A Claude Code team workflow on Jira, Linear or GitHub

Updated

Your tracker already holds what Claude Code needs to know: the task, its parent and the decisions already on it. CRUX lets Claude Code read them over MCP and gives each status in your workflow its own instructions, so nobody writes a prompt. This page is for teams that use Claude Code with Jira Cloud, Linear, Plane or GitHub Issues.

What goes wrong with Claude Code alone in a team

On Monday Ryan, a developer, and Claude Code agreed to keep retries in the queue worker. On Thursday he opens a new session for the next task. That session has no record of Monday, so it moves the retries into the API handler.

For every task he types the same lines again: use the review skill, run the tests, open a PR. By the afternoon he has three sessions in three tabs and cannot say which one waits for his answer. Megan, who reviews his pull request, sees the result but not the plan behind it.

Setting up Claude Code for a small team

  1. Sign up for CRUX and create a project tied to your tracker project, for example Jira Cloud, Linear, Plane or GitHub Issues. Each person connects their own tracker account.
  2. A project admin sets up the workflow with /crux:crux-config, described below.
  3. Each developer opens "Connect your coding agent" in the dashboard, picks the project and Claude Code, and presses "Create my token". The token is shown once, works for one project and acts as you.
  4. Each developer adds the two files below and installs the plugin.

Commit .mcp.json to the repository. It reads the token from ${CRUX_MCP_TOKEN}, so the file holds no secret and every teammate uses their own token.

{
  "mcpServers": {
    "crux": {
      "type": "http",
      "url": "https://crux.team/mcp/p/PROJECT_ID",
      "headers": {
        "Authorization": "Bearer ${CRUX_MCP_TOKEN}"
      }
    }
  }
}

Your own token goes into .claude/settings.local.json, which stays out of git. The dialog writes this file for you, with your token in place of YOUR_TOKEN. It also trusts the project's crux server and lets its tools run without asking. Check that the file is ignored with git check-ignore .claude/settings.local.json before you commit. If you would rather keep the token out of files, set CRUX_MCP_TOKEN in your shell profile or a secrets manager instead.

{
  "env": {
    "CRUX_MCP_TOKEN": "YOUR_TOKEN"
  },
  "enabledMcpjsonServers": [
    "crux"
  ],
  "permissions": {
    "allow": [
      "mcp__crux__*"
    ]
  }
}

Restart Claude Code in the repository, approve the project's server, then check it:

claude mcp list

Team best practices beyond a shared CLAUDE.md

Anthropic's advice for a team is a good start: commit a project CLAUDE.md, share .claude/settings.json and the project's MCP servers in .mcp.json, and follow its best practices. Do that first.

A shared rules file tells every session how the team writes code. It does not say who answers the agent's question, where the agreed plan stays, or who moves the task to the next status. CRUX covers those parts:

QuestionA shared CLAUDE.md and skillsCRUX
Where the task's context comes fromPasted into each sessionRead over MCP: the task, its parent and the decisions already on it
Who writes the prompt for a stepEach developerNobody: each status in your workflow has its own instructions
Where an agreed plan staysIn one session's historyOn the task, as an approved artifact
Who answers the agent's questionWhoever runs the sessionThe person whose role owns the status
Who sees what is in progressThe person at that terminalThe team, on the board

The CRUX plugin for Claude Code

The plugin needs a CRUX account and a connected project, because every skill calls your project's CRUX MCP server. It is open source under the MIT license:

claude plugin marketplace add cruxteam/skills
claude plugin install crux@cruxteam

Its skills:

  • crux-config sets up the project's workflow through an interview.
  • crux-run claims one workflow step of a task and opens a session for it.
  • crux-dispatch shows the board for your role and opens one session for each task you pick.
  • crux-worktree moves the work into the task's own git worktree.
  • crux-retro runs a retrospective over a sprint and turns the recommendations you pick into tasks.

A project admin runs /crux:crux-config from a checkout of the project's repository. Its config tools are hidden from other members. The skill asks who owns each tracker status and where the approval checkpoints go, then opens a pull request, and the workflow starts once you merge it. A project without a repository imports the workflow directly.

Start a step and approve at checkpoints

A person starts every step. In Claude Code, type:

/crux:crux-run CR-12 developer

Here CR-12 is the task and developer is the agent. The skill claims the step and creates the git worktree .claude/worktrees/CR-12 on the branch CR-12, so parallel tasks stay apart. In a terminal it prints one command: run it in a new tab to open the session. In the Claude desktop app it shows a session card to click instead. The session fetches its instructions from CRUX, so you paste no prompt.

At a checkpoint the agent shows you its understanding, its plan or its result, and posts the artifact to the task only after you approve. You move the task to the next status yourself. The board shows each step as not started, running or needs you.

Project management for Claude Code: keep it in your tracker

Your tasks, statuses and sprints stay in the tracker your team already uses, and there is no second board to keep in sync. CRUX writes comments and artifacts to the task and never changes your tracker's settings. The tracker stays the one place where the team sees what is planned, what is in progress and what was decided.

What Claude Code can read and change

Through MCP the agent reads a task with its child tasks, comments and attachments, and searches other tasks. It creates and updates tasks and adds comments. It changes a status or an estimate only with your consent.

CRUX has no bot account, so the agent acts as you. Text that the agent writes starts with a robot emoji, so readers can tell it apart.

Claude Code reads the project wiki from your repository checkout, like any other file. CRUX does not serve the wiki over MCP.

The same project works with Codex and Cursor, and the dialog also has entries for Gemini CLI, OpenCode and Zed.

Limits

  • CRUX is in beta, and one founder builds it.
  • CRUX is a hosted service only. It does not support Jira Server or Data Center.
  • Claude Code runs on your machine and on your own subscription. CRUX sells no usage credits.
  • A token works for one project and gives the agent your access to it. You revoke it in the dashboard under MCP clients.

Next step

Start free, then follow the article Connect your coding agent.

FAQ

Is this the Claude Team plan?

No. The Team plan is Anthropic's paid Claude subscription for organizations, and it includes Claude Code. CRUX is a separate service that works with the Claude Code plan each developer already has.

Is this the same as Claude Code agent teams?

No. Agent teams is an experimental Claude Code feature in which one session leads other Claude Code sessions on a shared task list. CRUX organizes the people on a team around the tracker: each person starts the steps of the statuses their role owns.

Does Claude Code remember the project between sessions?

Each session still starts empty. Each step posts an approved artifact to the task, and the next session reads it over MCP. The article on a project wiki for AI agents covers the rest of a project's memory.

Does the agent pick up tasks on its own?

No. A person starts each step.

What does it cost?

One person uses CRUX free. A team of 2 to 10 people pays $49 per active project a month, and a team of 11 or more pays $120. The first month of every project is free, and during the beta nothing is charged. The details are on the pricing page.