript, Issue 12556
- DomTerm: opensource.com
- cat9: GitHub letoram/cat9
- Hyperia: dev.to
- ht: GitHub andyk/ht
- terminal-mcp: claudemarketplaces.com
Not covered
- Hands-on use. Every claim comes from documentation or from reports.
- Whether Warp’s permissions or redaction reach a third-party CLI agent in practice. The documentation does not say that they do, and it was not tested.
- tmux control mode, Windows Terminal, Alacritty, foot and Konsole.
- Claude Squad, and Zed’s ACP (Agent Client Protocol).
- The pages for DomTerm, cat9, Hyperia, ht and the MCP servers, which were not read again.