runelay

runelayvsExecutorBenchmark · 14 Nov 2026

runelay vs Executor, measured

Same 11 MCP servers (125 tools) behind each, both self-hosted with Docker on the same machine.

2.6xfewer context tokens
10xfaster tool calls
60xfaster search
8.4xless idle RAM

Context your agent pays for at connect

No gateway, 11 servers direct44,43442x
GitHub’s MCP server alone3,5483.4x
Executor2,7292.6x
runelay1,0451x

Every measurement

runelayExecutor
  1. Context tokens at connect125 tools behind the gateway 1,045 2,729 tokens 2.6x fewer
  2. Discover a tool + its schematokens for the 2 calls 576 851 tokens 1.5x fewer
  3. Tool call, medianthrough the gateway, localhost 6.6 68 ms 10x faster
  4. Search all 125 toolsmedian, same query on both 3.8 231 ms 60x faster
  5. Idle RAMcontainer, 60 s after start 86 723 MiB 8.4x less
  6. Clean to working endpointimage already pulled 2.1 4.9 s 2.3x faster

Image size, updated. Since runelay 0.3.3 (15 Nov 2026) the image is a 66 MB download, down from 278 MB, vs 80 MB for Executor. Measured from the registry. The other rows are from the 14 Nov run with 0.3.2.

Composio. Not in this run: self-hosting Composio is an Enterprise arrangement, and its cloud is not comparable to two local containers.

Test setup

8 vCPU · 16 GB RAMDocker, localhostrunelay 0.3.2 · Executor 2.0.0-beta.5114 Nov 2026
  1. Tokens: tiktoken o200k_base (cl100k_base within 3%) over instructions and tool definitions sent at connect.
  2. Latency: median of 100 runs after 5 warm-ups.
  3. Search across all apps. Scoped to one app, Executor takes 28 ms (7x).
  4. Setup: median of 3 runs.
  5. Your hardware and tool mix will differ. Different feature sets, not a quality ranking.
Try it on your machine Download runelayInstall on your server