Skip to content

armgit / letta

Public Imported from GitHub
0

Platform for stateful agents: AI with advanced memory that can learn and self-improve over time.

fix: patch minimax (#9099)

Sarah Woodersfb69a96cd668
diff --git a/letta/llm_api/anthropic_client.py b/letta/llm_api/anthropic_client.py
index 8acd65dd9..7f133afb1 100644
--- a/letta/llm_api/anthropic_client.py
+++ b/letta/llm_api/anthropic_client.py
@@ -356,9 +356,17 @@ class AnthropicClient(LLMClientBase):
) -> Union[anthropic.AsyncAnthropic, anthropic.Anthropic]:
api_key, _, _ = self.get_byok_overrides(llm_config)
+ # For MiniMax provider, use minimax_api_key from settings
+ if not api_key and llm_config.provider_name == "minimax":
+ api_key = model_settings.minimax_api_key
+
# For claude-pro-max provider, use OAuth Bearer token instead of api_key
is_oauth_provider = llm_config.provider_name == "claude-pro-max"
+ # Only use custom base_url for MiniMax (Anthropic-compatible API)
+ # The Anthropic SDK adds /v1/messages internally, so we only override for non-Anthropic providers
+ base_url = llm_config.model_endpoint if llm_config.provider_name == "minimax" else None
+
if async_client:
if api_key:
if is_oauth_provider:
@@ -370,8 +378,8 @@ class AnthropicClient(LLMClientBase):
"anthropic-beta": "oauth-2025-04-20",
},
)
- return anthropic.AsyncAnthropic(api_key=api_key, max_retries=model_settings.anthropic_max_retries)
- return anthropic.AsyncAnthropic(max_retries=model_settings.anthropic_max_retries)
+ return anthropic.AsyncAnthropic(api_key=api_key, base_url=base_url, max_retries=model_settings.anthropic_max_retries)
+ return anthropic.AsyncAnthropic(base_url=base_url, max_retries=model_settings.anthropic_max_retries)
if api_key:
if is_oauth_provider:
@@ -383,8 +391,8 @@ class AnthropicClient(LLMClientBase):
"anthropic-beta": "oauth-2025-04-20",
},
)
- return anthropic.Anthropic(api_key=api_key, max_retries=model_settings.anthropic_max_retries)
- return anthropic.Anthropic(max_retries=model_settings.anthropic_max_retries)
+ return anthropic.Anthropic(api_key=api_key, base_url=base_url, max_retries=model_settings.anthropic_max_retries)
+ return anthropic.Anthropic(base_url=base_url, max_retries=model_settings.anthropic_max_retries)
@trace_method
async def _get_anthropic_client_async(
@@ -392,9 +400,17 @@ class AnthropicClient(LLMClientBase):
) -> Union[anthropic.AsyncAnthropic, anthropic.Anthropic]:
api_key, _, _ = await self.get_byok_overrides_async(llm_config)
+ # For MiniMax provider, use minimax_api_key from settings
+ if not api_key and llm_config.provider_name == "minimax":
+ api_key = model_settings.minimax_api_key
+
# For claude-pro-max provider, use OAuth Bearer token instead of api_key
is_oauth_provider = llm_config.provider_name == "claude-pro-max"
+ # Only use custom base_url for MiniMax (Anthropic-compatible API)
+ # The Anthropic SDK adds /v1/messages internally, so we only override for non-Anthropic providers
+ base_url = llm_config.model_endpoint if llm_config.provider_name == "minimax" else None
+
if async_client:
if api_key:
if is_oauth_provider:
@@ -406,8 +422,8 @@ class AnthropicClient(LLMClientBase):
"anthropic-beta": "oauth-2025-04-20",
},
)
- return anthropic.AsyncAnthropic(api_key=api_key, max_retries=model_settings.anthropic_max_retries)
- return anthropic.AsyncAnthropic(max_retries=model_settings.anthropic_max_retries)
