AI harnesses are the software layer that turns a language model into a working agent. Letta Code is Letta’s memory-first coding agent harness: instead of starting fresh every session, Letta Code agents carry persistent memory, learn skills from experience, and rewrite their own system prompt. As of 20 September 2026.
What is Letta Code?
Letta Code is developed by Letta, founded by the authors of MemGPT and the sleep-time compute approach (now called “dreaming”). The project is open source on GitHub at letta-ai/letta-code under the Apache-2.0 licence. The harness is installed via npm install -g @letta-ai/letta-code; afterwards, letta starts a session in the project directory, and letta --new-agent --personality tutorial starts a guided introduction. Besides the CLI, a desktop app for macOS, Windows and Linux and a web client at chat.letta.com are available, including on mobile.
Letta Code as an AI harness
The key difference from classic coding harnesses: agents are, per the documentation, long-lived and “more like people than tools.” They learn through memory blocks - building blocks of the system prompt that the agent itself reads and rewrites - and through learned skills from global (~/.letta), project-scoped (.agents/skills) and agent-owned sources. MemFS versions this entire context via git and can be synced to your own GitHub repository with /memory-repository set. For complex tasks, seven built-in subagents are available (general-purpose, fork, recall, reflection, history-analyzer, memory, init), which run as their own subprocess with their own context; any other existing Letta agent can also be deployed as a subagent. Four permission modes - strict, standard, acceptEdits, unrestricted - control which tool calls are auto-approved.
Models and model access
Letta Code is explicitly model-agnostic: /connect sets API keys for Anthropic, OpenAI, Google Gemini, Mistral, DeepSeek, Amazon Bedrock, Azure OpenAI and other providers, and /model switches the active model. For a custom endpoint, Letta Code documents support for OpenAI-compatible providers with a freely chosen base URL and API key, as long as that endpoint supports chat completions and tool calling. Model access can therefore generally be routed through a central proxy such as our AI Gateway by entering its OpenAI-compatible endpoint as the base URL.
Enterprise use: data protection and governance
Agents and their memory can either run in Letta Cloud or be fully self-hosted: letta server starts the Apache-2.0-licensed app server on any machine, from a laptop through a GitHub Action to your own VM. Permission modes and project-wide allow/deny rules in .letta/settings.json govern automated tool execution, and a cross-agent memory guard protects other agents’ memory stores. The public documentation makes no statement on telemetry or central usage-data collection at this point. Regardless of the deployment model: source code and conversation content go to whichever model provider is chosen - this belongs in your own records of processing activities in advance, particularly when using Letta Cloud rather than a self-hosted server.
Orchestrating Letta Code with herdr
herdr is a terminal runtime for several coding agents running at once. Per herdr’s documentation, it detects Letta Code through the terminal screen manifest and provides session integration through it, making Letta Code sessions visible alongside other supported harnesses in the same multi-agent overview.
Where it fits: developer harness and enterprise harness
Letta Code targets developers who want long-lived, learning agents in a terminal, an IDE-adjacent desktop app, or a browser. For business users across an entire company, CompanyGPT is the fitting harness: tools via MCP servers, context via companyRAG, roles from Entra ID or Keycloak, and audit logs. Our AI Gateway provides the shared cost and governance layer across different harnesses and model providers. You’ll find further agent harnesses in our harness overview. If you’d like to check how a memory-first agent like Letta Code fits into your development environment and data protection concept, we’re glad to help.