Supported runtimes and backend management

Loomcast provides an abstraction layer that lets you drive several local LLM runtime servers through a single, unified interface.

---

1. Runtime comparison

Capability	Ollama	LM Studio	llama.cpp (llama-server)
Backend enum	ollama	lmstudio	llamacpp
Default port	11434	1234	8080
API protocol	Native REST API (/api/chat, /api/pull, /api/tags)	OpenAI-compatible API (/v1/chat/completions, /v1/models)	OpenAI-compatible API (/v1/chat/completions, /v1/models, /health)
Model installation	Automatic through the API (POST /api/pull)	Manual, inside the LM Studio application	Manual, at server startup (llama-server -m)
Integrity verification	Available (SHA-256 digest returned by the Ollama API)	Unavailable (/v1/models exposes no digests)	Unavailable (the REST API provides no digests)
Concurrent models	Several models managed per instance	Depends on what the application is configured for	One model loaded at a time per llama-server instance

---

2. Known limitations per runtime

Ollama (ollama)

• Strengths: native support for model downloads with real-time progress reporting, and full integrity verification through the SHA-256 digest (loomcast verify).

LM Studio (lmstudio)

• Installation: LM Studio exposes no public HTTP endpoint for downloading models remotely through an API. Models must be searched for and loaded beforehand in the LM Studio GUI.
• Verification: loomcast verify reports an unverified status, because the LM Studio OpenAI-compatible API includes no cryptographic signatures.

llama.cpp (llamacpp)

• Single model: a llama-server instance serves exactly one model, chosen when the executable starts (-m /path/to/model.gguf).
• API installation: there is no hot download through the REST API. Trying to install a model that is not loaded returns an error message telling you to restart the server with the model you want.

---

3. Switching between runtimes

Loomcast lets you switch backends through three distinct mechanisms.

A. With the CLI flag (--backend / -b)

To override the runtime for a single CLI command:

# Check which models are loaded in LM Studio
loomcast list --backend lmstudio

# Download a model into Ollama
loomcast install qwen2.5:0.5b -b ollama

# Chat with a model served by llama-server
loomcast chat gemma-2-2b "Explain the principles of Clean Architecture." --backend llamacpp

B. From the Python SDK (runtime_type)

In your Python scripts, pass the runtime_type parameter:

import loomcast

# Request against LM Studio
response_lms = loomcast.chat(
    model="qwen2.5-coder-7b",
    messages=[{"role": "user", "content": "Hello!"}],
    runtime_type="lmstudio",
)

# Get the runtime adapter directly
manager = loomcast.get_manager()
llamacpp_adapter = manager.get_runtime("llamacpp")

if llamacpp_adapter.is_available():
    print("llama-server is running on port 8080!")

C. Through global configuration (default_backend)

To change the default backend permanently:

From the CLI

loomcast config --set-default lmstudio

In config.toml

default_backend = "lmstudio"

Through an environment variable (temporary, or for the shell session)

export LOOMCAST_DEFAULT_BACKEND="llamacpp"