MCP Plumbing Landscape
A technical reference for builders
What infrastructure is emerging around MCP, what’s gaining traction, and what it means for anyone building on or around the protocol.
Week of March 26, 2026. All data sourced from PT-Edge‘s live tracking of 166K+ AI repos across GitHub, PyPI, npm, Docker Hub, HuggingFace, and Hacker News.
Context: The Token Burn Problem
MCP was designed for rich IDE integrations where context window cost wasn’t the primary constraint. Every tool a server exposes injects its full JSON schema — parameter names, types, descriptions, constraints — into the LLM’s context on every turn. A server with 30 tools can burn thousands of tokens just on schema descriptions before the model does anything useful.
As agents get more autonomous, run longer sessions, and connect to more servers simultaneously, this overhead becomes a genuine bottleneck. The community is now patching this from multiple angles, and it’s driving some of the fastest-growing repos in the entire AI ecosystem.
Beyond token efficiency, the ecosystem is also hitting scaling pain around server management, authentication, and transport interop. The pattern mirrors what REST APIs went through over the last decade — except compressed into months.
1. Token Efficiency Tools
The hottest frontier. These tools reduce or eliminate the context window cost of MCP tool schemas.
mcp2cli — knowsuchagency/mcp2cli
1,684 stars, +18.7% this week · Python · MIT
Turns any MCP server, OpenAPI spec, or GraphQL endpoint into a CLI at runtime with zero codegen
The agent calls bash commands instead of consuming tool schemas — the LLM doesn’t see the schema at all
Claims 96–99% token savings versus native MCP tool use
How it works under the hood:
At invocation, connects to the MCP server (HTTP/SSE or stdio) or parses an OpenAPI/GraphQL spec
Dynamically builds a CLI command tree from discovered tools/endpoints — all in-memory, no codegen step
Maps CLI flags to tool parameters
Calls the underlying API and returns results to stdout
Caches schemas for 1 hour in
~/.cache/mcp2cli/
“Bake mode” saves named connections:
mcp2cli bake create petstore --spec https://api.example.com/spec.jsonthenmcp2cli @petstore list-pets --limit 10Supports OAuth (authcode + PKCE, client credentials), API key headers, and secret injection via
env:andfile:prefixesHas a TOON output encoding mode that yields 40–60% fewer tokens than JSON for uniform arrays
Agent skill integration via
npx skills add knowsuchagency/mcp2cli
dbhub — bytebase/dbhub
2,287 stars · TypeScript
“Zero-dependency, token-efficient database MCP server”
Supports Postgres, MySQL, SQL Server, MariaDB, SQLite
Explicitly optimized for minimal schema overhead when agents talk to databases
The key design choice: instead of exposing every possible database operation as a separate tool with verbose schemas, it exposes a minimal surface area and lets the agent write SQL directly
context7 — upstash/context7
50,614 stars, +937 this week
Not MCP plumbing per se, but solves the adjacent problem: up-to-date code documentation formatted specifically for LLM consumption
Instead of the agent ingesting raw docs and burning context on irrelevant sections, context7 pre-processes documentation into LLM-optimized chunks
Used as a reference source by coding agents to avoid hallucinating outdated APIs
2. Gateways & Aggregators
As teams spin up 5, 10, 20+ MCP servers, managing them becomes its own problem. This is the “service mesh” moment.
mcpo — open-webui/mcpo
4,041 stars · Python
MCP-to-OpenAPI proxy from the Open WebUI team
Wraps any MCP server in a standard REST API so non-MCP clients can consume it
Single job, minimal surface area, well-executed
metamcp — metatool-ai/metamcp
2,100 stars, +1.5%/week · TypeScript
Aggregator, orchestrator, middleware, and gateway in one Docker container
The kitchen-sink approach — routes requests across multiple MCP servers, handles tool discovery, provides a unified interface
archestra — archestra-ai/archestra
3,548 stars · TypeScript
Enterprise-oriented: guardrails, MCP registry, gateway, and orchestrator
Targeting teams that need governance and policy enforcement over their MCP fleet
mcp-context-forge — IBM/mcp-context-forge
3,393 stars, +1%/week · Python
IBM’s entry into the space
Unified endpoint sitting in front of any MCP, A2A, or REST/gRPC APIs
Centralized discovery, guardrails, and access control
The “enterprise middleware” DNA is obvious — this is IBM doing what IBM does
mcphub — samanhappy/mcphub
1,866 stars · TypeScript
Centralized hub for routing to multiple MCP servers with flexible routing strategies
Essentially an MCP load balancer and service directory
Unla — AmoyLab/Unla
2,051 stars · TypeScript
MCP gateway that transforms existing APIs into MCP servers with zero code changes
The inverse of mcpo: instead of MCP→REST, it’s REST→MCP
Features routing, caching, and tool aggregation
casibase — casibase/casibase
4,461 stars · Go
Open-source “AI Cloud OS” — knowledge base plus MCP/A2A management platform with admin UI
Broader scope than pure gateway, but includes MCP server fleet management
3. Credential & Auth Management
MCP’s auth story is still immature. The spec added OAuth support, but practical tooling lags badly. This category’s growth rate is the market screaming.
