I built a VS Code extension to solve my biggest frustration with AI coding tools

Picture this.

You are deep into a complex bug. You have spent 20 minutes explaining your codebase to Claude Code — the architecture, the tech stack, what you have already tried. You are making real progress. The AI finally understands exactly what you need.

Then credits run out. Mid-sentence.

So you switch to Gemini. And you have to explain everything again.

Gemini hits its rate limit. You open Codex. And you have to explain everything again.

This is not a rare edge case. This is the daily reality of working with multiple AI coding tools in 2026. Every tool has credit limits. Every tool has rate limits. None of them share context with each other.

I got frustrated enough to build a solution.

What I built

Session Bridge AI is a VS Code extension that maintains a running SESSION.md in your project — continuously updated, always ready to hand off to any AI tool so you can continue exactly where you left off.

The core insight

The obvious solution — syncing AI tools directly — is impossible. They all have different APIs, different context formats, different limitations.

The simpler insight: maintain a single source of truth that lives in your project. A plain markdown file that any AI tool can read and immediately understand.

That file is SESSION.md.

Every time you save context, Session Bridge AI:

Reads your git diff — what code actually changed
Reads your currently open files — what you are looking at
Combines it with your manually logged notes
Sends everything to your chosen AI provider
Generates a structured SESSION.md

When your current AI tool dies, you open SESSION.md, paste it as the first message to the next tool, and continue exactly where you left off.

What SESSION.md looks like

Here is a real example generated by the extension:

## Problem
Implement JWT-based authentication for a REST API

## Approach
Express.js + PostgreSQL, bcrypt for password hashing, JWT for stateless auth

## Completed
- POST /register — complete
- POST /login — complete, returns access + refresh tokens
- DB schema — finalized

## In Progress
- GET /me endpoint — needs auth middleware

## Next Steps
- Complete auth middleware
- Add token refresh endpoint
- Write integration tests

## Key Decisions
- Chose JWT over sessions for statelessness
- Refresh tokens stored in DB for revocation support

## Files Modified
- src/routes/auth.ts — login and register routes
- src/middleware/authenticate.ts — JWT verification
- src/db/schema.sql — users and tokens tables

## How To Continue
Complete the GET /me route in src/routes/user.ts using the
authenticate middleware. Then add POST /auth/refresh for token renewal.

Any AI tool can read this and immediately understand exactly what is happening in your project.

Features

Multi-provider support — Gemini, Claude, and OpenAI
Automatic git diff capture — sees what code actually changed
Open files capture — knows what you are currently working on
Auto-save — saves context every N messages automatically
Manual save — click the status bar button anytime
Secure key storage — API keys stored in VS Code secret storage, never in plaintext
Persistent buffer — survives VS Code restarts
Configurable threshold — set auto-save frequency to your preference

The workflow

Start working with Claude Code / Codex / Gemini / Amazon Q
Log progress every few exchanges:
Ctrl+Shift+P → “Session Bridge: Log Message” Credits die unexpectedly? → Open SESSION.md → Paste it as the first message to your next AI tool → Continue exactly where you left off

Getting started in 3 steps

Step 1 — Install

Search Session Bridge AI in VS Code Extensions, or install directly from the Marketplace.

Step 2 — Get a free Gemini API key

Go to Google AI Studio and create a free API key. No credit card required.

Step 3 — Set your provider

Ctrl+Shift+P → Session Bridge: Set AI Provider & API Key

Select Gemini, paste your key, done.

Why I built this

I was working on a personal project using a combination of Claude Code, Gemini, and Amazon Q. Every time I switched tools, I lost all the context I had built up. I was spending more time re-explaining my codebase than actually writing code.

I looked for existing solutions. There were tools for sharing context within a single tool — CLAUDE.md, .cursorrules — but nothing that worked across tools as a live session handoff.

So I built it myself.

The entire extension — from problem statement to published Marketplace listing — was built in a single focused session. The irony of building a tool to survive AI credit exhaustion while using AI tools that kept running out of credits is not lost on me.

What’s next

Keyboard shortcuts for faster logging
Auto-detect which AI tool is active
Team sharing via git for collaborative sessions
Automatic chat interception — no manual logging needed

If you have ideas or run into bugs, open an issue on GitHub. PRs are welcome.

Find more of my work at rj-gamer.github.io/rajatjog-website