# Pair-program interactively in a terminal with approvals

> Pair-program interactively in a terminal with approvals: 13 companies fully do it, 25 with a limit, 101 carry the tag. Every listing in Agent harnesses with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/agent-harnesses.pair-in-terminal/
- Markdown: https://www.anchorterminal.com/niches/agent-harnesses.pair-in-terminal/index.md (~3,300 tokens)
- Slim: https://www.anchorterminal.com/niches/agent-harnesses.pair-in-terminal/index.min.md (~680 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/agent-harnesses.pair-in-terminal/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Four or more companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 13 | 25 | 101 | 15 | 8 |

## What counts as done

The requested change is applied and reviewable as a diff, each risky command was shown before it ran, and the session ends when the person says so.

How agents ask for it: add a priority field to tasks; refactor this module; find the bug in auth.py; explain this codebase

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Antigravity CLI (Google LLC)**: The Default preset asks before sensitive operations outside the terminal sandbox, which is on by default with no network, and Deny beats Ask beats Allow rules.
- **Crush (Charmbracelet, Inc.)**: By default Crush asks for permission before running tool calls, and permissions allow lists which tools may skip the prompt.
- **Gemini CLI (Google LLC)**: general.defaultApprovalMode defaults to default, which prompts for approval; YOLO can only be turned on from the command line and security.disableYoloMode can forbid it.
- **GitHub Copilot CLI (GitHub, Inc.)**: Tools prompt for permission unless listed with --allow-tool or --allow-all-tools, deny rules still apply, and the local sandbox can be turned on with /sandbox enable.
- **Junie CLI (JetBrains s.r.o.)**: The action allowlist runs matching commands automatically and asks for the rest, including each part of a chained or nested command, and Brave Mode is opt-in.
- **Kimi Code CLI (Moonshot AI (the global Kimi terms name NOVASCENT PRIVATE LIMITED, Singapore))**: Default `manual` permission mode asks before tool calls, the dangerous-command guard covers risky shell, `.env` and git control paths, and `--yolo` or `--auto` relax it only when asked for.
- **Kiro CLI (Amazon Web Services, Inc.)**: Most tools default to `ask`, `permissions.yaml` rules use deny > ask > allow, and shell commands are parsed per sub-command before matching.
- **mini-swe-agent (No legal entity named. The MIT licence names Kilian A. Lieret and Carlos E. Jimenez)**: `confirm` mode is the default: the model proposes one bash action and the person confirms or rejects each one.
- **MiniMax Code (Nanonoble Pte. Ltd.)**: TUI modes Ask, Auto and Full access; Auto classifies routine actions and asks when risk is high, Ask confirms every tool action.
- **Mistral Vibe CLI (Mistral AI (RCS Paris 952 418 325))**: Default `accept-edits` agent auto-approves edits and asks for shell and other sensitive tools; `ask` asks for everything, and calls outside the working directory prompt.
- **Muse Code (Meta Platforms, Inc.)**: Approval and an OS sandbox are on by default; the Auto-review profile lets a judge clear eligible requests and falls back to asking, and dangerous shell patterns always stop for review.
- **Mirror CLI (Reflection AI)**: Approval mode defaults to `approve` (manual), remembered approvals live in `tool_approvals.json`, and `/worktree` and `/sandbox` isolate work.
- **Rovo Dev CLI (Atlassian Pty Ltd)**: Most tools default to `ask` with Once, Session or Always choices, bash commands get per-command permissions, and worktree mode keeps edits out of the checkout.
- **Warp (Warp Agent CLI and the Oz Agent API) (Denver Technologies, Inc. (d/b/a Warp))**: The CLI shows inline request cards for commands and edits, with profile allowlists and denylists, though `--auto-approve` bypasses the denylist unless a setting is changed.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Aider (Aider AI LLC)**: Edits land as reviewable git commits and shell commands are suggested for confirmation, but users want tool calling and find it a poor fit without git knowledge (acc-33), and it is unmaintained.
- **Auggie CLI (Augment Computing, Inc.)**: Allow, deny, ask-user and webhook rules exist with deny winning, but what the CLI does with an unmatched tool is undefined and the MCP page says Auggie runs commands and tools automatically.
