Thanks to visit codestin.com
Credit goes to github.com

Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 

README.md

Benchmarks

Compare mcpfs (filesystem) against CLI tools and raw MCP JSON-RPC.

Run

# Mount some servers first
mkdir -p /tmp/mnt
mcpfs mount github /tmp/mnt/github -- mcpfs-github
mcpfs mount vercel /tmp/mnt/vercel -- mcpfs-vercel
mcpfs mount docker /tmp/mnt/docker -- mcpfs-docker

# Run benchmarks
./bench/bench_tokens.sh /tmp/mnt
./bench/bench_latency.sh /tmp/mnt 10
./bench/bench_compose.sh /tmp/mnt

What's measured

Benchmark Measures Key metric
bench_tokens.sh Context cost (chars/tokens) per interface Filesystem is 3-10x cheaper
bench_latency.sh Read latency: cold, warm, CLI, MCP API call dominates
bench_compose.sh Lines of code for same task across interfaces Unix pipes vs custom scripts

Expected results

Tokens: Filesystem reads return pre-formatted summaries. CLI returns full API JSON. For a typical "list repos" query, filesystem uses ~500 tokens vs ~5000 for raw API.

Latency: Dominated by upstream API calls (100-500ms). FUSE overhead adds <5ms. Cold MCP start (Go compile) adds ~2s but pre-compiled binary is <50ms.

Composability: Filesystem enables grep, awk, comm, wc across all services. Same cross-service queries require per-service CLI flags and custom scripting.