Six terminal-based AI coding agents compete for the developer's shell in 2026 Q2. This note compares their feature surfaces — not their model quality (which changes monthly) but their architectural bets: what contract does each offer between the developer, the agent, and the tools the agent can invoke?

Scope note: this covers the six agents with the deepest CLI surfaces as of April 2026. Other notable tools (Aider, Cursor Agent, Windsurf/Cline, Continue) are omitted either because their primary surface is an IDE extension rather than a terminal CLI, or because their CLI surface is thin enough that comparison at this level of detail is not informative. See Terminal AI Agents 2025 for the broader landscape including IDE-first tools. The 2026-10-15 addendum adds pi (missed here) and DeepSeek Harness (released after Q2), and compares how each harness uses more than one model.

Notable: Copilot CLI defaults to a Claude model, not a GitHub/OpenAI model. Gemini CLI, OpenCode, Goose, and Aider are open source. Goose and Aider are added in the June 2026 update as they have full CLI surfaces comparable to the original six. On FreeBSD 15.0, only Claude Code, Gemini CLI, and Goose run natively; Codex CLI and OpenCode have hard platform checks that block FreeBSD.

Claude Code and Codex CLI have the deepest non-interactive / scripting surfaces. Gemini CLI matches on output format and session management. Copilot CLI is the simplest invocation surface. Kiro's CLI is the least publicly documented.

Gemini CLI has the most mature OS-level sandboxing (four backends). Codex CLI has the cleanest tier model and a standalone codex sandbox tool for running arbitrary commands under the agent's sandbox policy. Claude Code has an OS-level sandbox, but a narrower one: it is off by default (/sandbox or sandbox.enabled) and covers shell commands only — Bash, PowerShell and Monitor, and the processes they start — using Seatbelt on macOS and bubblewrap on Linux and WSL2. Sandboxed commands can write only to the working directory, a per-user temp directory and --add-dir directories; reads stay open machine-wide unless denied; network goes through a domain-filtering proxy. The file tools (Write, Edit), MCP servers and hooks run outside it, under permission rules alone. On 2.1.299 a sandboxed touch ~/x fails with Operation not permitted while the Write tool creates the same path. A blocked command can be retried unsandboxed via dangerouslyDisableSandbox unless sandbox.allowUnsandboxedCommands is false. On a platform the runtime doesn't support, Claude Code runs commands unsandboxed without saying so unless sandbox.failIfUnavailable is set; FreeBSD, where the harness audit below runs it, is not a supported platform. The runtime is open source as @anthropic-ai/sandbox-runtime (srt); see the 2025-10-20 announcement and the sandboxing docs. Worktree isolation remains filesystem-only. Kiro and OpenCode have no sandbox primitives.

Codex CLI's codex mcp-server is unique — it can run as an MCP server, meaning another agent can consume Codex as a tool. Claude Code has the most flexible multi-config story (--mcp-config accepts multiple files).

As of April 2026, Claude Code had the most complete hook surface in this set (7 event types + conditional if hooks); the surface has since grown sharply — official docs list 33 distinct hook events as of October 2026. Copilot CLI is close with 6 hook types including an errorOccurred hook that no other agent has. Kiro has 5 triggers including AgentSpawn and veto via exit code 2 on PreToolUse. Earlier versions of this note incorrectly attributed file-event hooks to Kiro; the official docs show no file-event triggers.

The SKILL.md convention (originated by Anthropic) is now read by Claude Code, Copilot CLI, and OpenCode — the closest thing to a cross-harness standard. Copilot CLI ships no built-in skills but discovers SKILL.md files in .claude/skills/ and .agents/, so project-level skills written for Claude Code work without changes. Gemini CLI has the most developed install/manage surface for skills. Codex CLI and Kiro use AGENTS.md as the project-level agent configuration (different convention, same purpose).

Kiro's auto-generated product.md, tech.md, structure.md are unique — the agent bootstraps its own project understanding on first run. Claude Code has the most structured memory system (typed categories with an indexed MEMORY.md). Gemini CLI and Copilot CLI both have auto-compaction at the 95% threshold.

