Claude Code for Teams: Sharing Context

Alek Perak

4 min read

·

3 days ago

--

Listen

Share

A rollout guide to Claude Code for teams: what the repo shares, what stays with each developer, and how to close the gap

By Alek Perak, founder of TokenRip

A fresh clone is the simplest test of a team’s Claude Code setup. If a second developer needs undocumented local files before the agent can build, test or understand the repo, part of the setup still lives with the first developer. Start the rollout by making that gap visible.

What does a team share in Claude Code?

A team shares what it commits to the repo, plus policy an admin pushes. Every developer who clones the repo gets:

• Project instructions: CLAUDE.md (or .claude/CLAUDE.md) and rules in .claude/rules/.
• Project settings: .claude/settings.json for permissions, hooks and plugins. Each person can override in .claude/settings.local.json.
• MCP servers: .mcp.json, written by claude mcp add — scope project. Each person approves a server before first use.
• Skills: .claude/skills/<name>/SKILL.md.

Admins have two more layers. A managed CLAUDE.md applies to every user on a machine. Server-managed settings, set in the claude.ai admin console, reach every member on Team and Enterprise (Claude Code docs (https://code.claude.com/docs/en/memory)).

What stays with each developer?

Session work stays personal. Anthropic’s docs place these on the developer’s own machine or account:

• Auto memory, in ~/.claude/projects/<project>/memory/.
• Personal instructions in ~/.claude/CLAUDE.md and CLAUDE.local.md.
• Session transcripts.
• MCP servers added at local or user scope.
• Claude Code projects (a beta), which belong to one user.

An open GitHub feature request sums it up: “Claude Code’s memory system is individual-only” (#38536 (https://github.com/anthropics/claude-code/issues/38536)). Plan around that from day one.

How to roll out Claude Code for teams in six steps

Review the shared files like code. Here is the order we’d follow:

1. Write the project CLAUDE.md. Run /init, then cut it to build commands, conventions and “always do X” rules. Aim for under 200 lines.
2. Commit the shared config. Add .claude/settings.json, .mcp.json and .claude/skills/ to the repo.
3. Install team plugins once. In /plugin, choose “Install for all collaborators on this repository.” Each developer still runs the install command once.
4. Push org rules through an admin. Use a managed CLAUDE.md or server-managed settings for rules every repo needs.
5. Audit now and then. Run /doctor prompt-audit to find outdated or conflicting instructions.
6. Decide where decisions go. This is the step most teams skip. See below.

At TokenRip we keep each CLAUDE.md around 150 lines. It works as a router: basics first, then pointers to task files. An AGENTS.md symlink sits next to each one, so a teammate on Codex or Cursor reads the same rules.

Is Claude Code included in the Claude Team plan?

Yes. Anthropic’s help center says “Claude Code is included with every Team plan seat,” and premium seats add usage (Claude help (https://support.claude.com/en/articles/11845131-use-claude-code-with-your-team-or-enterprise-plan)). Developers run claude, choose the subscription login and pick the Team plan. Skills and plugins your organization provides then load in Claude Code (v2.1.273 or later). Memory and sessions still stay with each member.

Where should team decisions live?

Decisions need a place every developer’s agent reads. The free option is a log in the repo:

1. Create docs/decisions.md.
2. Add one line to CLAUDE.md: “When the team makes a decision, append a dated line to docs/decisions.md with the reason.”
3. Review it in pull requests. Put handoffs in the PR description.

This holds when everyone works in one repo and decisions can wait for a merge. It gets harder in three cases:

• A decision must reach another agent before a merge.
• Two developers’ agents could start the same task.
• The work isn’t in a repo at all.

Does agent teams mode replace this?

No. Agent teams is an experimental mode in which one lead session spawns teammates. They share a task list and message each other. Turn it on with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. The task list lives in ~/.claude/tasks/ on one machine, so it’s a team of agents inside one session, not a team of people (agent teams docs (https://code.claude.com/docs/en/agent-teams)).

A shared workspace for the rest

Disclosure: I’m the founder of TokenRip. My cofounder and I keep our shared context in a TokenRip workspace, and it’s the first thing both of our Claude Code sessions load: pinned decisions, files, open tasks and the last handoff note. A decision one of us pins reaches the other’s agent on its next load, with no merge needed. Each task is held under a lease while an agent works on it, which keeps a second agent off it.

The repo still carries the connection. Run this once and commit the .mcp.json it writes, so every new developer gets the server:

claude mcp add --transport http tokenrip --scope project https://api.tokenrip.com/mcp

Each developer signs in with their own account from /mcp. Teammates join a workspace free. More at https://tokenrip.com.

Onboard from a clean checkout

Ask a teammate to clone the repository in a new directory and follow only the committed setup instructions. Have them list the rules that loaded, identify the test command and inspect the configured server names. They should authenticate with their own account. Record any step that required asking the original developer, then decide whether it belongs in shared instructions or remains an explicit personal setup step.