NEW v0.4.0 is live on PyPI

Persistent sprint memory
for AI coding agents.

Sprint Agent stops your AI agent from re-reading 50 files at every session. Drops a tiny .sprint/ folder in your repo, and the agent picks up exactly where it left off.

$ pip install sprint-agent
12,300
Tokens / session — without
re-read 50 files every time
2,100
Tokens / session — with
↓ 83% reduction
2 min
Context loading
vs 8–15 min · 4–7× faster
+50%
Real work / 30-min daily
28 min vs 15–22 min before
The waste

Every AI session starts the same way.

Your agent rebuilds context from scratch. Same 12k tokens. Same 8–15 minutes. Every. Single. Time.

Without Sprint Agent

● Wasteful
  • ~12,300 tokens of re-exploration per session
  • No memory of yesterday's blockers or decisions
  • Agent re-asks the same questions

With Sprint Agent

● Persistent
  • ~2,100 tokens — 83% reduction in context
  • Persistent sprint memory in .sprint/
  • Agent knows what shipped, what's next, what's blocked
In action

Two minutes from install to first sprint.

Single command creates a .sprint/ folder with backlog, current sprint, and the right config file for your agent.

~/my-saas — sprint-agent zsh
What's inside

Markdown over databases. Git-native. Zero deps.

No background server, no SQLite, no Docker. Just files your agent reads and your editor opens.

Token cost
−83%

Less context, more work

Measured on a B2B SaaS with 5M+ company records. 10 production sessions. Tokens at ~4 bytes/token.

AGENT.md
~600B
sprint S01
~1.4KB
backlog
~480B

Single source of truth

One .sprint/ folder. Markdown files. Version-controlled. Diff-able. Greppable. Editable in any editor.

.sprint/ ├── AGENT.md # auto-loaded ├── backlog.md # prioritized ├── sprints/S03_*.md └── retros/RETRO_S03.md

30-min dailies

One ticket per session. Forces shipping over polishing.

Friday retros

One command, one template. Score, document, plan next.

Zero deps

Pure Python stdlib. Single 700-line file. Pip or curl, your choice.

Compatibility

One install. Every AI coding agent.

Auto-generates the config file each agent expects on disk. Pick yours.

Auto
Claude Code
CLAUDE.md
Auto
Cursor
.cursorrules
GitHub Copilot
.github/copilot-instructions.md
Auto
OpenAI Codex
AGENTS.md
Gemini
read .sprint/AGENT.md
Aider
.aider.conf.yml
Auto
Windsurf
.windsurfrules
Any agent
point at .sprint/AGENT.md
Battle-tested

Six sprints. 70+ tickets. One developer.

Real numbers from six months building DataFrancePro — a B2B SaaS with 5M+ company records, daily 30-min sessions with Claude Code.

−83%
context tokens / session
4–7×
faster session starts
+50%
real work / 30-min daily
0
runtime dependencies
Why this exists

Built out of frustration. Open-sourced because
it kept solving the same problem.

I was burning 12 minutes every Claude Code session re-explaining the same project before any actual work happened. CLAUDE.md alone wasn't enough — I needed structure: backlog, current sprint, what's blocked, what shipped yesterday. So I built one tiny CLI that drops a .sprint/ folder in the repo. After 6 months and 70+ tickets shipping DataFrancePro with this workflow, it works. Sharing the source.
M
Miloud Belarebia
Builder · @miloudbelarebia
01

Markdown beats databases

Files diff in git, open in any editor, survive every tool change. Zero migrations, zero servers.

02

Designed for solo + AI

Not a team tool. Opinionated 30-min daily, single ticket per session. Trades flexibility for shipping velocity.

03

Read every line

Pure Python stdlib. ~700 lines. Single file. Nothing weird running in the background.

FAQ

Common questions.

How is this different from just writing a CLAUDE.md?

CLAUDE.md tells the agent what the project is. Sprint Agent adds where you are in the work — current sprint, today's ticket, what's blocked, what shipped yesterday. The agent picks up mid-sprint instead of starting fresh.

Why markdown instead of a database?

Markdown is readable by humans and AI agents without parsers. It diffs cleanly in git, opens in any editor, survives every tool change. Zero dependencies, zero migrations, zero servers.

Does it work offline?

Yes. Sprint Agent never makes a network call. It just reads and writes local files in .sprint/. Your AI agent reads them on its next session.

What about teams of 2+ developers?

Sprint Agent is designed for solo dev + AI pairs. Teams should keep using Linear, Jira, GitHub Projects. The opinionated 30-min daily and single-ticket-per-session model breaks at team scale.

Can I use it with my own custom agent?

Yes. Any agent that reads project files works — just point it at .sprint/AGENT.md at session start. Works out of the box for Claude Code, Cursor, Copilot, Codex, Gemini, Aider, Windsurf.

How do I update or contribute?

Bug reports and feature requests via GitHub Issues. Discussions in Discussions. Pull requests welcome — read CONTRIBUTING.md first.

Get started

Ship like you have an agile team. Solo.

Install in two minutes. Free, open-source, MIT-licensed.

$ pip install sprint-agent && sprint-agent init
$ curl -O https://raw.githubusercontent.com/miloudbelarebia/sprint-agent/main/sprint_agent.py