| icon |
title |
details |
link |
linkText |
📦 |
Caveman Mode compression |
3-layer compression of prompts, tool output, and file reads. ~85% reduction on tool output. $1.70–$6.92 saved per session vs Claude Code. |
/reference/tools |
How it works |
|
| icon |
title |
details |
link |
linkText |
🔑 |
20+ providers, 6 OAuth flows |
Claude Pro, ChatGPT Plus, GitHub Copilot, Gemini, Antigravity, plus every major API. One CLI, every backend. |
/getting-started/auth |
Authenticate |
|
| icon |
title |
details |
link |
linkText |
🌳 |
Session branching |
Fork at any turn, navigate the tree, never lose context. Auto-save JSONL sessions per cwd. |
/reference/slash-commands |
/tree, /fork |
|
| icon |
title |
details |
link |
linkText |
🧠 |
Plan mode + subagents |
Read-only exploration, structured plans, then 7 parallel worktree-isolated subagents to execute. |
/reference/plan-mode |
Plan and act |
|
| icon |
title |
details |
link |
linkText |
🛡️ |
Native sandbox |
macOS Seatbelt, Linux Landlock, Windows Restricted Tokens. Permission modes cycle on Shift+Tab. |
/reference/permissions |
Permission profiles |
|
| icon |
title |
details |
link |
linkText |
🔌 |
MCP everywhere |
stdio + Streamable HTTP + in-process. ToolSearch defers schemas. caveman-code can also serve as an MCP server. |
/reference/mcp |
MCP servers |
|
| icon |
title |
details |
link |
linkText |
🪝 |
Hooks (Claude Code-compatible) |
12 lifecycle events with the exact settings.json schema as Claude Code. Paste your config and it Just Works. |
/reference/hooks |
Hook reference |
|
| icon |
title |
details |
link |
linkText |
💾 |
Memory via cavemem |
Native integration with cavemem. Episodic→semantic consolidation. Bridges Claude Code's MEMORY.md. |
/reference/memory |
Memory layer |
|
| icon |
title |
details |
link |
linkText |
🆓 |
MIT, open source |
No telemetry by default. No vendor lock-in. Self-host the daemon. Beat Claude Code on cost in a single binary. |
/comparison |
vs the field |
|