Letta Code
A stateful agent harness for long-lived agents that retain memory, identity, and learned context.
Per-dimension scores and reasoning
Evidence shows permission modes (standard/auto-approved/auto-denied) and user confirmation mechanisms (/connect for API keys), but implementation details are not provided. Data flow transparency: README mentions secrets obfuscation and memory repository sync, but does not detail data flow. Sensitive data handling: secrets feature exists, but encryption or storage methods are not specified. Dependency security: CI and dependency locking (bun.lockb not provided) exist, but no vulnerability scanning evidence. External effects: hooks and custom scripts exist, but permission controls are not detailed. Rollback: MemFS uses git tracking, but rollback mechanism is not explicit. Source attribution: license and copyright notices exist, but publisher is unverified.
Self-consistency: README and package.json descriptions align, CI configuration shows multiple tests, but test results are not provided. Dependency availability: dependency list is complete, but lockfile or mirrors are not provided. Failure messages: CI has error logging steps, but user-facing error handling documentation is not provided.
Audience and scenarios: README describes multiple usage scenarios (CLI, desktop, browser, messaging integrations), but detailed tutorials are not provided. Capability boundaries: permission modes and tool disabling options exist, but limitations are not explicit. Trigger precision: slash commands and crons exist, but precise trigger conditions are not provided. Environment fit: supports multiple platforms and remote environments, but detailed configuration guides are not provided.
Information architecture: README is well-structured with feature tables and documentation links. Install notes: npm installation and desktop app download are provided. Naming stability: package name and command are stable. Examples and FAQ: tutorial agent and documentation links exist, but FAQ is not provided. Known limitations: not explicitly listed. License: Apache-2.0 with brand asset exclusion. Versioning and changelog: version number exists but changelog is not provided. Maintenance responsibility: CI and release process exist, but maintainers are not explicit.
Output usability: supports multiple output formats (text/json/stream-json), but examples are not provided. Marginal value: provides unique features (memory, skills, subagents), but comparison with other tools is not provided. Cost-benefit: requires API keys and cloud services, but pricing information is not provided.
Claim traceability: README features have documentation links, but implementation code is not provided. Cross-source corroboration: CI tests and documentation exist, but independent verification is not provided. Fact-inference separation: README distinguishes feature descriptions and documentation links, but does not explicitly separate facts and inferences.
- Publisher identity is unverified; be cautious about supply chain risks.
- Dependency security lacks vulnerability scanning evidence; consider checking lockfile.
- Permission mode implementation is not evident in source; requires runtime verification.
- Data flow and sensitive data handling details are insufficient; further review needed.
What does this agent do, and when should you use it?
Letta Code is a stateful agent harness for building agents intended to evolve over time rather than operate as isolated chats. The same agent can be reached through a local CLI, desktop apps for macOS, Windows, and Linux, chat.letta.com, and Telegram, Slack, Discord, or custom channels. Its documented controls cover memory rewriting, skills, prompts, mods, subagents, hooks, schedules, and message search. With Letta Cloud, agent state and conversations are stored in the cloud while the harness can run on connected machines or cloud environments. This makes it most relevant when continuity across sessions, interfaces, or execution hosts matters.
After you run letta in a project directory, /connect configures your OpenAI, ChatGPT, Anthropic, or Z.ai coding plan API keys, and /model switches models. MemFS tracks context, including memory blocks, through Git; /memory-repository set [email protected]:... syncs that context to a custom GitHub repository. The harness exposes /search for messages and agents, /skills for skills, /sleeptime for periodic dreaming, /doctor for memory-quality audits, and /palace for viewing memory. A machine can be registered with letta server --env-name "...", discovered with letta environments list --online-only, and targeted with letta -p --agent <agent-id> --environment "..." "...".
- A developer building a project assistant that must retain context and revise its memory or skills over many sessions.
- A user who wants to continue working with one agent while moving between a terminal, desktop app, browser, and Slack.
- An automation owner who needs an agent to operate over time through heartbeats, crons, or periodic dreaming.
- An engineering team routing the same agent between a laptop, cloud VM, Mac Mini, or managed sandbox.
- A developer who wants Git-tracked agent context and the option to sync it to a custom GitHub repository.
What are this agent's strengths and limitations?
- It treats memory blocks, skills, prompts, and mods as long-lived state that agents can rewrite, rather than limiting adaptation to a single conversation.
- One agent is documented as reachable through CLI, desktop, browser, mobile browser, and multiple messaging channels.
- MemFS tracks all context through Git and can sync it to a custom GitHub repository.
- Built-in subagents, hooks, search, and self-managed schedules provide concrete building blocks for long-running workflows.
- You must configure your own LLM API keys for the named providers; provider pricing and quota behavior are not documented here.
- Letta Cloud, remote multi-environment operation, and secrets require a Letta sign-in, so those workflows are not purely local.
- Permissions, hooks, skills, and autonomous schedules require operational configuration and governance by the adopter.
- The supplied documentation does not specify CLI runtime prerequisites by operating system, resource requirements, or production failure-recovery behavior.
How do you install or deploy this agent?
Install with npm install -g @letta-ai/letta-code. In a project directory, run letta; for the tutorial agent, run letta --new-agent --personality tutorial. Run /connect to configure your LLM API keys, then choose a model with /model. For Letta Cloud, remote environments, or secrets, run /login in the CLI or sign in through the desktop app.
How do you use this agent?
Start letta in the relevant project directory and interact with the agent. Use /skills to inspect loaded global, project-scoped, or agent-scoped skills; install an external skill with letta skills install <skill> and list a specific agent's skills with letta skills list --agent <agent-id>. To use a remote host, run letta server --env-name "work-laptop" there, list available hosts with letta environments list --online-only, then send work using --environment "work-laptop".
FAQ
Is Letta Cloud required?
Which model providers are explicitly supported?
/connect explicitly lists OpenAI, ChatGPT, Anthropic, and Z.ai coding plan, and /model is used to swap models.How is agent memory stored or moved?
/memory-repository set [email protected]:....Can an agent work on a schedule?
/sleeptime configures periodic dreaming.