sprinty

A disciplined-sprint MCP server for AI coding agents — Claude Code, Codex, and Gemini.

Sprinty gives an agent first-class tools to run a sprint with structure that can't silently rot: structured sprint → subsprint → item objects, an immutable append-only ledger anchored to real git commits, programmatic close-gates that re-run your tests before a sprint can close, a regex search over the record, and a live follow-along dashboard.

The point: the agent doesn't drift, and the record doesn't lie. IDs are minted server-side, items can't exist without gates, done rejects a commit that doesn't exist, and sprint_close refuses to close while anything is unresolved or a gate fails.

Install

Sprinty ships as a native plugin for each agent. The MCP server itself runs via npx -y sprinty-mcp (the npm package is sprinty-mcp; the server, tools, and plugins are all named sprinty).

Claude Code — clients/claude/ is a plugin (.claude-plugin/plugin.json) bundling the MCP server and skills. Add the MCP directly:

claude mcp add sprinty -- npx -y sprinty-mcp

Codex — clients/codex/ is a plugin (.codex-plugin/plugin.json + .mcp.json + skills), installed through a marketplace (clients/codex/marketplace.json). Or add the server to ~/.codex/config.toml:

[mcp_servers.sprinty]
command = "npx"
args = ["-y", "sprinty-mcp"]

Gemini CLI — clients/gemini/ is an extension (gemini-extension.json + GEMINI.md + skills):

gemini extensions install ./clients/gemini

The skills are authored once in skills/ and symlinked into each client, so all three share one body of guidance.

Related MCP server: anchor-mcp

The loop

sprint_new(goal)
  -> subsprint_new(description, goals[], gates[])
  -> add(subsprint, description, code_locations[], gates[])
  -> done(commit_id, gate_results[]) | split(...) | deprecate(reason)
  -> sprint_close()

Full tool reference: skills/using-sprinty/SKILL.md. How to run a sprint: skills/how-to-run-a-sprint/SKILL.md.

Storage

One append-only JSONL ledger file per sprint under .sprinty/ in the repo you're working on, with a .sprinty/current pointer naming the active sprint (this enforces one-open-sprint unicity). .sprinty/ is per working tree, so git worktrees run independent sprints. It is local state — keep it gitignored.

Develop

npm install
npm test            # builds, then runs unit + e2e tests
npm run test:coverage
npm run build

License

Apache-2.0 © Elie Bursztein