Get started

Letta Code

Explore Letta Code, a memory-first, model-agnostic agent harness

Letta Code allows you to create self-improving, stateful coding agents that can learn from experience and improve with use.

Unlike Claude Code, Letta Code is open source, model agnostic (you can use it with any model!), and most importantly, is stateful, meaning that you can use the same agent across many coding sessions, and have it learn and improve over time.

You can also use Letta Code as a CLI harness to connect any Letta agent (even non-coding agents) to your terminal.

Installation

Requirements:

Node.js (version 18+)
A Letta Developer Platform account

npm install -g @letta-ai/letta-code

Then navigate to your project and run:

cd your-project
letta

On first use, you’ll be prompted to log in via OAuth. Alternatively, set a valid LETTA_API_KEY in your environment.

Key features

Stateful memory — Your agent persists across sessions. It remembers your codebase, preferences, and past interactions. Run /init to bootstrap project knowledge, and /remember to save important context.

Model agnostic — Use Claude, GPT, Gemini, or any supported model. Switch models mid-conversation with /model. Your agent’s memory stays intact.

Skills — Agents can learn reusable skills from experience. Skills transfer across sessions and projects, letting your agent get better over time.

Subagents — Delegate complex tasks to specialized subagents that run autonomously. Explore codebases, plan implementations, or parallelize work.

Headless mode — Run Letta Code non-interactively for scripting and CI. Compose agents as UNIX programs.

Learn more

Quickstart — Set up your agent and learn the basics
Memory — Understand the hierarchical memory system
Skills — Create reusable skills your agent can learn
Commands — Full CLI and interactive command reference
Models — Switch models and configure toolsets
Headless mode — Run non-interactively in scripts and CI

Additional resources

Self-hosting — Connect to a self-hosted Letta server
Permissions — Configure tool approval settings
Subagents — Delegate tasks to specialized agents
GitHub — Source code and issue tracker