+ return anthropic.AsyncAnthropic(api_key=api_key, base_url=base_url, max_retries=model_settings.anthropic_max_retries)
+ return anthropic.AsyncAnthropic(base_url=base_url, max_retries=model_settings.anthropic_max_retries)
if api_key:
if is_oauth_provider:
@@ -419,8 +435,8 @@ class AnthropicClient(LLMClientBase):
"anthropic-beta": "oauth-2025-04-20",
},
)
- return anthropic.Anthropic(api_key=api_key, max_retries=model_settings.anthropic_max_retries)
- return anthropic.Anthropic(max_retries=model_settings.anthropic_max_retries)
+ return anthropic.Anthropic(api_key=api_key, base_url=base_url, max_retries=model_settings.anthropic_max_retries)
+ return anthropic.Anthropic(base_url=base_url, max_retries=model_settings.anthropic_max_retries)
@trace_method
def build_request_data(
diff --git a/letta/schemas/providers/base.py b/letta/schemas/providers/base.py
index 6b5b722df..f24794d6b 100644
--- a/letta/schemas/providers/base.py
+++ b/letta/schemas/providers/base.py
@@ -192,6 +192,7 @@ class Provider(ProviderBase):
GroqProvider,
LettaProvider,
LMStudioOpenAIProvider,
+ MiniMaxProvider,
MistralProvider,
OllamaProvider,
OpenAIProvider,
@@ -244,6 +245,8 @@ class Provider(ProviderBase):
return LMStudioOpenAIProvider(**self.model_dump(exclude_none=True))
case ProviderType.bedrock:
return BedrockProvider(**self.model_dump(exclude_none=True))
+ case ProviderType.minimax:
+ return MiniMaxProvider(**self.model_dump(exclude_none=True))
case _:
raise ValueError(f"Unknown provider type: {self.provider_type}")
diff --git a/letta/schemas/providers/minimax.py b/letta/schemas/providers/minimax.py
index 616cf11d5..707771133 100644
--- a/letta/schemas/providers/minimax.py
+++ b/letta/schemas/providers/minimax.py
@@ -73,7 +73,7 @@ class MiniMaxProvider(Provider):
configs.append(
LLMConfig(
model=model["name"],
- model_endpoint_type="minimax",
+ model_endpoint_type="anthropic",
model_endpoint=self.base_url,
context_window=model["context_window"],
handle=self.get_handle(model["name"]),
diff --git a/letta/server/server.py b/letta/server/server.py
index bfdb21a85..7cb815a80 100644
--- a/letta/server/server.py
+++ b/letta/server/server.py
@@ -67,6 +67,7 @@ from letta.schemas.providers import (
GroqProvider,
LettaProvider,
LMStudioOpenAIProvider,
+ MiniMaxProvider,
OllamaProvider,
OpenAIProvider,
OpenRouterProvider,
@@ -342,6 +343,13 @@ class SyncServer(object):
api_key_enc=Secret.from_plaintext(model_settings.xai_api_key),
)
)
+ if model_settings.minimax_api_key:
+ self._enabled_providers.append(
+ MiniMaxProvider(
+ name="minimax",
+ api_key_enc=Secret.from_plaintext(model_settings.minimax_api_key),
+ )
+ )
if model_settings.zai_api_key:
self._enabled_providers.append(
ZAIProvider(
diff --git a/tests/test_providers.py b/tests/test_providers.py
index 216006823..3b013a357 100644
--- a/tests/test_providers.py
+++ b/tests/test_providers.py
@@ -154,8 +154,8 @@ async def test_minimax():
assert model.context_window == 200000
# All MiniMax models have 128K max output
assert model.max_tokens == 128000
- # All use minimax endpoint type
- assert model.model_endpoint_type == "minimax"
+ # MiniMax uses Anthropic-compatible API endpoint
+ assert model.model_endpoint_type == "anthropic"
@pytest.mark.skipif(model_settings.azure_api_key is None, reason="Only run if AZURE_API_KEY is set.")