Claude Desktop is most teams’ first AI-native chat surface. It’s where the prompt-engineer-y stuff happens, where the long conversations live, where you’d ask the agent to think with you about a hard problem before writing the code. So it’s also the natural place to ask: “given everything in the sprint, what should we be doing right now?”

Until recently the answer was: copy-paste sprint state into the chat by hand. Now Claude Desktop speaks MCP, and SprintFlint runs a public MCP server. So the answer is: connect once, ask whatever you want.

This is the practitioner’s guide. Not the marketing pitch. What actually changes for an engineering manager (or a tech lead, or a solo dev running their own sprints) when sprint state lives in Claude Desktop, and what’s not worth using it for.

The 30-second setup

We’re keeping this short on purpose. Three steps:

1. Generate a SprintFlint API token from your account settings.

2. Open ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows). Add this block:

{
  "mcpServers": {
    "sprintflint": {
      "url": "https://mcp.sprintflint.com",
      "headers": { "Authorization": "Bearer YOUR_TOKEN" }
    }
  }
}

3. Restart Claude Desktop. The SprintFlint tools appear in the tool menu (top-right of the chat). You’re done.

If you’d rather not edit JSON by hand, the MCP Config Generator emits the same block (and the equivalent for Cursor, Claude Code, and Zed) with one click.

Five real use cases

Setup done. Now what’s actually worth doing? Five patterns we’ve seen settle into regular use after a week.

1. The Monday-morning “where are we?” prompt

The first prompt of the week, before standup, before opening the board:

“Pull the active sprint from SprintFlint. List tickets by status. Highlight anything blocked or stale. Tell me what I should be paying attention to today.”

Claude calls list-sprints → list-issues → reads each get-issue for tickets older than 3 days → comes back with a prioritised triage. The ML doesn’t add information you couldn’t get yourself, but it compresses what would be five tab-switches and a notepad into one chat.

This is the use case that justifies the setup on its own. Five minutes saved per Monday × 50 Mondays = a working day.

2. Sprint-aware “what should I work on next?”

The mid-day question every engineer asks themselves at least once.

“What should I pick up next? My current branch is auth-rate-limit, I’m waiting on review for it. Look at the SprintFlint sprint and recommend the next ticket given my context.”

Claude pulls the active sprint, filters to tickets assigned to you (or unassigned, depending on team norm), reads the descriptions, and recommends one with reasoning. Useful because the agent has both the code context (your current branch, your open files if you’ve shared them) and the sprint context. Neither side alone is enough.

3. Bug-report → ticket pipeline

Customer reports a bug in Slack. You paste the message into Claude Desktop:

“Customer report: ‘when I try to delete an empty project, the page hangs and I have to refresh.’ File a SprintFlint ticket — well-formed acceptance criteria, link to the customer’s email, set priority to High since they’re an enterprise account.”

Claude calls create-issue with a properly-structured payload — title, description in markdown, acceptance criteria as a checklist, priority field set. You read it back, edit if needed, accept. Five-minute task done in 30 seconds, and the ticket is consistent with everything else on the board because the agent has read the existing tickets first.

4. PM-ish status updates without becoming PM

The Wednesday afternoon Slack ping from the GM: “how’s the sprint looking?” You don’t want to spend 20 minutes in the board. So:

“Write me a 5-line sprint status for the GM. Don’t speculate — only use what’s in SprintFlint. Mention the goal, percent done, biggest open risk.”

Claude pulls the sprint goal, computes the done/total ratio from list-issues, finds the oldest in-progress ticket as a stand-in for “biggest risk”, writes the five lines. You proofread, send. The agent doesn’t make up numbers because it can’t — it only has what the API gave it.

5. Retro prep without homework

Friday before retro:

“Pull the closing sprint. List tickets that took more than 5 days from in-progress to done. Find tickets that got moved between sprints (look at the comments for clues). Suggest 3 retro discussion topics.”

Some of this requires guessing — moved between sprints isn’t always cleanly tracked — but the agent will say so, and the topics it suggests are 80% of the way to a retro you didn’t have to prep for.

What it’s NOT good at

Three failure modes worth flagging early so you don’t burn a week on them.

“Strategic” or stakeholder questions

Don’t ask Claude “should we add this feature to next sprint?” The agent has the sprint state, not the business context, the customer-success queue, the marketing calendar, or the mental model of why the team chose what it chose. The right shape is: agent surfaces tickets, human prioritises.

Anything destructive

We didn’t ship delete-issue or delete-sprint on day one — both have an asymmetric failure mode where a hallucinated argument silently destroys work. So if you ask Claude to “clean up old tickets”, the right answer is “here are the candidates, you click delete in the UI”. Update operations are available (update-issue), because every mutation is reversible from the UI in two clicks.

Long-horizon forecasting

The MCP tools surface the current sprint cleanly. They don’t expose historical velocity by sprint, cycle-time distributions, or burndown curves — those are easier to read on a chart than to summarise via tool calls. For forecasting, use the Sprint Forecaster and Cycle Time Calculator, which are designed for that shape of data.

A note on tokens

Bearer tokens have full org access. If you connect Claude Desktop on your work laptop and your home laptop and your iPad, that’s three places your token lives. Best practice:

• One token per device, named (claude-desktop-macbook, claude-desktop-home)
• Revoke immediately if a device is lost
• Don’t share your token with teammates — they should generate their own (free tier supports it)

The account settings page lists every active token with its last-used time. Revisit monthly.

The honest summary

Claude Desktop with MCP is not magic. It doesn’t make decisions for you. It doesn’t replace the board UI for tasks that need spatial reasoning (drag-drop, multi-ticket selection). What it does is reduce the friction between “I have a question about my sprint” and “I have an answer” from five clicks to zero. Five clicks isn’t much. But you do it forty times a day.

Connect once. See if it sticks. The setup takes 30 seconds; if it doesn’t help by the end of the first week, it won’t.

Try it

• Generate a token: SprintFlint account settings (free tier available)
• Copy the config: MCP Config Generator
• Read the protocol primer: Sprint state in your agent — what changes when MCP makes the board reachable from the editor
• See per-editor instructions: /integrations