Bring your own model, run agents locally or on any provider

MyAgens runs on Claude by default, but you are not locked in. This guide covers running agents against a model on your own hardware. If you are choosing between the hosted backends instead, or setting up failover for usage limits, see Choose an AI backend.

Local and self-hosted models

Point MyAgens at a local model through LM Studio or Ollama, or at any OpenAI-compatible or Anthropic-compatible endpoint. That keeps everything on your hardware with no cloud key, which is ideal for private or offline work, as long as you run a capable enough model locally.

Per-agent routing

Each agent can use its own model, so you can spend where it counts: a routine Lead on a cheap or local model, Atlas on a frontier one. Switch a model with /model.

How to set it up

1. Start your local server: in LM Studio load a model and start its server, or run ollama serve with a model pulled.
2. In MyAgens, open the panel's model settings, or run /model in Telegram, and point an agent at the endpoint, for example http://127.0.0.1:1234/v1 for LM Studio or http://127.0.0.1:11434 for Ollama.
3. Choose which agent uses it. Set it globally, or per Lead so one specialist runs local while Atlas stays on Claude.

Give the local model a real task before you rely on it. Small models handle chat fine but struggle with the multi-step tool use agents lean on, so confirm it holds up on your actual work. The Ollama backend is plain chat rather than an agentic CLI, which is exactly why it stays fast and fully independent of any cloud provider, and also why it is the wrong choice for long tool-driven jobs.

A local model as a fallback

A local model earns its keep as a fallback target even if it is not your primary. Give an agent a frontier primary and an Ollama fallback, and a usage limit on the hosted side means a slower answer rather than no answer. Choose an AI backend covers how failover is triggered and what it carries over.

Offline extras

Semantic memory and voice can run offline too, with local embeddings and local speech models, so a fully self-hosted setup stays fully private.

For picking a capable local model and setting honest expectations about what one can do, there is a deep dive at Running Claude Code on local models.