FlowBack

Your AI co-pilot for picking up where you left off.

Comment

INSPIRATION

Developers lose 20–45 minutes every time they switch back to a project after a break — trying to remember what they were doing, what broke, and what to do next.

Multiply that by 3–5 context switches a day, and you're burning hours just regaining momentum.

We built flowBack to eliminate that cognitive overhead entirely.

WHAT IT DOES

Run one command before you step away.

flowBack scans your recently modified files, sends them to an AI, and saves a structured briefing:

  • Your goal
  • Where you got stuck
  • Your next 3 steps

When you return, run: flowback resume

…and you're immediately back in the zone.

It also tracks errors during development:

  • Analyzes root cause and gives you 3 fix steps
  • Detects recurring errors (loop detection) and tells you how to break the cycle
  • Visualizes your error patterns and skill gaps as an interactive force-directed graph

✅ No setup
✅ No server
✅ Works from any terminal

HOW WE BUILT IT

  • litellm — multi-provider LLM abstraction (Gemini, Claude, GPT-4o, Groq, Ollama)
  • Gemini 2.0 Flash — recommended default (free tier, fast)
  • Python + Typer — CLI framework
  • Rich — terminal output formatting
  • SQLite — local history storage (~/.flowback/history.db)
  • FastMCP (Anthropic SDK) — Claude Code MCP integration
  • FastAPI + React + Vite + Tailwind — optional web UI
  • force-graph.js — self-contained interactive skill graph (no server needed)

WHAT MAKES IT DIFFERENT

Most developer tools are about writing code faster.
flowBack is about resuming work faster.

Every briefing is structured for re-entry, not just documentation:

  • What was my goal?
  • Where did I get stuck?
  • What are my next 3 steps?

Native Claude Code Integration

flowBack integrates as an MCP server.

You can ask:

“What was I working on?”

…and get your last briefing instantly.

Fully Flexible AI Layer

  • Works with any LLM
  • Supports fully local models via Ollama
  • Zero data leaves your machine (if you choose)

CHALLENGES WE FACED

  • Ensuring consistent structured JSON output across multiple LLM providers
  • Designing error fingerprinting to detect recurring issues without false positives
  • Keeping workflow frictionless:
    • Too broad → noisy context
    • Too narrow → missing context
  • Building MCP server for natural language access inside Claude Code

ACCOMPLISHMENTS

  • Full context briefing generated in under 30 seconds
  • Works across 50+ LLM providers via litellm
  • Loop detection for recurring developer errors
  • Interactive skill gap graph from real usage data (self-contained HTML)
  • Claude Code MCP integration:
    • pause
    • resume
    • track_error
    • skill_gaps

WHAT WE LEARNED

Developers don’t want another note-taking tool.
They want to not lose momentum.

The positioning shift changed everything:

❌ “Developer journal”
✅ “Resume your code in 30 seconds”

The structure matters as much as the AI:

Goal → Stuck Point → Next Steps

This is not a summary.
This is a re-entry system.

WHAT'S NEXT

  • Team mode
    Shared briefings so anyone can pick up a PR

  • Git-aware context
    Detect changes since last commit automatically

  • IDE plugins
    VS Code, Cursor — resume inline without leaving editor

  • Weekly skill digest
    Email/Slack summary of recurring error patterns

Built With

Share this project:

Updates