LLM Providers¶

RLM Code connects to 20+ model providers through a unified LLMConnector backed by a ProviderRegistry. Providers are organized into two connection modes: Local (no API key required) and BYOK (Bring Your Own Key).

Module¶

rlm_code.models.llm_connector      -- Connection and generation
rlm_code.models.providers.registry  -- Provider metadata and resolution
rlm_code.models.providers.model_catalog  -- Curated model lists

Architecture¶

graph TD
    User["User: /connect provider model"] --> LLM[LLMConnector]
    LLM --> Registry[ProviderRegistry]
    Registry --> Resolve["Resolve provider + model"]
    Resolve --> Adapter{"Adapter Type"}
    Adapter -->|ollama| Ollama[Ollama API]
    Adapter -->|openai| OpenAI[OpenAI SDK]
    Adapter -->|anthropic| Anthropic[Anthropic SDK]
    Adapter -->|gemini| Gemini[Google GenAI SDK]
    Adapter -->|openai_compatible| Compatible["OpenAI-Compatible HTTP"]

Provider Registry¶

The ProviderRegistry holds a ProviderSpec for each supported provider and resolves provider names, aliases, and model identifiers.

ProviderSpec Dataclass¶

from rlm_code.models.providers.registry import ProviderSpec

spec = ProviderSpec(
    provider_id="openai",
    display_name="OpenAI",
    adapter_type="openai",
    aliases=["gpt"],
    api_key_env_vars=["OPENAI_API_KEY"],
    base_url_env_var="OPENAI_API_BASE",
    docs_url="https://platform.openai.com/",
    requires_api_key=True,
    connection_type="byok",
    category="US Labs",
    example_models=["gpt-5", "gpt-4o", "gpt-4o-mini"],
)

Registry Methods¶

from rlm_code.models.providers.registry import ProviderRegistry

registry = ProviderRegistry.default()

Local Providers¶

Local providers run models on your own hardware. No API key is required.

Ollama¶

# Connect
/connect ollama qwen2.5-coder:7b
/connect ollama llama3.2:3b
/connect ollama gpt-oss:20b

LM Studio¶

vLLM¶

SGLang¶

Hugging Face TGI¶

MLX¶

llama.cpp Server¶

OpenAI-Compatible (Generic)¶

Use this for any local server that exposes an OpenAI-compatible /v1/chat/completions endpoint.

BYOK Providers (Cloud)¶

Bring Your Own Key providers require an API key, typically set via environment variable.

US Labs¶

OpenAI¶

Curated models:

gpt-5.3-codex, gpt-5.2, gpt-5.2-pro, gpt-5.2-codex,
gpt-5.1, gpt-5.1-codex, gpt-5.1-codex-mini,
gpt-4o-2024-11-20, gpt-4o, gpt-4o-mini, o1, o1-mini

Anthropic¶

Curated models:

claude-opus-4-6, claude-opus-4-5-20251101,
claude-sonnet-4-5-20250929, claude-haiku-4-5-20251001,
claude-sonnet-4-20250514, claude-opus-4-20250514, claude-haiku-4-20250514

Google Gemini¶

Curated models:

gemini-3-pro-preview, gemini-3-flash-preview,
gemini-2.5-pro, gemini-2.5-flash, gemini-2.0-flash, gemini-flash-latest

xAI¶

Curated models: grok-3, grok-3-mini, grok-2, grok-beta

Perplexity¶

Curated models: sonar-pro, sonar, sonar-reasoning

China Labs¶

DeepSeek¶

Curated models:

deepseek-ai/DeepSeek-V3.2, deepseek-ai/DeepSeek-R1,
deepseek-chat, deepseek-coder, deepseek-reasoner

Moonshot (Kimi)¶

Curated models: moonshot-v1-128k, moonshot-v1-32k, moonshot-v1-8k, kimi-k2

Alibaba (DashScope / Qwen)¶

Curated models: qwen-max, qwen-plus, qwen-turbo, qwen2.5-coder-32b-instruct, qwen2.5-72b-instruct

Model Hosts¶

OpenRouter¶

Curated models:

anthropic/claude-sonnet-4, openai/gpt-4o,
google/gemini-2.0-flash, meta-llama/llama-3.3-70b-instruct

OpenCode Zen¶

Curated models: glm-4.7-free, grok-code, kimi-k2.5-free, gpt-5-nano, minimax-m2.1-free, big-pickle

Groq¶

Curated models: llama-3.3-70b-versatile, llama-3.1-8b-instant, mixtral-8x7b-32768, gemma2-9b-it

Together AI¶

Curated models:

meta-llama/Llama-3.3-70B-Instruct-Turbo,
Qwen/Qwen2.5-Coder-32B-Instruct,
deepseek-ai/DeepSeek-R1,
mistralai/Mixtral-8x22B-Instruct-v0.1

SiliconFlow¶

Fireworks AI¶

Cerebras¶

Curated models: llama-3.3-70b, qwen-3-coder-480b

Other Labs¶

Mistral AI¶

Curated models: mistral-large-2411, mistral-medium-2505, mistral-nemo, codestral-latest, mistral-small-latest

LLM Connector¶

The LLMConnector class manages connections and response generation.

Constructor¶

from rlm_code.models.llm_connector import LLMConnector
from rlm_code.core.config import ConfigManager

connector = LLMConnector(config_manager=ConfigManager())

Key Methods¶

Usage Tracking¶

The connector automatically tracks token usage across all provider calls:

# Before a task
before = connector.usage_snapshot()

# Run task...
response = connector.generate_response("Explain this code")

# After
after = connector.usage_snapshot()
delta = LLMConnector.usage_delta(before, after)
print(delta)
# {"total_calls": 1, "prompt_tokens": 150, "completion_tokens": 300}

The full usage summary includes per-model breakdowns:

summary = connector.get_usage_summary()
# {
#   "totals": {"total_calls": 5, "prompt_tokens": 1200, "completion_tokens": 2400},
#   "by_model": {
#     "anthropic/claude-sonnet-4-5-20250929": {"calls": 3, ...},
#     "ollama/qwen2.5-coder:7b": {"calls": 2, ...},
#   },
#   "last_call": {"timestamp": "...", "model": "...", ...},
# }

Adapter Types¶

Each provider uses one of five adapter backends:

pip install "openai>=2.8.1"      # For OpenAI and compatible providers
pip install anthropic             # For Anthropic Claude
pip install "google-genai>=1.52.0"  # For Google Gemini

Model Catalog¶

The SUPERQODE_MODEL_CATALOG provides curated model lists for each provider, used by the TUI's interactive model picker:

from rlm_code.models.providers.model_catalog import get_superqode_models

models = get_superqode_models("anthropic")
# ["claude-opus-4-6", "claude-opus-4-5-20251101", ...]

The catalog uses curated provider/model lists and takes priority over the example_models field in ProviderSpec.

Provider Resolution¶

The connector resolves providers through a multi-step process:

1. Explicit provider -- If model_type is given, resolve via ProviderRegistry.get().
2. Model prefix inference -- If the model name contains / (e.g., anthropic/claude-sonnet-4), extract the provider prefix.
3. Alias matching -- All aliases are checked case-insensitively (e.g., claude resolves to anthropic).

# All of these resolve to the Anthropic provider:
connector.connect_to_model("claude-sonnet-4-5-20250929", model_type="anthropic")
connector.connect_to_model("anthropic/claude-sonnet-4-5-20250929")
connector.connect_to_model("claude-sonnet-4-5-20250929", model_type="claude")

Environment Variables Reference¶