Claude Code for teams comes down to one rule: the repo is the shared memory, and everything else is per person. A team shares what it commits (CLAUDE.md, rules, skills, MCP config, settings) plus policy an admin pushes to every machine, and it shares nothing that Claude learns or does in a session. We run a two-founder company on Claude Code, so below is what the docs say a team gets, what it doesn't, and how we cover the gap. Docs checked 2026-10-21.
What does Claude Code share with a team natively?
Anything you commit to the repo, plus organization policy. Every teammate who clones the repo gets:
- Project instructions:
./CLAUDE.mdor./.claude/CLAUDE.md, plus rules in.claude/rules/(memory docs). - Project settings:
.claude/settings.jsonfor permissions, hooks and plugins, with a personal override in.claude/settings.local.json(settings docs). - MCP servers:
.mcp.json, written byclaude mcp add --scope project; each person approves these servers before first use (MCP docs). - Skills:
.claude/skills/<name>/SKILL.md(skills docs).
Admins get two more layers. A managed CLAUDE.md, deployed with MDM or set through the claudeMd key in managed settings, applies to every user on the machine and can't be skipped with claudeMdExcludes. Server-managed settings, set by Owners in the claude.ai admin console, reach every member on Team and Enterprise (server-managed settings).
What stays private to each developer?
Everything Claude learns and everything you do in a session. A feature request on GitHub opens with the same point: "Claude Code's memory system is individual-only" (#38536).
- Auto memory in
~/.claude/projects/<project>/memory/, on your machine. - Personal instructions:
~/.claude/CLAUDE.mdandCLAUDE.local.md. - Session transcripts, stored locally per project.
- MCP servers added at local or user scope, in
~/.claude.json. - Claude Code projects (public beta on Pro and Max): "A project belongs to one user" (projects docs).
- Agent teams (experimental): one session's teammates, with a task list in
~/.claude/tasks/on your machine (agent teams docs).
That line is why developers keep posting their own shared memory servers and repo-based memory systems for Claude Code.
How do I roll Claude Code out to a team?
Treat the repo as the shared memory and review it like code:
- Run
/init, then cut the project CLAUDE.md to build commands, conventions and "always do X" rules, under 200 lines. - Commit
.claude/settings.json,.mcp.json(servers added with--scope project) and skills in.claude/skills/. - For a plugin everyone needs, choose Install for all collaborators on this repository in
/plugin. That commits it to.claude/settings.json, and each collaborator still runsclaude plugin install <name>@<marketplace> --scope projectonce (plugin docs). - Ask your admin to put org-wide rules in a managed CLAUDE.md or server-managed settings.
- Run
/doctor prompt-auditnow and then to find outdated or conflicting instructions.
Our version of step 1 is stricter than the docs. Every CLAUDE.md we have is around 150 lines and works as a router: the basics, then pointers to task-specific files that load when that's the work. We keep auto memory off on every machine, because it applied stale decisions to new work; Claude Code memory has the detail. The shared rules are the same files for both of us, and the AGENTS.md next to each is a symlink, so a teammate on Codex or Cursor would read them too.
Is Claude Code included in the Claude Team plan?
Yes. "Claude Code is included with every Team plan seat," and premium seats add usage for heavier workloads (Claude help). Run claude, choose "Claude account with subscription", pick the Team plan and authorize. Skills and plugins from your Claude account, including ones your organization provides, then load in Claude Code automatically (v2.1.273 or later), so org-wide skills reach every member without git. Memory still doesn't: each member's auto memory and sessions stay their own.
What is Claude Code agent teams mode?
An experimental mode where one session, the lead, spawns teammates that each run in their own context window, share a task list and message each other (agent teams docs). Turn it on with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1; split panes need tmux or iTerm2. Teammates load CLAUDE.md, MCP servers and skills but not the lead's conversation, and the mode uses a lot more tokens than one session. It's a team of agents, not of people: one session, one team, task list on your machine. We haven't tried it yet; our parallelism comes from two people's agents working one shared task list, which is the next section.
What does a team still need after all that?
| What a team needs | What Claude Code does natively | Gap |
|---|---|---|
| Same rules for everyone | Project CLAUDE.md, rules, settings via git; managed CLAUDE.md | None for one repo |
| Same tools and workflows | .mcp.json and .claude/skills/ via git |
None for one repo |
| Learnings shared across the team | Auto memory is per person, per machine | Not shared |
| See a teammate's session | Cloud session link with Team visibility | View only, not live |
| Shared task list across people | Agent teams task list is one session, one machine | Not shared |
| Teammates on Codex or Cursor | AGENTS.md shares instructions | No shared state |
The free fix for the third row is a docs/decisions.md log in the repo with one CLAUDE.md line ("When the team makes a decision, append a dated line to docs/decisions.md with the reason"), reviewed in pull requests, with handoffs in the PR description. That holds when everyone works in one repo and decisions can wait for a merge. It breaks when a decision has to reach another agent before a merge, when two people's agents could start the same task, or when the work isn't in a repo at all.
How two founders run it: one workspace, every agent
We're two people and more agents than that, and the thing that makes it work is that none of the shared context lives in any one tool. All of our docs are in Tokenrip workspaces, and the workspace is what both of our Claude Codes load first: the pinned decisions, the files, the open tasks and the last handoff note. When one of us makes a call, it's pinned, and the other's agent reads it on its next load without a merge or a message. Tasks get filed by whoever or whatever finds the work; a Grok Bot spots an issue and files a task, one of our Claude Codes picks it up under a lease so the other can't take the same one, and the result is recorded where everything else can see it. ChatGPT reads the same workspace when it builds a deck. None of this touches auto memory, which we keep off.
The repo still carries the rules, and .mcp.json carries the connection: claude mcp add --transport http tokenrip --scope project https://api.tokenrip.com/mcp, committed, so a new teammate gets the server and signs in with their own account from /mcp. Teammates join a workspace free and bring their own agents, Claude Code or otherwise. To connect yours, paste this into it; you give an email address and a six-digit code.
Connect me to Tokenrip. Fetch https://tokenrip.com/setup?ref=claude-code-for-teams and follow it.
For the longer version of how we work, read how we work with agents.