llms.txt@docs/public · diff
git:20260428.3b4ea56 to git:20261117.6b20bc9
4 added, 6 removed. Audit B to B.
What this update changes in the checks
No check changed: every fixed check gives the same answer before and after this update.
A tripped check is a label, not a verdict: read the matched text and the lines below. This week's updates that newly trip a check.
# Cave CLI — terminal coding agent
- > Cave 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. Three-layer Cave Mode
- > compression gives ~85% reduction on tool output and $1.70-$6.92 of
- > Anthropic spend saved per typical session.
+ > Cave is a terminal coding agent in a single MIT-licensed binary, frozen
+ > in August 2026. Three-layer Cave Mode compression shrinks tool output
+ > (measured offline on the bundled fixtures).
This file is the LLM-friendly entry point to Cave's documentation. Each
linked page is plain markdown with a "Copy for LLMs" button at the top
that strips the navigation chrome.
## Install
```
curl -fsSL https://cave.sh/install | bash
```
## Top-level docs
- [Quickstart](https://cave.sh/docs/getting-started/quickstart): zero-to-first-prompt in 30 seconds
- [Install](https://cave.sh/docs/getting-started/installation): macOS, Linux, Windows, Docker, Homebrew, npm
- [Auth & Providers](https://cave.sh/docs/getting-started/auth): 20+ providers, 6 OAuth flows
- [Models](https://cave.sh/docs/getting-started/models): default models, thinking levels, architect/editor split
## Reference
- [Tools](https://cave.sh/docs/reference/tools): built-in tools, Cave Mode compression, cost transparency
- [Slash Commands](https://cave.sh/docs/reference/slash-commands): built-ins + authoring markdown commands
- [Skills](https://cave.sh/docs/reference/skills): markdown skills loaded by description match (Claude Code-compatible)
- [Subagents](https://cave.sh/docs/reference/subagents): worktree-isolated parallel agents via Task tool
- [Memory (cavemem)](https://cave.sh/docs/reference/memory): cavemem integration, episodic-to-semantic consolidation
- [MCP](https://cave.sh/docs/reference/mcp): client + cave-as-server, stdio/HTTP/inproc transports
- [Hooks](https://cave.sh/docs/reference/hooks): 12-event lifecycle (Claude Code settings.json schema)
- [Permissions](https://cave.sh/docs/reference/permissions): five modes, Seatbelt/Landlock/Restricted Tokens
- [Plan Mode](https://cave.sh/docs/reference/plan-mode): read-only exploration, structured plan, /act to execute
- [Daemon](https://cave.sh/docs/reference/daemon): cave serve, multi-client attach, worker mode
- [Recipes](https://cave.sh/docs/reference/recipes): YAML pipelines, Goose-style schema
## Migration
- [From Claude Code](https://cave.sh/docs/migration/from-claude-code): paste your config, it works
- [From Codex](https://cave.sh/docs/migration/from-codex): AGENTS.md, .mcp.json, sandbox parity
- [From Aider](https://cave.sh/docs/migration/from-aider): repo map and edit-format-per-model
## Cookbook & API
- [Cookbook](https://cave.sh/docs/cookbook): GitHub Actions, multi-agent review, daemon pair programming
- [Comparison](https://cave.sh/docs/comparison): Cave vs Claude Code / Codex / Aider / Crush / opencode
- [Troubleshooting](https://cave.sh/docs/troubleshooting): install, auth, sessions, tools, hooks, MCP
- [API](https://cave.sh/docs/api): SDK, JSON-RPC, OpenAPI, extension API
## Differentiators (one-line each)
1. Cave Mode compression: 3 layers, ~85% reduction on tool output
- 2. 20+ providers + 6 OAuth flows (Claude Pro, ChatGPT Plus, Copilot, Gemini, Antigravity)
+ 2. 20+ providers, OAuth for ChatGPT Plus, Copilot, Gemini, Antigravity
3. Session branching (`/tree`, `/fork`) — no major competitor has this
4. Plan mode + 7 parallel worktree-isolated subagents
5. Native MCP client + Cave can serve as MCP server
6. Native sandbox: macOS Seatbelt, Linux Landlock, Windows Restricted Tokens
7. Aider-grade repo map (PageRank) + edit-format-per-model
8. Daemon (`cave serve`) survives SSH drops; multi-client attach
9. Shadow-git checkpoints + `/rollback N`
10. cavemem-backed memory with episodic-to-semantic consolidation
11. Claude Code authoring formats (settings.json, commands, skills, agents) — paste-and-go compat
12. MIT licensed; no telemetry by default
## Repo
- Source: https://github.com/JuliusBrussee/caveman-cli
- Issues: https://github.com/JuliusBrussee/caveman-cli/issues
- Discord: https://discord.gg/cave-cli
- Releases: https://github.com/JuliusBrussee/caveman-cli/releases