# Sprinty

> Agentic sprints done right. Sprinty is an open-source MCP server and pair of agent skills that help coding agents plan work, coordinate ownership, and finish with verified results. It records the plan, dependencies, item claims, progress, gate evidence, artifacts, and changelog in a local append-only ledger.

## Get started
- [Project examples](https://github.com/ebursztein/sprinty/tree/main/docs/examples): website, research, and parallel-agent starting prompts.
- [Captioned demo](https://ebursztein.github.io/sprinty/demo.html): 45-second visual walkthrough using illustrative project data.
- [Copyable setup prompt](https://ebursztein.github.io/sprinty/agent-start.txt): instructions a user can paste into a coding agent to install skills and configure the MCP server.
- [Source and install guide](https://github.com/ebursztein/sprinty): supported client integrations and tool reference.
- [Sprint workflow skill](https://github.com/ebursztein/sprinty/blob/main/skills/how-to-run-a-sprint/SKILL.md): item-driven planning, claims, evidence, and closure.
- [Tool contract skill](https://github.com/ebursztein/sprinty/blob/main/skills/using-sprinty/SKILL.md): exact inputs, bounded responses, ownership, timing, and recovery.
- [npm server package](https://www.npmjs.com/package/sprinty-mcp): stdio MCP server.

## Install
One-command setup for Codex or Claude Code (run in the project): curl -fsSL https://ebursztein.github.io/sprinty/install.sh | sh
Requires Node.js 20+, Git, npx and the client CLI. Choose a client if both are installed. Preserves existing MCP entries; restart the client afterward. Claude shared-subagent hooks require the plugin separately.
Skills only: npx skills add ebursztein/sprinty
Server: npx -y sprinty-mcp@latest
These are separate installations. Skills supply guidance; the MCP server supplies tools. The skills CLI does not install Claude plugin hooks.

## Start or resume
Call info({ git_dir }) first. Pass the canonical repository explicitly. Use a gitignored data_dir, normally <repository>/.sprinty. Resume existing work with sprint_resume({ git_dir, data_dir }). Start a new sprint only when the user has work to track and no suitable sprint exists. Follow next_tools and bounded list/get tools.

## Work
Create a feature subsprint and dedicated items before coding. Claim an item before editing or completing it. Record item_progress for meaningful updates, and agent_heartbeat during active work between calls. Complete items with a real resolving commit and gate evidence. sprint_close rechecks gates and writes the canonical changelog.

## Collaboration and recovery
Independent local MCP processes coordinate through one shared canonical repository/data directory. Sessions have distinct named identities and exclusive claims. Use agent_list() and items_claimed({ mine:true }) to inspect ownership. Normal connections remember identity automatically; the Claude plugin's hook supplies native subagent IDs for shared connections. Other shared clients can use explicit agent tokens. Native identity recovery requires a stable caller ID; do not assume a new process owns old claims. Worktrees keep separate bindings. Potential stalls are warnings; explicit takeover requires an explanation and eligible stalled work.

Compaction does not erase the disk ledger. It is the work record, not a backup of the entire conversation. After context loss, call info and sprint_resume to recover the plan, current work, and evidence. Keep secrets and local sprint data out of git.

## ETA
The dashboard and tools share an activity-aware estimate. It separates queue wait, claim duration, and estimated active effort; sleep contributes at most five minutes per observed activity event. It learns from at least three eligible completions. Paused or inactive-owner states suppress calendar finish forecasts. Ranges are effort estimates, not statistical confidence intervals.