// Your machine. Your models. Your code.
Proverbs is a fully self-contained AI coding system — built-in inference server, browser IDE, persistent memory, local model training, and 70+ commands. No Ollama. No cloud. No subscriptions. Nothing leaves your machine.
// What it does
Built-in llama.cpp inference server, browser IDE, persistent project memory, and 70+ slash commands.
Drop a .gguf model into ~/.proverbs/models/ and start coding.
No Ollama. No external services. No internet required after setup.
/ide./remember saves facts about your project permanently. Auto-injected into every session so Proverbs never forgets your stack, conventions, or decisions./train tokenizer, /train embed, /train prune./goal <condition> — Proverbs loops autonomously until the condition is met. Like Claude Code's agentic mode, fully local./backend auto finds it automatically./cwd change./scan detects your stack, import aliases, conventions, and git history. Injected into every prompt automatically.~/.proverbs/plugins/ to add custom tools the model can call — same interface as built-in tools.ANTHROPIC_API_KEY or OPENAI_API_KEY — Proverbs calls them directly when you want cloud quality. No proxy needed./branch saves a named conversation checkpoint. /checkout restores it. Never lose a working state./rewind N rolls back N conversation turns. /undo restores any file write. Full time-travel for both chat and code./agents spawns 3 parallel sub-instances each attacking the task from a different angle. /fork runs a sub-task non-blocking./security runs a full OWASP audit — injection, XSS, hardcoded secrets, missing auth — entirely on local inference. /review high for code review./sandbox on blocks destructive bash patterns before they reach your shell. Rate limiting, auth middleware, audit log — all built in./profile shows live latency, tokens/sec, cache hit rate, memory usage, and backend stats. Know exactly what's happening under the hood./update downloads the latest release and hot-swaps itself with zero downtime. No npm, no package manager required.// Browser interface
Run /ide from the terminal to open a full development environment at localhost:4400.
File tree, tabbed editor, integrated AI terminal, git panel — all talking to the same local inference server.
// Get started
One command. The built-in server starts automatically — just drop a .gguf model into
~/.proverbs/models/ and run proverbs. No Ollama. No external setup.
Node.js 18+ and installs it if missing via HomebrewProverbs inference server (llama.cpp) — no Ollama requiredqwen2.5-coder-1.5b) so you can run immediately — swap for any .gguf anytimeproverbs command globally. Run proverbs to chat or proverbs /ide for the browser IDE.Node.js LTS silently if not presentProverbs inference server — no Ollama requiredproverbs CLI globally and registers it in PATHproverbs to chat or proverbs /ide for the browser IDE..gguf model into ~/.proverbs/models/ — or let the installer fetch onenpm install -g proverbs-aiproverbs in any project folder, or proverbs /ide for the browser IDE.// How it stacks up
Claude Code is excellent — Proverbs is what you run when you want no cloud dependency, full data privacy, and the ability to train your own model.
| Feature | Proverbs | Claude Code |
|---|---|---|
| Cost | Free forever | Per token |
| Works offline | ✓ | ✗ |
| Data privacy | 100% local | Cloud processed |
| Browser IDE | ✓ (VS Code-style, /ide) | ✗ |
| Persistent project memory | ✓ (/remember, auto-injected) | CLAUDE.md only |
| File read / write / patch | ✓ | ✓ |
| Bash execution | ✓ | ✓ |
| Git integration | ✓ | ✓ |
| Web search | ✓ (no API key) | ✓ |
| Session persistence | ✓ (save/resume/branch) | ✗ |
| Train your own model | ✓ (DPO, GRPO, LoRA, SimCSE) | ✗ |
| Prefix KV cache | ✓ (40–80% TTFT reduction) | — |
| Goal / autonomous mode | ✓ (/goal condition) | ✓ |
| Code review | ✓ (/review low|med|high) | ✓ |
| Conversation branches | ✓ (/branch, /checkout) | ✗ |
| Plugin system | ✓ (JS files in plugins/) | MCP servers |
| Self-critique loop | ✓ | ✗ |
| Test runner integration | ✓ (jest, vitest, pytest) | ✗ |
| Security auditor | ✓ (OWASP, local) | ✗ |
| Multi-agent mode | ✓ (/agents, /fork) | ✓ |
| Inference backend | Built-in llama.cpp · LM Studio · Jan · network server | Anthropic cloud only |
| Context window | 32k–128k (model dependent) | 200k |
| Reasoning quality | Good (qwen2.5-coder 1.5b–7b, codestral 22b) | Excellent (Claude 4) |
| Cloud model fallback | Direct Anthropic / OpenAI | Anthropic only |