writing/cli-coding-agents-and-launchpad.md

CLI coding agents and Launchpad

2026-09-01

“CLI coding agent” in 2026 usually means an agentic edit/run loop in the terminal — not Tab autocomplete, and not another VS Code chat panel. The agent reads the repo, edits files, runs shell tools, and iterates behind permission gates.

Most people stall in the gap: they install three CLIs, paste a “best of 2026” matrix into Notes, and still open the IDE for every change. Stop that. Pick one harness that matches how you already pay for models, run one real repo task, and only then worry about the workspace around it.

This post maps the terminal-agent landscape and where Launchpad fits as a macOS home for those CLIs — not as another LLM.

What a CLI coding agent is (2026)

Terminal agent loop vs IDE Tab vs extension

Harness \> model

The unit of product is the harness — tools, permissions, context packing — not only which logo is on the model card. Open Interpreter’s 2026 framing is explicit about emulating harnesses so lower-cost models can behave like lab CLIs (OI blog). Lab CLIs (Claude Code, Codex, Antigravity) sell the opposite: stay inside one polished harness tied to a subscription.

What this post is not

Not a Copilot ranking. Not a no-code “AI coworker” bake-off. SEO listicles that mash IDE agents, extensions, and terminal agents into one scoreboard are for naming only — not for stars, SWE-bench %, or M&A rumors.

Landscape map (categories, not a tier list)

Lab CLIs

• Claude Code: Anthropic agentic coding across terminal CLI, IDE, desktop, and web (product, docs). Asks before file/command changes.
• Codex CLI: OpenAI coding agent in the terminal (plus IDE/app/cloud siblings); ChatGPT plan or API key; interactive and codex exec (GitHub, docs, upgrades post). Distinguish CLI vs Web vs IDE extension.
• Antigravity CLI (ex–Gemini CLI): Google’s terminal agent; invoke with agy; shares harness/settings with Antigravity 2.0 (product, intro, GitHub). On May 19, 2026 Google announced the transition; Jun 18, 2026 consumer Gemini CLI paths for Google AI Pro/Ultra and free/individual Code Assist stop serving; Enterprise/Standard and paid API key paths are called out as continuing (Google Developers Blog). Keep “Gemini CLI” in search copy as history; ship Antigravity for the current product.

Model-agnostic OSS

• Aider: OSS terminal pair programmer; maps the repo; edits files; auto-commits; undo via git; model-agnostic (aider.chat, GitHub).
• OpenCode: OSS AI coding agent (TUI + desktop + IDE extension); multi-provider (opencode.ai, docs, GitHub). Homepage MAU/star claims are vendor — re-verify before citing.
• Goose: Block-origin OSS general-purpose agent; desktop + CLI + API; deep MCP (goose-docs.ai, GitHub).
• Open Interpreter: OSS terminal agent; Codex-fork lineage; harnesses for low-cost/open models; CLI + desktop (openinterpreter.com, 2026 blog).

Editor-tied CLI

• Cursor CLI: Cursor agent from the terminal (agent binary); modes Agent / Plan / Ask; print/automation for scripts/CI; sandbox controls; same MCP as the editor; & prefix hands off to Cloud Agent (CLI overview).

Workspace / ADE hosts

• Warp: ADE with its own Warp Agent and first-class third-party CLI agent hosting (docs, CLI agents overview). Two stories: Warp’s agent vs Warp as host for Claude/Codex/OpenCode/Amp/Cursor/Goose/Antigravity/etc.
• Launchpad: terminal-first macOS workspace for CLI agents — not itself an LLM agent. Native Apple Silicon app (Tauri v2); PTY tabs/splits; project-scoped file browser that never writes cd into the terminal (agent-safe); git panel; CodeMirror editor. Built as a comfortable home for Claude Code, Aider, and friends (GitHub, docs, adamwickwire.com).

Amp as agent + orb environment

Amp is a coding agent plus environment: web orbs (per-thread machines), macOS/iOS apps, and CLI sharing threads (Amp manual). The CLI is one surface, not the whole product.

Continue — consolidation hygiene

Continue was an OSS VS Code/JetBrains assistant. continue.dev now says Continue joined Cursor; the Apache 2.0 codebase remains on GitHub (continuedev/continue, The New Stack). Do not pitch it as a living CLI competitor. Footnote for landscape memory only.

