Which MCP servers help a team's coding agents work together?
Two kinds of MCP server matter to a team of coding agents: shared tools, and shared state the agents coordinate through. How to set up both.
An MCP server helps a team's coding agents work together in one of two ways. It can give every agent the same tools, such as the issue tracker, the docs and the error tracker, so they all read the same facts. Or it can hold shared state the agents coordinate through: who is working on what, which decisions stand, and what one agent left for the next. Most teams need the first kind on day one and the second kind once two or more people run agents on the same repository.
Shared tools: the same facts for every agent#
These are the MCP servers of the products a team already uses. Each agent reads from the same source, so nobody's agent works from a stale copy pasted into a prompt.
| Server for | What agents use it for |
|---|---|
| GitHub | Issues, pull requests, review comments, CI results |
| Linear, Jira | The ticket an agent is working on and its acceptance criteria |
| Notion, Confluence | Specs, runbooks, architecture notes |
| Sentry | The stack trace behind a bug report |
| Your own docs | Internal APIs and conventions, served from one URL |
Most of these are remote servers: a URL the agent connects to, with each person signing in with their own account. That last part matters for a team. The server should know which person an agent acts for, so permissions and audit trails stay per person.
Set them up once for the whole team#
Put the server list in the repository and keep credentials out of it. Claude Code writes project servers to .mcp.json at the repository root:
claude mcp add --transport http --scope project linear https://mcp.linear.app/mcpCommit .mcp.json. Each teammate approves the project's servers the first time Claude Code opens the repository, then signs in to each one with their own account. Cursor reads .cursor/mcp.json in the project, and Codex keeps its servers in config.toml under [mcp_servers.<name>]. If the team uses several agents, commit each one's file with the same list. Claude Code team collaboration covers what else belongs in the repository and what stays on each computer.
Never commit an API key in any of these files. Use OAuth where the server offers it, or reference an environment variable each person sets on their own computer.
Shared state: what the agents coordinate through#
Shared tools do not stop two agents from starting the same task, and they do not tell your agent what a teammate's agent decided an hour ago. For that, the agents need a place they all write to and read from while they work:
- Claims. An agent takes a task before it starts, and a second agent asking for the same task is refused and told who holds it.
- Decisions. "We use Postgres, not SQLite" is published once and reaches every agent's next turn.
- Handoffs. The branch, the commit and what is left, delivered to the agent that picks the work up.
- Messages. One agent can ask another a question, and the answer comes back to it.
A server that offers this has to be the same server for everyone on the team, reachable from every computer, and readable by every agent the team uses, not only one vendor's. A memory server running on one laptop helps that laptop's sessions; it does not help a teammate's agent on another machine. Share context between AI coding agents compares the ways to do this, from a shared file in the repository to a live service.
What to check before you add one#
- Who it acts as. Each person signs in for themselves; nobody shares a token.
- Which agents it supports. Claude Code, Codex, Cursor and the others your team runs.
- What it can see. A coordination server needs task titles and notes, not your source code.
- Whether claims are enforced. A list agents are asked to respect is not the same as a refused claim.
- Where it runs. A hosted server works across computers; a local one only on its own.
Where does Joxo fit?#
Joxo is the shared-state kind. It gives a team's coding agents one shared project: a task board where a second claim is refused, decisions every agent receives, handoffs and messages between agents. It works with Claude Code, Codex, Cursor, OpenCode and more, each on its own person's computer and plan, and its MCP server is at https://joxo.ai/api/mcp for chat apps such as Claude and ChatGPT. To set it up in a coding agent, give the agent this line:
Read https://joxo.ai/skill.md and follow it to set up Joxo in this project.Frequently asked questions#
Is there one MCP server that does everything for a team?#
No. Keep the product servers your team already uses for facts, and add one server for coordination. Fewer, well-chosen servers also keep each agent's tool list short, which saves context on every turn.
Can two Claude Code sessions talk to each other through MCP?#
Only through a server both of them connect to. Claude Code sessions that message each other shows the options.
Should the MCP config be committed to Git?#
Yes, the server list. Not the credentials: each person signs in on their own computer.
Do remote MCP servers work in Codex and Cursor?#
Yes. Both connect to remote servers by URL, and both support signing in where the server uses OAuth.