Request flow

From MCP client to Google API and back.

Overview

graph TB
    subgraph Client
        CC[Claude Code / MCP Client]
    end

    subgraph "MCP Server"
        subgraph Transport
            STDIO[Stdio Transport]
            HTTP[HTTP Transport - POST /mcp]
            SSE[SSE Stream - GET /mcp]
        end

        PROTO[Protocol Layer - JSON-RPC 2.0]
        META[Metadata - _meta and Trace Context]
        DISPATCH[Request Dispatch]

        subgraph "Tool Execution"
            POLICY[Policy Engine]
            HELPERS[Helper Dispatch]
            EXEC[Execute - URL building and API calls]
            TASKS[Task Manager]
        end

        subgraph Discovery
            CACHE[Discovery Cache]
            TOOLS[Tool Builder]
        end
    end

    subgraph "Google Workspace"
        GAPI[Google APIs]
        DISC[Discovery Service]
    end

    CC -->|stdio or HTTP| Transport
    Transport --> PROTO
    PROTO --> META
    META --> DISPATCH
    DISPATCH --> POLICY
    POLICY --> HELPERS
    HELPERS --> EXEC
    EXEC --> GAPI
    DISPATCH --> TASKS
    DISPATCH --> TOOLS
    TOOLS --> CACHE
    CACHE --> DISC

How a request flows

Transport -- The client sends a JSON-RPC message over stdio or HTTP
Protocol -- The message is parsed, validated, and assigned an error code category
Metadata -- _meta is extracted for protocol version, client info, and W3C Trace Context
Dispatch -- The method is routed to the appropriate handler (tools/call, tasks/get, etc.)
Policy -- The policy engine checks service allow-list, method denylist, parameter constraints, and read-only mode
Helper dispatch -- Services with helpers (Drive, Docs, Sheets, Slides, Gmail) route to purpose-built tool functions instead of the generic executor
Execute -- The Google API URL is built, OAuth token is obtained (cached), and the request is sent
Response -- The Google API response is returned to the client as MCP tool result content

Multi-user deployment

Each user gets their own server instance with isolated credentials and policy:

graph LR
    subgraph "User A"
        PA[Policy A] --> DA[Pod A]
        CA[Credentials A] --> DA
    end
    subgraph "User B"
        PB[Policy B] --> DB[Pod B]
        CB[Credentials B] --> DB
    end
    DA --> G[Google APIs]
    DB --> G

Module map

Module	Purpose
`main.rs`	CLI args, templates, interactive wizard, policy checker
`handler.rs`	rmcp ServerHandler: tools, prompts, resources, completions, tasks, elicitation, subscriptions
`server.rs`	Tool dispatch business logic, request explanation
`tools.rs`	Tool list builder, lazy filtering, compact schema mode
`execute.rs`	Google API execution, URL building, pagination, resumable uploads, rate limiting
`helpers.rs`	Google Docs enrichment (9 tools)
`sheets_helpers.rs`	Google Sheets enrichment (14 tools)
`drive_helpers.rs`	Google Drive enrichment (9 tools)
`slides_helpers.rs`	Google Slides enrichment (9 tools)
`gmail_helpers.rs`	Gmail enrichment (10 tools)
`format.rs`	Markdown/Plain to Docs batchUpdate, doc to Markdown reverse converter
`cache.rs`	LRU + TTL in-memory cache for Sheets reads
`rate_limit.rs`	Per-service sliding-window rate limiter
`policy.rs`	JSON policy engine, constraint enforcement
`http.rs`	Axum HTTP server, SSE, probes, metrics, webhooks
`resources.rs`	MCP resources: gws:// URI scheme, resource templates
`completions.rs`	MCP completions: autocomplete for resource URIs and prompt arguments
`elicitation.rs`	MCP elicitation: structured user input
`subscriptions.rs`	MCP subscriptions: Drive watch channels, webhooks
`prompts.rs`	MCP prompts: external Markdown files, argument substitution
`tasks.rs`	Task lifecycle for resumable uploads and chunked downloads
`metrics.rs`	Prometheus counters, histograms, gauges
`meta.rs`	Request metadata, W3C Trace Context
`auth.rs`	OAuth2 credential chain with token caching
`audit.rs`	Structured JSONL audit log writer
`image_gen.rs`	Gemini image generation and Drive upload
`marp.rs`	Marp Markdown to Google Slides conversion

Edit this page on GitHub