Skip to main content
Glama
JacobRyu

Custom MCP Server

by JacobRyu
OverviewSchemaSourceRelated ServersDiscussionsAdmin
Python
Remote

Custom MCP Server

A custom MCP (Model Control Proxy) server that sits between clients and an OpenAI-compatible chat-completion backend. It provides:

  • Input validation, normalization, and chunking

  • Session-based conversational memory (in-memory store, swappable)

  • Sliding-window context management with rough token capping

  • Prompt construction with injectable system instructions

  • Resilient async agent calls (timeout + single retry on 429/5xx)

This is the MVP scope (milestones M1 + M2). Streaming, real summarization, persistent stores, and auth are deferred.

Requirements

  • Python 3.11+

  • uv (package & environment manager)

Related MCP server: MCP Server with LLM Integration

Install

uv sync --all-extras
# activate the environment
source .venv/bin/activate

Configure

cp .env.example .env
# edit .env and set AGENT_API_KEY, AGENT_API_BASE_URL, AGENT_MODEL

Run

uv run uvicorn app.main:app --reload --port 8000

Endpoints

POST /agent/input

curl -s http://localhost:8000/agent/input \
  -H 'Content-Type: application/json' \
  -d '{"user_input": "Hello!"}'

GET /session/{id}

curl -s http://localhost:8000/session/<session_id>

DELETE /session/{id}

curl -X DELETE http://localhost:8000/session/<session_id>

GET /health

curl -s http://localhost:8000/health

Test

uv run pytest -q

Layout

app/
  main.py              FastAPI app, logging, health
  config.py            Settings (pydantic-settings)
  api/                 HTTP routers
  models/schemas.py    Pydantic request/response models
  core/                input_processor, context_manager, prompt_builder, agent_client
  storage/             SessionStore protocol + in-memory implementation
  utils/               logging + token estimation
tests/                 pytest suite (agent backend mocked)
docs/                  Design documentation (architecture, data model, API, ADRs)

Design documentation

See docs/ for the full design package:

This server cannot be deployed

One endpoint for every MCP toolRayrun runs every MCP server behind one URL and holds the credentials. Each call is decided per client, tool, and argument, then recorded.ray.runAd

Maintenance

ActivityMaintained
ResponsivenessNo issues

Resources

Related MCP Connectors

Related MCP Servers

  • MCP Platform

    jck411
    F
    license
    Not graded
    quality
    Not graded
    maintenance
    High-performance Model Context Protocol server supporting multiple LLM providers (OpenRouter, OpenAI, Groq) with WebSocket API and conversation history persistence.
    -

  • MCP Server with LLM

    MelaLitho
    A
    license
    Not graded
    quality
    D
    maintenance
    Enables chat with multiple LLM providers (OpenAI and Anthropic) while maintaining persistent conversation memory. Provides extensible tool framework for various operations including echo functionality and conversation storage/retrieval.
    MIT