OpenCode has the broadest provider story (75+ via Models.dev). Codex CLI has the cleanest multi-provider config (config.toml profiles). Claude Code's provider flexibility is limited to API key and Vertex/Bedrock setup wizards. Kiro is locked to AWS Bedrock.

Codex CLI is the only agent with cloud-execution (codex cloud) and MCP-server modes. Copilot CLI supports ACP (Agent Client Protocol) — a competing open standard to MCP for agent interop. Claude Code's /remote-control provides browser-to-terminal bridging.

1. Non-interactive mode with -p / exec — every agent has a script-friendly mode. The shell integration pattern is settled.
2. MCP as the tool-extension protocol — all six support MCP. Configuration format varies (JSON, TOML, settings) but the transport is standard.
3. Project instruction files — CLAUDE.md, GEMINI.md, AGENTS.md, .kiro/steering/. Every agent auto-loads a project-level context document. No standard name.
4. Permission tiering — every agent has at least a read-only / ask / auto-approve ladder. The number of tiers varies (2–6) but the shape is the same.
5. Session persistence + resume — all but OpenCode support continuing a prior conversation.

1. Sandboxing. Gemini CLI and Codex CLI make OS-level sandboxing the agent's boundary (Seatbelt, gVisor, Landlock). Claude Code sandboxes only shell commands, opt-in, and leans on worktree isolation and permission hooks for the rest. Kiro and OpenCode punt entirely.
2. Hooks and lifecycle events. Claude Code and Kiro have rich hook surfaces; the others have basic or no hooks.
3. Skills portability. SKILL.md is becoming a cross-harness convention (Claude Code, Copilot CLI, OpenCode). AGENTS.md is a competing convention (Codex CLI, Kiro, OpenCode).
4. Cloud execution. Only Codex CLI has a cloud-execution tier (codex cloud).
5. Model portability. OpenCode supports 75+ providers; Kiro locks to Bedrock.

The diagram below positions the six agents by their primary architectural bet. Clusters group agents that share the same divergent commitment; edges to the central hub show the feature each agent contributes to the convergent baseline.

Each agent's claims were independently verified by a subagent checking official documentation or local --help output. This section records what was checked, what version, and what was corrected.

Correction (2026-10-15): the April check of Claude Code used claude --help, which does not surface the /sandbox command or the sandbox.* settings, so the sandbox rows read "-" and the prose said there was no OS-level sandbox. A reader pointed out the 2025-10-20 announcement; the rows, the prose and Divergent bets item 1 are corrected, with the write boundary reproduced on 2.1.299 (see the sandbox-isolation drawer). The landscape diagram still clusters Claude Code by its primary bet, hooks.

All corrections have been applied to the tables above. The note is accurate as of the versions listed; flag surfaces change frequently and should be re-verified quarterly.

Channel note: Claude Code 2.1.114 is from the npm/=latest= channel (FreeBSD). Homebrew's stable channel ships 2.1.98 as of the same date. All flags in the comparison tables are present in both versions; the gap is power-user flags (--effort, --from-pr, --chrome, etc.) that the note intentionally omits as out of scope.

Vetted against FreeBSD nexus 15.0-RELEASE amd64 (the machine that publishes this site).

A reader called the list above too vendor-shaped: one flagship per large lab, and no pi or DeepSeek. Fair. The literature's word for what this note compares is harness — everything around the model: prompts, tools, skills, memory, the control loop — and evaluations increasingly report results per harness–model pair, because in some benchmarks changing the harness moves scores nearly as much as changing the model. Scoping by "has a terminal CLI" and by vendor is what left out the two harnesses below. The tables above stay as the Q2 snapshot; a Q4 note will compare harnesses directly.

Each subsection below was checked by a verification agent against the npm registry, the GitHub API, shipped docs and changelogs, and --help / --version runs; no model requests were sent.

• Claude Code Workshop 2026 — deep-dive on Claude Code's surface.
• Claude Code Features — 2026 Q2 — speculative notes on Claude Code's direction.
• Terminal AI Agents 2025 — prior landscape survey; this note is the Q2 2026 update.
• Cloudflare Agents Week 2026 — where the agent infrastructure (sandbox, memory, identity) sits relative to these CLI surfaces.