- **AXLR (Underpass AI)**: The console shows approval cards, but after setup approvals.autonomous is on and exec_sandbox off, so commands run with the person's rights unless the settings are changed first.
- **Claude Code (Anthropic, PBC)**: Since 2.1.283 interactive sessions start in auto mode, where a classifier rather than the person reviews most actions; the Bash sandbox is off by default, and skip-permissions mode deleted a home directory (acc-14).
- **CLI Manager (solhun (Jeong Kyeonghun))**: It hosts Claude Code, Codex and Gemini CLI terminals side by side with worktree workspaces, but approvals and diffs belong to the child harness.
- **Cline (Cline Bot Inc.)**: A prompt runs in yolo mode by default; pass --auto-approve false or --act to get act mode, which asks before tools; in ACP mode the default is already false.
- **Conductor (Conductor)**: Each task gets a workspace, branch, terminal and diff for Claude Code, Codex, Cursor or OpenCode, but no new downloads since 4 November 2026 and the local app winds down in early 2027.
- **Cursor CLI (Anysphere, Inc.)**: Shell(...), Read and Write permission tokens in cli.json allow or deny actions and --sandbox toggles a sandbox, but the docs do not state what runs without approval by default in an interactive session.
- **Droid CLI (The San Francisco AI Factory, Inc.)**: Autonomy Off, Low, Medium and High gate what runs without asking and command rules can block, but the sandbox is off until sandbox.enabled is set and included usage is stated only as rolling limits without numbers.
- **Emdash (General Action, Inc.)**: Each task runs a CLI agent in its own worktree with a diff to review and a pull request button, but approvals and sandboxes are the child agent's.
- **goose (Agentic AI Foundation (Linux Foundation))**: Manual and Smart Approval modes exist, but Autonomous mode, which approves everything, is the default.
- **Harness (Autonomous, Inc.)**: Every agent runs in a real tmux pane the person can read and type into, across machines, but approvals and diffs are the child harness's.
- **Kilo Code CLI (Kilo Code Inc. (the terms are a contract with Anaconda, Inc. on behalf of itself and its affiliates, including Kilo Code Inc.))**: Allow, ask and deny rules per tool exist and unmatched checks default to ask, but most tools ship with "*": "allow", so only exceptions such as .env reads prompt out of the box.
- **Letta (Letta)**: The interactive CLI starts in `unrestricted`, auto-allowing tools; the person must switch to `standard` or `strict` before risky commands are shown.
- **OpenAI Codex (OpenAI OpCo, LLC)**: The Auto preset edits and runs in the workspace and asks outside it, with `/permissions` and worktrees, but acc-22 found Codex edits through scripts unreviewable as they happen and acc-26 ran out of Plus quota mid-task.
- **OpenCode (Anomaly Innovations, Inc.)**: The default policy is `allow` for every action except `.env` reads, so nothing is shown before it runs until the user adds `ask` rules.
- **Orca (Lovecast Inc.)**: Orca hosts Claude Code, Codex and any CLI agent in chat or raw terminal; approvals belong to the agent below, and the README is the only source read.
- **Paseo (Mohamed Boudra)**: The app and `paseo attach` stream an agent and plugins answer permissions, but approval behaviour is the harness's and was not documented in the pages read.
- **Prime Agent (Prime Intellect, Inc.)**: The TUI executes model-written Python and shell with the user's rights and asks nothing first, so risky commands are not shown before they run.
- **Qwen Code (Qwen team, Alibaba Group (no legal entity is named in the repository or the docs))**: The default is Auto, where an LLM classifier approves safe actions, so a person sees only what the classifier escalates; acc-25 ran it on a local model successfully.
- **Sculptor (Imbue, Inc.)**: The person prompts in the app and reviews diffs, but Claude Code runs sandboxed with tool permissions auto-approved, so commands are not shown before they run.
- **T3 Code (T3 Tools Inc.)**: Pairing happens in the desktop, web or phone app with Supervised, Auto-accept edits, Auto and Full access modes, but the initial default is Full access.
- **Traycer (Traycer AI)**: A desktop app, not a terminal; new conversations default to Full access and Auto mode's judge is experimental.
- **Unstoppable (Unstoppable Domains, Inc.)**: A desktop IDE rather than a terminal; every chat starts in Full access until the person picks Ask or Accept edits, though deny rules still apply.
- **Verdent (Verdent AI, Inc.)**: A desktop app, not a terminal; Agent Mode is the default and some actions prompt for approval, but which ones is not listed.

