Skip to content
Atlas by Comet Lab
Capability layer

MCP and connectors, compared

The Model Context Protocol is an open standard for connecting models to tools and data, so a capability built once works with every MCP-aware client. Around it sit frameworks for building servers, libraries for using them, and platforms that handle sign-in to third-party apps.

Learn the idea How AI connects to company systems, and what MCP is on Horizon

Open this layer in the explorer

At a glance

Every tool in this layer, with where it fits and the facts that usually decide it.

Tool Where it fits Licence Self-hostable Hosted API Maturity
Model Context Protocol Open protocol The right approach when a capability should be reusable across many models and clients, rather than wired into one app. Open source Yes No Growing
Arcade Agent authorisation A fit when agents must take actions as a specific end user and per-user OAuth and permissions have to be correct. Proprietary No Yes Growing
Composio Managed connector platform A strong fit when an agent needs to act across many third-party apps and you want auth and connector upkeep handled for you. Proprietary No Yes Growing
FastMCP MCP server framework The default when you are building a custom MCP server in Python and want to write tools, not protocol plumbing. Open source Yes No Growing
Klavis AI MCP infrastructure A fit when you want reliable, authenticated MCP servers for common apps and the option to self-host them. Open source Yes Yes Growing
mcp-use MCP client library A fit when you are building an agent and want to consume existing MCP servers without writing client code. Open source Yes No Growing
Pipedream Connector platform A fit when an agent needs the widest possible catalogue of third-party app integrations with auth managed. Proprietary No Yes Established
Smithery MCP registry A fit when you want to find and connect existing MCP servers quickly, without hosting them yourself. Proprietary No Yes Growing

How to choose

Start from your situation. Where a line says “our”, it comes from Comet Lab’s own builds.

  1. A capability many models and clients should share

    Build it as an MCP server, as we now do by default. For one private integration called by one app, a direct API call is simpler.

  2. Writing your own MCP server in Python

    What we write MCP servers with: the code stays about the capability, not the protocol.

  3. Your agent uses existing MCP servers

    mcp-use connects an LLM to MCP servers in a few lines; Smithery is a registry for finding them and a host for running them.

  4. The agent acts across many third-party apps

    Auth and connector upkeep handled for you. Pipedream has the widest catalogue; Klavis can be self-hosted.

  5. The agent acts as a specific user

    Per-user OAuth and permissions, handled correctly.