This is a toolkit which mounts an S3 bucket as a read-only virtual filesystem. It provides filesystem-like tools for reading and searching files in AWS S3 buckets. Ideal for use in coding agents like Amp, Claude Code and Codex
s3.mp4
Perfect for navigating large blob storage and performing agentic search over compressed files.
1. Generate a manifest (indexes your bucket for fast searches):
bun run util/generate-s3-manifest.ts <bucket> [prefix] [region]2. Start the MCP server:
S3_BUCKET=your-bucket S3_REGION=us-east-1 bun run src/server.ts3. Configure Amp (add to .amp/settings.json):
{
"amp.mcpServers": {
"s3_filesystem": {
"url": "http://localhost:7011/sse"
}
}
}- read - Read files or list directories
- glob - Find files by pattern (
**/*.json) - grep - Search file contents (requires cache)
- filter_metadata - Filter by size, date, storage class
- Fast - Manifest-based directory listings (no S3 API calls)
- Smart - Auto-decompresses
.gzand.zstfiles - Cached - LRU cache for frequently accessed files
Uses standard AWS credential resolution:
AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY~/.aws/credentials- IAM roles
GPL-3.0