## Couldn't check

The docs were closed to us, or silent. These carry no mark either way.

- **Parallax (Conifer Solutions, Inc.)**: The macOS app that runs Claude Code and Codex side by side has no documentation; only the review service pages were read.
- **Qoder CLI (BRIGHT ZENITH PRIVATE LIMITED)**: docs.qoder.com answered 502 from the fetch proxy on three attempts and a DNS failure through WebFetch, so no page was read.

- not this job: Agent Zero (Agent Zero, s.r.o.): A web UI, not a terminal, and Tool Access defaults to allow with no approval before shell commands.
- not this job: Amp (Amp Frontier Corporation): The tools page says Amp does not ask for approval before running tools and that control needs a custom policy plugin, so risky commands are not shown before they run.
- not this job: Clawboo (Sanreds): Clawboo is a web dashboard and board, not a terminal, and its coding runtimes start with their own permission prompts turned off.
- not this job: DeepSeek Harness (dsh) (DeepSeek (the LICENSE names DeepSeek, and no legal entity is named in the repository or the docs)): dsh's interactive surface is a local web app or an ACP editor session, not a terminal UI.
- not this job: dsh-superagent (DeepSeek) (DeepSeek (the LICENSE names the DeepSeek Harness Team; DeepSeek Harness's terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)): A daemon that answers issues and pull requests; no terminal pairing.
- not this job: Devin (Cognition AI, Inc.): Devin works in its own cloud VM; the terminal products are separate and not graded here.
- not this job: Pi (Earendil Inc.): The security page says Pi does not ask for approval before every tool call and has no permission system, so a risky command is not shown before it runs.
- not this job: Harness² (Google LLC): Not interactive; executors run with approvals and sandbox bypassed on benchmark tasks.
- not this job: HarnessRouter (Lumentree Corporation): A server-side API; no terminal session.
- not this job: Jules (Google LLC): A cloud agent with a CLI dashboard, not a terminal pair programmer, and read-only since 11 November 2026.
- not this job: OpenAI Agents API (OpenAI OpCo, LLC): It is a hosted API; there is no terminal client or interactive approval loop.
- not this job: OpenHands (All Hands AI): The CLI is no longer maintained; the interactive surface is a web UI, and acc-23 saw it edit README without asking in build mode.
- not this job: OpenMausBot (Supamaus Software Private Limited): It is a chat app with approval cards, not a terminal pairing session.
- not this job: Paperclip (Paperclip Labs, Inc.): It is a web control plane; the terminal belongs to the harness it drives.
- not this job: Vibe-Trading (HKUDS): Not a coding harness; the bash tool and generated backtests run without per-action approval.

## What stops agents finishing

- weekly subscription limits stop work mid-task (acc-26, acc-21)
- skip-permissions mode plus a bad shell command deletes files outside the repository (acc-14)
- edits made through scripts cannot be reviewed as they happen (acc-22)

## Around it

Before: code-hosting.clone. After: code-hosting.commit, agent-harnesses.fix-failing-ci.

## Where this comes from

Weight 15 of 100 in Agent harnesses: 8 of 33 accounts (acc-20, acc-21, acc-22, acc-26, acc-27, acc-32, acc-33, acc-14) are about daily interactive use or its failures

As data: https://www.anchorterminal.com/api/v1/use-cases/agent-harnesses.pair-in-terminal.json, or the MCP tool get_use_case with id agent-harnesses.pair-in-terminal. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
