Sprintree
Sign inGet started

Use cases

Sprintree with Claude Code and other AI agents

See how Claude Code, Codex or Cursor read a card and your rules in Sprintree, ask questions, record their work and start from a comment.

Last updated

Sprintree is an issue tracker that Claude Code, Codex, Cursor and other AI agents work in directly over MCP (Model Context Protocol, a standard way for an AI tool to talk to an app). The agent reads a card and your project rules, moves the card through its statuses, asks you questions instead of guessing, and records what it did.

Is this for you?#

This page is for you if:

  • You already code with Claude Code, Codex, Cursor or a similar agent, and you are tired of re-explaining the project in every session.
  • You want a written record of what an agent changed, which checks it ran and which commits it made.
  • You want to answer an agent's question in a few seconds, without opening the terminal.

It is not for you if you want a tool that runs a model for you. Sprintree does not. Your own agent does the work, and Sprintree is where the work is written down.

What problem does it solve?#

An AI agent starts every session with no memory. Without a tracker, you paste the task and the rules into the prompt each time, and "done" is only something the agent said in a chat.

In Sprintree the card holds the task, the project holds the rules and decisions, and the agent writes the result back to the card. The next session, or the next teammate, starts from the same place.

A day with an agent on one card#

Say you use Claude Code on a project with the key WEB.

  1. The agent reads before it works. You say, "Take the next To do card in WEB." The agent calls get_work_context for the card. One answer holds the card, its parent chain, its description, open blockers, acceptance criteria and the project rules that apply. It then moves the card to In progress. See Rules and decisions for agents.

  2. It asks instead of guessing. Halfway, it cannot choose between two payment providers. It asks a question on the card with two options and a recommendation, and keeps working on anything that does not depend on the answer. You tap one option in the Inbox or in Telegram. See Agent questions.

  3. Your decision is saved. The answer can become a decision page in the Memory Drawer, with the reason, so later sessions read it. See Memory Drawer.

  4. You steer from the browser. You read the card and notice a missing case. You comment /claude add a test for the empty cart. The running Claude Code session picks it up and replies. See Live events.

  5. It commits and the commit finds the card. The agent puts the card key in its commit message, such as fix: copy buttons (WEB-123). With a GitHub repository connected, the commit shows in the Code section of the card. See GitHub.

  6. It reports its work. The agent writes a comment with the result, calls record_work, and moves the card to In review. The record lists the checks, the commits and a short summary. See Work records.

  7. You review. You open the card, read the record and the commits, and move it to Done.

Set it up in six steps#

  1. In your terminal, install the plugin:

    claude plugin marketplace add SNebiyev/sprintree-agent && claude plugin install sprintree@sprintree

  2. Run /mcp, pick sprintree and sign in. On the approval page, tick Account settings too if you want /claude comments and triggers (step 6) to work. For Codex, Gemini CLI, Cursor, Claude.ai and ChatGPT, follow Connect an AI agent.

  3. Ask the agent to list your projects, then to read one item. Start with Read access only if you can.

  4. Write a rule or two in the Memory Drawer, such as "Each bug fix adds a failing test first".

  5. Link your repository, so commits show up. See GitHub.

  6. Optional: turn on live events and make a trigger for work that reaches In review.

Questions developers ask#

Can my AI agent update cards on its own?#

Yes, if you allow it. When you sign in with OAuth, you choose what the agent may do: Read is always on, Write lets it create and change cards, comment and move statuses. Leave Write off for a read-only agent. The agent acts as you and can do only what you can.

Does it work with Codex, Cursor and ChatGPT, or only with Claude Code?#

Any MCP client can connect. Claude Code and Codex have a Sprintree plugin, and Gemini CLI has a Sprintree extension. Cursor, Claude.ai, Claude desktop and ChatGPT use the MCP address https://sprintree.com/mcp. Live events and /claude comments need Claude Code with the plugin. Codex and Cursor have no live monitor, but they can list your waiting jobs and take one after you confirm it.

Will it follow my project rules?#

Rules marked Apply to every task reach the agent on every card. Other rules and decisions reach it when the card, or a parent of the card, links to them. The agent can also list the whole drawer with get_project_rules. If your own words disagree with a rule, your words win.

Can the agent start by itself when a card changes?#

Yes, with a trigger. A trigger is your own rule in one project: when a card moves to a status, gets a label, or you comment /claude and a name, Sprintree makes a job for your open Claude Code session. Sprintree never runs the agent itself. If no session is open, the job waits 24 hours and then expires. Changes your own agent makes through MCP do not fire your triggers, so it cannot start a loop.

Can I trust a work record?#

A record is what the agent reported, not a fact Sprintree checked. A commit is shown as a link when Sprintree saw it in a connected GitHub repository, and as "not seen on GitHub" when it did not. Open the commits when it matters.