Letta Code

// Overview

Letta Code is a stateful agent harness from the team behind Letta (formerly MemGPT). Where most coding agents start every session from an empty context, a Letta Code agent carries memory, identity and a sense of experience over time. It learns over long horizons by rewriting its own memory blocks, skills, prompts and even the harness itself through mods, and the agents are designed to be self-configuring — the documentation's own advice is to ask the agent to change its skills, hooks or permissions rather than editing config by hand.

It can be driven interactively or left running as an always-on agent that works proactively. The same agent is reachable from a local CLI, a desktop app for macOS, Windows and Linux, a browser (including mobile) at chat.letta.com, and messaging integrations for Telegram, Slack and Discord, with a documented path for custom channels. Context — memory blocks included — is tracked in git through MemFS, so it can be synced to a GitHub repository you own.

Around that sit the features you would expect from a modern harness: skills loaded from global, project and agent scopes; built-in subagents (general-purpose, forked, recall, history-analyzer) that agents can call in the background, including calling other agents or themselves; hooks that run scripts at points in the agent loop; permission modes for auto-approval and auto-denial; and heartbeats and crons so agents can schedule their own work. Signing in to Letta Cloud adds remote computers — the harness runs on any connected machine while agent state stays in the cloud — and obfuscated secrets. The repository is Apache-2.0.

// What it does

• Persistent memory blocks and skill learning, with /sleeptime for periodic consolidation, /doctor to audit memory quality and /palace to inspect it
• MemFS — all context tracked in git and syncable to your own GitHub repository
• Skills at three scopes (global ~/.letta, project .agents/skills, agent-scoped in MemFS) plus /skill-creator
• Built-in subagents (general-purpose, forked, recall, history-analyzer) and agent-to-agent calls
• One agent, many front ends: CLI, desktop app, browser and mobile, Slack, Telegram and Discord
• Hooks, permission modes, and self-managed crons and heartbeats
• Message search across all agents and conversations with /search
• Remote computers via Letta Cloud — run the harness on a laptop, a VM, GitHub Actions or a Mac Mini against the same agent state

// Getting started

Letta Code installs from npm as a global CLI. A desktop app is available too, and agents created in one show up in the other.

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

cd my-project
letta

# or start with the guided agent
letta --new-agent --personality tutorial

/connect
/model

letta skills install https://github.com/owner/repo

letta server --computer-name "work-laptop"
letta computers list --online-only
letta -p --agent <agent-id> --computer "work-laptop" "hello from that machine"

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Keep a coding agent's context, conventions and hard-won fixes across sessions instead of re-explaining them each morning
• Run an always-on agent that picks up work on a schedule and reports back over Slack or Telegram
• Move the same agent between a laptop, a VM and CI while its memory stays in one place
• Version an agent's context in git so memory changes are reviewable like code

// How Letta Code compares

Letta Code alongside other open-source autonomous coding agents tools AI/TLDR tracks, ranked by GitHub stars.