onecli — onecli/onecli
1,301 stars, +54.1% this week — fastest percentage grower in the entire MCP ecosystem · MIT
Open-source credential vault for AI agents
Agents access services without the raw keys ever being exposed to the LLM or written into prompts
Solves the problem that most MCP servers either hardcode credentials, require them in environment variables, or don’t handle auth at all
fastapi_mcp — tadata-org/fastapi_mcp
11,644 stars · Python
Expose FastAPI endpoints as MCP tools, with auth baked in
The “with Auth!” in the description is doing heavy lifting — most MCP server frameworks ship auth as an afterthought or not at all
Handles OAuth flows and API key validation at the MCP layer
Klavis — Klavis-AI/klavis
5,660 stars · YC X25 · Python
MCP integration platform focused on reliable tool use “at any scale”
The YC backing signals this is a serious company bet, not a weekend project
Emphasis on production reliability and enterprise-grade auth
4. Transport Bridges
MCP has multiple transport mechanisms (stdio, HTTP/SSE, Streamable HTTP) and most servers only implement one. These tools bridge the gaps.
mcp-proxy — sparfenyuk/mcp-proxy
2,338 stars · Python
Bridges Streamable HTTP and stdio MCP transports
Sounds boring, is essential — lets HTTP-only clients talk to stdio-only servers and vice versa
Critical for connecting cloud-hosted agents to locally-running MCP servers
markdownify-mcp — zcaceres/markdownify-mcp
2,444 stars · TypeScript
Converts almost anything to Markdown via MCP
Universal format adapter — PDFs, HTML, Office docs, images → Markdown → LLM-ready
google_workspace_mcp — taylorwilsdon/google_workspace_mcp
1,771 stars, +4.2%/week
Gmail, Calendar, Docs, Sheets, Slides, Chat, Forms, Tasks, Search, and Drive — all via MCP
One of the most comprehensive single-server implementations out there
The growth rate suggests there’s massive demand for “give my agent access to my whole Google workspace” as a single integration
xiaohongshu-mcp — xpzouying/xiaohongshu-mcp
12,217 stars, +345 this week
MCP server for Xiaohongshu (RedNote, China’s Instagram-equivalent)
Surprisingly large star count for a single-platform connector — indicates the size of the Chinese MCP developer community
Companion tool XHS-Downloader (10.3K stars) handles content extraction
5. Domain-Specific Bridges
Where MCP meets actual professional workflows. These are not general-purpose — they wire AI agents into specific tools that practitioners already use daily.
chrome-devtools-mcp — ChromeDevTools/chrome-devtools-mcp
31,572 stars, +1,413 this week
Chrome DevTools for coding agents
Lets AI agents inspect, debug, and interact with web pages through the Chrome DevTools Protocol via MCP
Massive and still accelerating — the single largest MCP-specific tool by stars after the official servers
The growth tells you web development is probably the highest-volume MCP use case right now
unity-mcp — CoplayDev/unity-mcp
7,573 stars, +4.3%/week
Bridge between Unity Editor and AI assistants (Claude, Cursor, etc.)
