Larb

English

ไทย

English

ไทย

Appearance

LarbThe security-first coding agent

Open-source, model-agnostic, governed. A terminal-native autonomous engineer you can point at any model, extend with signed skills, and run without rate-limit cliffs or vendor lock-in — safe to open an untrusted repo by default.

Get started
Architecture
GitHub
Larb
🔒

Safe by default, powerful by consent

Trust-before-anything boot — nothing is read-as-code, executed, or sent over the network before you make a trust decision. Real container isolation; dangerous capabilities are opt-in and scoped.

🔌

No lock-in, ever

One config line points Larb at Anthropic, OpenAI, Gemini, DeepSeek, Groq, Mistral, xAI, OpenRouter, Together, Perplexity, or a local Ollama model. You own your keys, data, and routing.

✅

Verify, don't trust the model

A mandatory verification loop runs your lint / build / tests after edits and feeds failures back until the task verifiably passes — output is shippable, not merely plausible.

💸

Cost is a first-class signal

Live token and dollar accounting with hard limits that halt the agent before overspend — never a surprise bill. Cache-aware pricing and cheap-worker routing keep long runs inexpensive.

🧩

Extensible, but governed

Community skills with signing, a declared-permission manifest, and tiered trust — plus MCP servers plugged in as permission-gated tools. Install ≠ trust — unsigned skills run in the tightest sandbox.

🔎

Open and auditable end-to-end

The client is not a black box. An append-only audit log records every model call, tool call, and permission grant. Apache-2.0.

What is Larb? ​

Larb is an open-source, model-agnostic, security-first autonomous coding agent. It plans, edits, runs, tests, and self-corrects until a task verifiably passes — inside a capability sandbox, under a hard spend governor, with every action logged.

The name is also a Lao/Thai dish (ลาบ) — a nod to the project's roots. This site is bilingual: English · ไทย.

⚠️ Status: 0.1.0-alpha. Early, BYO-key developers only. APIs and config may change. See the roadmap and the README's Known limitations.

The wedge, in one sentence ​

Everyone else makes you choose between open, safe, unlocked, and cheap. Larb refuses that trade-off. See the comparison for how it stacks up against Claude Code, Codex CLI, Gemini CLI, DeepSeek, Aider, and OpenClaw.

Explore ​

  • Getting started — install, point Larb at a model, your first run, and the safety model to understand before you rely on it.
  • Architecture — the orchestrator loop, provider abstraction, sandbox, context engine, skills, and cross-cutting governors, with diagrams.
  • Comparison — the competitive landscape and where Larb wins.
  • Roadmap — what's shipped (phases 0–7) and the ambitious plan ahead.
  • Security model — the differentiator: the attack classes Larb designs out.