Claude Code MCP
MCP connects your Claude Code agent to tools and data. It's powerful — and it has a clear edge. Here's how to set it up and where shared context picks up where MCP leaves off.
MCP is one of the most useful things you can add to Claude Code — it's how the agent reaches beyond your codebase into databases, APIs, issue trackers, and more. It's also frequently misunderstood as "the way to give the agent context." It isn't, quite. MCP gives the agent access. Your team's decisions are a different thing. This page covers both.
MCP in 60 seconds
The Model Context Protocol is an open standard for connecting AI agents to external systems. An MCP server wraps a system — GitHub, a Postgres database, a filesystem, a SaaS API — and exposes typed tools. Claude Code launches the servers you configure and hands those tools to the agent, which can then read and act on those systems during a task.
Setting up MCP in Claude Code
You declare MCP servers in config — typically a project
.mcp.json — naming each server and the command to start
it:
{
"mcpServers": {
"github": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"] },
"postgres": { "command": "mcp-server-postgres", "args": ["$DATABASE_URL"] }
}
}On the next session, the agent has GitHub and database tools available. Popular servers exist for GitHub, filesystems, search, databases, and many SaaS products — the official MCP server directory is a good place to start.
MCP vs shared context
Here's the distinction that matters. MCP lets the agent answer "what does the database contain right now?" It does not let the agent answer "why did our team model it this way, and what did we decide not to do?" The first is live access; the second is accumulated, owned knowledge. MCP delivers the former. It was never designed to be the latter.
| MCP servers | Context tree | |
|---|---|---|
| Provides | Live access to systems + tools | Your team's decisions, ownership, rationale |
| Answers | "What's the current state?" | "What did we decide, and why?" |
| Freshness | Real-time, per query | Versioned, reviewed, owned |
| Lives | Behind a tool call | In a Git repo every agent reads |
Use them together
The strongest setup runs both: MCP for the agent to reach your systems, and a context tree for the agent to understand your team's choices about them. An agent triaging a GitHub issue uses the GitHub MCP server to read the issue and a context node to know which decisions and owners apply.
First-Tree is the orchestration platform that supplies the context half of that pair.
It's an open-source platform for engineering teams shipping with humans and AI agents side by side: agents coordinate in shared chat threads, GitHub issues and PRs become the work queue the right agent picks up, and a living context tree of decisions, designs, and ownership is the memory every agent reads — wired into Claude Code via a SessionStart hook. MCP connects the agent to your systems; First-Tree coordinates the work and makes sure each agent acts on those systems the way your team intended.
If you're standing up MCP for a whole team, pair it with the team best practices and a platform that orchestrates the work — shared chat threads, GitHub as the work queue, and a context tree every agent reads — which is the combination the AI agent teams approach is built around.