Market read: The field splits into (1) agent CLIs, (2) model-agnostic OSS CLIs, (3) workspace/ADE layers that host agents. Launchpad competes on comfortable shell + files + git, not model harness. Warp is the closest commercial cousin on the workspace axis.

How to choose

Pick the row that matches this week’s pain. Run one real task in one repo. Expanding the desk before that is collecting apps, not shipping.

Patterns that matter in practice

Permissions, sandboxes, approve-before-shell

Claude Code asks before file and command changes (product). Cursor CLI documents sandbox controls and a sudo path that keeps the password off the model (CLI docs). Shell agents still inherit your credentials. Treat approve/deny as part of the workflow, not friction to disable on day one.

Git as undo log vs git as human UX

Aider makes git the undo log: edit, commit, revert (aider.chat). Launchpad’s git panel is human UX — stage/commit/push/amend/rebase without typing flags while an agent runs in the PTY (README, docs). Complementary, not competing.

MCP / skills extension layer

Cursor CLI shares MCP with the editor. Claude Code extends via tools and MCP servers. Goose advertises a large MCP extension set. Antigravity CLI surfaces plugins, MCP, skills, hooks, and subagents. Amp adds MCP alongside built-in tools. Same trust rules as any DIY MCP post: install less, review more.

Headless vs interactive

Interactive TUI for exploration. Print/exec/CI modes for automation. Do not confuse “agent can run unattended” with “agent should run unattended on prod.”

Cloud / async handoff

Cursor CLI’s & prefix hands off to a Cloud Agent. Amp orbs keep threads across web/apps/CLI. Codex narrates local-to-cloud via ChatGPT account. Async handoff is a market pattern — not a Launchpad feature.

Risks without scare-copy

Renames and sunsets

Gemini CLI → Antigravity with a hard consumer date. Continue acquired and wound down as a standalone recommend. Products rename; your muscle memory does not. Prefer primary docs over listicles that still rank dead tools.

Do not trust listicle benchmarks

Secondary “best CLI 2026” posts disagree on leaders and paste volatile stars and percentages. Useful for names. Useless as scoreboards.

Agent + credentials = blast radius

A terminal agent with your shell is powerful because it is dangerous. Keep approval gates. Sandbox when the product offers it. Do not paste secrets into prompts “just this once.”

Workspace lock-in vs agent lock-in

Launchpad is macOS / Apple Silicon today — platform limit, stated plainly (README). Lab CLIs lock you to a harness and a bill. Those are different locks. Choose knowingly.

Workspace ≠ intelligence

Launchpad adds no LLM. Quality depends on which CLI and model you run inside it (adamwickwire.com, README). A prettier PTY will not fix a weak harness.

Launchpad as the Mac home for CLI agents

Launchpad’s job is to host the agent without fighting it.

From the primary sources (GitHub, docs, adamwickwire.com):

• Terminal-first macOS app with PTY tabs and splits
• Project-scoped file browser that stays visual — it does not inject cd into the agent session
• Git panel for human git UX while the agent works
• Built as a comfortable home for Claude Code, Aider, and other CLI agents
• Distributed as a notarized Apple Silicon DMG (and build-from-source); check releases for the current tag

Minimal proof path: install the DMG → open a project → spawn a PTY → run claude or aider (or another CLI you already trust) → browse files and use the git panel without polluting the agent session → optional split panes for a second agent tab.

That is the product claim. Screenshots and a specific demo task belong on the site and in the repo — not invented here.

Bottom line

CLI coding agents are terminal harnesses that edit and run. Lab CLIs win when you already pay that lab. OSS CLIs win when you need BYOK or local models. Cursor CLI wins when you already live in Cursor. Workspaces like Launchpad and Warp win when the pain is the shell around the agent, not the model itself.

Pick one CLI that matches your billing. Give it one real repo task today. If the PTY keeps getting stepped on, put it in Launchpad (Mac) or Warp (cross-platform). Prefer primary docs over another best-of listicle.

Sources

Cited inline. Prefer product/docs primaries for agent CLIs; Launchpad claims only from the WalrusQuant README, docs site, and adamwickwire.com. Secondary CLI roundups are for naming. Do not invent stars, SWE-bench %, or deal prices.