Letta Code adds embedded local server with Ollama and LM Studio support
Letta Code can now run fully locally with an embedded server, removing the login and Docker requirement while keeping memory sync via `/memory-repository`. That gives developers a local-first agent harness with optional Ollama and LM Studio support instead of forcing everything through Letta’s hosted API.
TL;DR
- Letta_AI's announcement says Letta Code now runs fully locally with an embedded server, which removes both the login and Docker requirement.
- Local mode still keeps an optional sync path, because Letta_AI's announcement says memory stays on disk by default but can be pushed to GitHub with
/memory-repository. - Local model support is part of the same update, with Letta_AI's announcement naming Ollama and LM Studio via pi-ai from badlogicgames' reply context.
- sarahwooders' explanation says Letta built its hosted agents service first and added the local harness later, after users asked for Letta Code without the Letta API.
Letta_AI's launch post is the core change, sarahwooders' follow-up explains the architecture tradeoff, and Letta_AI's Memory view demo shows that the local harness work is landing alongside more opinionated memory tooling. There is also a teaser for Letta_AI's livestream post covering local mode, Letta Code improvements, and a still-murky /experiments area.
Embedded local server
The headline change is simple: Letta Code now bundles an embedded server, so the product no longer assumes a hosted control plane. That turns setup from a service-plus-container workflow into a local app workflow.
Sarah Wooders explains that Letta built "somewhat backwards," starting with the agents API service and only later building the bridge back to local execution. Her thread says the shared service made agents decoupled from any one environment, but user demand for the same harness without the Letta API pushed the team to ship local mode.
Local memory and model backends
The storage model is local-first. Memory lives on the machine, while /memory-repository adds an optional GitHub sync path instead of making cloud sync mandatory.
The model backend story also changed in the same release. Letta_AI's announcement says local LLM support is now built in for Ollama and LM Studio through pi-ai from badlogicgames' reply context.
Memory structure
Letta is exposing more of the agent's internal context layout in the UI. The new Memory view surfaces three pieces directly:
- context type: core, external, skills
- context links
- version history
sarahwooders' context note adds the more interesting claim: agents own their own context, including system-prompt-level memory blocks, skills, and external files, and different agents diverge in how they organize that context over time.
Multi-agent UI and /experiments
The launch was paired with a livestream agenda that named four things: a new Letta app version, Letta Code improvements, local mode, and a "mysterious" /experiments area. The post does not explain what /experiments contains, but it does mark it as part of the current product surface.
A separate clip shows a "Toggle between your AI employees" interface, which suggests the local harness is shipping alongside a more explicit multi-agent workspace metaphor, not just a single-agent terminal.