AgentBridge

AgentBridge

A custom MCP (Model Context Protocol) server that exposes personal productivity data — Google Calendar, Gmail, and local notes — as agent-callable tools, paired with a LangGraph ReAct agent that decides when to use them, with Human-in-the-Loop approval, guardrails, and persistent memory — plus both a terminal and a web chat interface.

Built as a portfolio project to demonstrate practical multi-agent orchestration, tool-calling protocols, and hybrid RAG — not just a single chatbot wrapper.

Architecture User query (terminal or Streamlit UI) │ ▼ LangGraph ReAct agent (Groq / Anthropic Claude) · conversation persisted to SQLite — survives restarts │ ├─ no tool needed → answers directly │ └─ tool needed │ ▼ [Guardrails] blocked? ──yes──▶ rejected automatically, agent told why │ no ▼ [Human-in-the-Loop] "Approve this tool call? (y/n)" │ (only proceeds on approval) ▼ MCP client ──HTTP──▶ MCP server (localhost:8765) │ ┌─────┼─────────────┐ ▼ ▼ ▼ Calendar Gmail Notes (local RAG: (Google (Google FAISS + sentence- API) API) transformer embeddings) What's in this repo File Purpose src/agentbridge/server.py MCP server — exposes 3 tools over HTTP src/agentbridge/rag.py Local RAG layer: chunks + embeds notes/*.md, semantic search via FAISS src/agentbridge/guardrails.py Safety checks run before any tool call reaches approval agent.py LangGraph ReAct agent — terminal chat, Human-in-the-Loop, SQLite persistence app.py Streamlit web UI for the same agent, with approve/reject buttons test_client.py Standalone script to sanity-check the MCP server without the agent notes/ Sample notes indexed by search_notes Tools exposed by the MCP server get_upcoming_events — reads Google Calendar (read-only scope) search_emails — searches Gmail using Gmail's query syntax (read-only scope) search_notes — semantic search over local .md notes using a local embedding model (no API key needed, runs offline after first download) Safety features Human-in-the-Loop: every tool call pauses for explicit approval before it runs (interrupt_before=["tools"] in LangGraph). Guardrails (guardrails.py): blocks tool calls before they even reach approval if they request an unreasonable amount of data (possible runaway loop) or search for sensitive terms like passwords. Persistent memory: conversation history is saved to agentbridge_memory.db (SQLite) via SqliteSaver, so context survives restarting the agent. Setup

1. Install dependencies bash cd agentbridge uv sync
2. Google Cloud setup (Calendar + Gmail) Go to https://console.cloud.google.com/ → create/select project "AgentBridge" APIs & Services → Library → enable Google Calendar API and Gmail API APIs & Services → Credentials → Create Credentials → OAuth client ID Application type: Desktop app Download the JSON, rename it exactly credentials.json, place it in the agentbridge/ root folder (same level as pyproject.toml) Google Auth Platform → Audience → Test users → add your own Google account email (required while the app is in "Testing" mode)

⚠️ Never commit credentials.json or token.json — both are already in .gitignore.

3. LLM provider

Default is Groq (free):

powershell $env:GROQ_API_KEY = "your-groq-key" # get one at console.groq.com/keys

To use Claude instead:

powershell $env:LLM_PROVIDER = "anthropic" $env:ANTHROPIC_API_KEY = "your-anthropic-key" 4. Run — Terminal

Terminal 1 — MCP server:

bash uv run python -m agentbridge.server

Terminal 2 — Agent:

bash uv run python agent.py 5. Run — Web UI (optional, instead of Terminal 2) bash uv run streamlit run app.py

Opens a browser chat window with clickable Approve/Reject buttons.

Try these queries What are my upcoming meetings? Do I have any unread emails? What did we decide about the transport layer? (answered from notes/) Multi-tool chaining: Do I have any meetings tomorrow, and are there any emails related to them? — the agent will call Calendar, then Gmail, asking for approval on each. Guardrails demo: Search my emails for my password reset code — gets blocked automatically, before any approval prompt. Resume bullet

Built AgentBridge, a custom MCP (Model Context Protocol) server exposing Google Calendar, Gmail, and local notes as agent-callable tools, integrated with a LangGraph ReAct agent featuring Human-in-the-Loop tool approval, rule-based guardrails, SQLite-backed persistent memory, a local FAISS RAG layer for semantic note search, and a Streamlit web interface.

Interview talking points MCP transport tradeoffs: started with stdio, switched to streamable-http after hitting Windows-specific pipe issues. Human-in-the-Loop: implemented via LangGraph's interrupt_before, pausing before the tools node and resuming with Command(resume=True) only on explicit approval. Guardrails: a small rule-based layer that runs before approval — can discuss how this differs from (and could evolve into) a full framework like Guardrails AI or NeMo Guardrails. Persistence: swapped InMemorySaver for SqliteSaver — can explain why production agents need durable state, not just in-memory. RAG vs Vectorless RAG: search_notes uses real vector search (FAISS + sentence-transformer embeddings) because notes are numerous and stable; a vectorless approach suits smaller, fast-changing data like a single email thread better.