Caveman Code vs Claude Code, Codex, Aider, Crush, opencode

May 18, 2026 · View on GitHub

This is the comparison table from the v2 master plan, kept current as features land. The pitch is short:

Caveman Code is the only terminal coding agent that beats Claude Code on cost, Aider on context selection, Codex on provider flexibility, and opencode on session UX — in a single MIT-licensed binary.

Capabilities

AxisCaveman Code v2Claude CodeCodexAiderCrushopencode
Token compression (3-layer Caveman Mode)yes (unique)nonorepo map onlynono
20+ provider OAuth (Claude Pro / ChatGPT / Copilot / Gemini)yes (unique)Anthropic onlyChatGPT onlyenv keys onlysubsetenv keys
Session branching + forkyesnofork onlygit onlynono
Native MCPyesyesyesnoyesyes
Native sandboxyespartialyes (best-in-class)nopartialpartial
Plan modeyesyesyesarchitectnoyes
Repo map (PageRank)yesnonoyes (best-in-class)nono
Edit-format-per-modelyesnonoyes (best-in-class)nono
Worktree-isolated subagentsyesyesyesnonono
Daemon / multi-clientyesnoyes (app-server)nonoyes (best-in-class)
Shadow-git checkpoints + /rollback Nyesnonogit onlynono
Containerized parallel sessionsyesnonononono
Cost transparency (per-msg $)yespartialpartialyes (best-in-class)nono
MIT open sourceyesclosedApacheApacheFSLMIT

Where each agent shines

  • Claude Code — first-party Anthropic, opinionated UX, polished out-of-box. Best if you only use Claude and don't care about cost.
  • Codex — OpenAI's terminal agent. Excellent sandbox primitive ("sandbox-as-utility"). Single-vendor by design.
  • Aider — pioneer of repo map + edit-format-per-model. Strongest at large-codebase context selection. Less ergonomic interactive UX.
  • Crush — fast, polished TUI (Charm). Mid-session model swap. Smaller ecosystem.
  • opencode — strong daemon / multi-client story. Newer; ecosystem still maturing.
  • Cave — borrows the best of all five and adds Caveman Mode compression + 20+ provider OAuth as native, unique differentiators.

Tokens — the headline

25-task MicroBench, gpt-5.5 on both sides, xhigh reasoning (2026-05-18):

AgentFresh tokensPass rateCost
Codex CLI1,010,18515/25 (60%)$0 (codex sub)
Caveman Code524,70314/25 (56%)$1.78

1.93× fewer tokens for ~equivalent pass rate. Reproduce in one command:

npx tsx research/evals/run-honest-bench.ts --tools caveman,codex

Raw CSV and per-task logs live in research/results/. Methodology spawns each CLI as a real child process — no SDK shortcuts.

Format compatibility

Caveman Code is a superset of Claude Code's authoring formats. Concretely, you can paste these directly into ~/.cave/:

  • ~/.claude/settings.json~/.cave/settings.json (hooks, permissions, statusLine identical schema)
  • ~/.claude/commands/*.md~/.cave/commands/*.md
  • ~/.claude/skills/<name>/SKILL.md~/.cave/skills/<name>/SKILL.md
  • ~/.claude/agents/<name>.md~/.cave/agents/<name>.md
  • .mcp.json (Codex / Claude Code standard) is read at the project root

See migration from Claude Code for the step-by-step.

Caveat — these comparisons evolve

Claude Code, Codex, Crush, and opencode all iterate weekly. We pin our compatibility target to Claude Code v2.1.119 schemas with a CI delta check; tracking the others is best-effort. If you spot drift, open an issue.