torbjorn/my-pal-mcp-server
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

refactor: moved temperature method from base provider to model capabilities

Browse Source 
refactor: model listing cleanup, moved logic to model_capabilities.py
docs: added AGENTS.md for onboarding Codex
This commit is contained in:
Fahad
2025-10-02 10:25:41 +04:00
parent f461cb4519
commit 6d237d0970
14 changed files with 460 additions and 512 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
providers/custom.py
View File

@@ -32,11 +32,20 @@ _TEMP_UNSUPPORTED_KEYWORDS = [
class CustomProvider(OpenAICompatibleProvider):
"""Adapter for self-hosted or local OpenAI-compatible endpoints.
The provider reuses the :mod:`providers.shared` registry to surface
user-defined aliases and capability metadata. It also normalises
Ollama-style version tags (``model:latest``) and enforces the same
restriction policies used by cloud providers, ensuring consistent
behaviour regardless of where the model is hosted.
Role
Provide a uniform bridge between the MCP server and user-managed
OpenAI-compatible services (Ollama, vLLM, LM Studio, bespoke gateways).
By subclassing :class:`OpenAICompatibleProvider` it inherits request and
token handling, while the custom registry exposes locally defined model
metadata.
Notable behaviour
* Uses :class:`OpenRouterModelRegistry` to load model definitions and
aliases so custom deployments share the same metadata pipeline as
OpenRouter itself.
* Normalises version-tagged model names (``model:latest``) and applies
restriction policies just like cloud providers, ensuring consistent
behaviour across environments.
"""
FRIENDLY_NAME = "Custom API"