Manage assets, control scenes, edit scripts, automate tasks — all from the AI agent
Game dev meets agents
ida-pro-mcp — mrexodia/ida-pro-mcp
6,708 stars, +3.4%/week · Python
AI-powered reverse engineering assistant bridging IDA Pro and LLMs through MCP
Niche audience (security researchers, malware analysts) but deeply passionate
jadx-ai-mcp — zinja-coder/jadx-ai-mcp
1,437 stars, +5.4%/week
Same concept for JADX (Android APK decompiler)
Security researchers and Android reverse engineers are early, enthusiastic MCP adopters
hexstrike-ai — 0x4m4/hexstrike-ai
7,377 stars · Python
MCP server for autonomous cybersecurity tool execution — 150+ security tools accessible to AI agents
Covers penetration testing, vulnerability scanning, network recon
The security community is aggressively adopting MCP as the interface layer between AI and offensive/defensive tooling
6. Agent Orchestration Platforms (MCP-native)
A layer above plumbing — these use MCP as the backbone for multi-agent coordination.
ruflo — ruvnet/ruflo
26,418 stars, +4,730 this week (+21.8%) · TypeScript
Agent orchestration platform built around Claude
Deploy multi-agent swarms, coordinate autonomous workflows
The growth rate is remarkable for an orchestration framework — suggests demand for “deploy many agents that share MCP tools”
n8n-mcp — czlonkowski/n8n-mcp
16,421 stars, +6.7%/week
MCP server that lets Claude Code/Desktop/Cursor build n8n workflows programmatically
Bridges the gap between AI agents and visual workflow automation
Effectively turns n8n’s 400+ integrations into agent-accessible tools
Skill_Seekers — yusufkaraaslan/Skill_Seekers
11,322 stars, +2.9%/week · Python
Converts documentation websites, GitHub repos, and PDFs into Claude AI skills
Automatic conflict detection when skills overlap
The meta-tool: it generates the tools that agents use
trigger.dev — triggerdotdev/trigger.dev
14,223 stars, +1.1%/week · TypeScript
Build and deploy fully-managed AI agents and workflows
MCP-native scheduling and execution infrastructure
7. Emerging Patterns Worth Watching
Subscription Arbitrage Infrastructure
sub2api — 8,528 stars, +23.3%/week
claude-relay-service — 10,104 stars, +3.3%/week
Chinese-language repos for pooling and proxying Claude/OpenAI/Gemini subscriptions
“Cost sharing” is the euphemism; subscription arbitrage is the reality
Both growing fast — indicates demand for AI API access that official pricing isn’t meeting
These will likely interact with MCP infrastructure as agents need authenticated API access at lower cost
The Physical Edge
xiaozhi-esp32 (24.9K stars) — MCP-based chatbot running on ESP32 microcontrollers
osaurus (4K stars) — AI edge infrastructure for macOS, local/cloud model runtime with MCP tool sharing
Signals that MCP is being adopted as the protocol for on-device and edge AI, not just cloud agents
Observability
mcp-use (9.4K stars) — “Fullstack MCP framework” that includes monitoring and observability for MCP tool calls
bifrost (maximhq, 2.9K stars) — AI gateway claiming 50x faster than LiteLLM with <100µs overhead, cluster mode, guardrails
The observability layer is thin right now but will grow fast as production MCP deployments scale
The Big Picture
The MCP ecosystem is recapitulating the entire API infrastructure stack in compressed time:
2024–early 2025: Individual servers. Everyone built their own MCP server for their own tool.
Late 2025–now: Too many servers → gateways, registries, and aggregators emerge.
Now: Performance problems → token-efficient adapters and CLI bridges.
Now, barely: Security gaps → auth vaults and credential management.
Next: Observability, rate limiting, billing, metering, and compliance tooling.
The fastest-growing tools are all solving problems the MCP spec doesn’t address. Token efficiency, credential management, server aggregation, transport bridging — the community is building the middleware layer that the protocol left as an exercise for the reader.
For builders, the implications are:
Design for token efficiency from day one. Minimal schemas, compact responses, no verbose descriptions unless requested. dbhub’s approach (thin surface area, let the agent write SQL) beats the “expose 50 tools” pattern.
You’ll need a gateway sooner than you think. If you’re deploying multiple MCP servers, mcpo and metamcp are the leading options depending on whether you want minimal or maximal.
Auth is undersolved and high-demand. Onecli’s 54% weekly growth with a simple credential vault is the clearest market signal in the whole ecosystem.
Transport bridging is unsexy but critical. If your server only speaks stdio, half your potential clients can’t reach it. mcp-proxy exists for a reason.
Watch the security community. They’re ahead of most verticals in MCP adoption. If you want to see where production MCP usage patterns are heading, watch the security tooling repos — they’re building the most sophisticated multi-tool agent workflows.
Generated March 26, 2026 from PT-Edge live data. Star counts and growth figures cover the 7-day window ending March 26, 2026.