Letta Code

Letta Code is a stateful agent harness. Its agents hold memory, an identity, and a record of past experience, and they change over long horizons by rewriting their own memory, skills, prompts, and — through mods — the harness itself. It runs interactively or behind always-on agents that work proactively.

You reach an agent from a local CLI, a desktop app for macOS, Windows, and Linux, a browser including mobile, or messaging integrations for Telegram, Slack, Discord, and custom channels. Agents created in the CLI appear in the desktop app and the other way round, and the same agent is reachable from every channel.

Context is versioned in git through MemFS: memory blocks and other context are tracked as files and can be synced to a GitHub repository you nominate. Skills load from three scopes — global under ~/.letta, project-scoped under .agents/skills, and agent-scoped inside MemFS — and are installed from a GitHub URL, ClawHub, or the Hermes Skills Hub. Built-in subagents (general-purpose, forked, recall, history-analyzer) run in the background, and an agent can call any other agent, including itself, as a subagent.

Agents stored in Letta Cloud keep their memory, identity, and conversations there while the harness runs on any connected computer: a laptop, GitHub Actions, a managed cloud sandbox, a remote VM, or a Mac Mini. letta server registers a machine under a name, letta computers list shows what is online, and a headless message can be routed to a named computer or to a cloud sandbox. Model keys and coding-plan subscriptions are configured with /connect, and /model swaps between them.

Features

• Self-editing memory: agents rewrite their own context, including system-prompt learning through memory blocks and skill learning from experience
• MemFS: all context is tracked in git and can be synced to a custom GitHub repository
• Memory tooling: /sleeptime for periodic dreaming, /doctor to audit memory quality, and /palace to view memory
• Message search: /search spans every message and agent, and agents can search their own or other agents' conversations
• Subagents: built-in general-purpose, forked, recall, and history-analyzer subagents, plus agent-to-agent calls
• Channels: CLI, desktop app, browser and mobile, Slack, Telegram, Discord, and custom channels
• Skills: global, project, and agent-scoped skills, installable from GitHub, ClawHub, or the Hermes Skills Hub
• Hooks: custom scripts run at set points in agent execution
• Permissions: permission modes with per-action auto-approve and auto-deny
• Scheduling: heartbeats, crons, and schedules the agent manages itself, with reviewable past runs
• Remote computers: any machine joins with letta server, and messages route to it by name
• Secrets: values exposed as environment variables across machines while kept out of the agent's context