open source · mcp for claude code

Your AI agent is
Relay shows you exactly why.

A latency profiler that reads your agent codebase, ranks every bottleneck by the time it costs, and applies the fix — all from a single Claude Code conversation.

MIT licensed/no account, no key/runs on your machine
agent-app — claude code
$claude
$Audit this codebase for latency issues
Relay  scanning 142 files…
   parsed 142 modules · 3 bottlenecks
ranked by time cost
01tools.py:882.4s
02planner.py:311.1s
03store.py:57840ms
$Apply the fix for #01
Relay  patched tools.py · async gather
$

Profiling that lives in your editor

[ capabilities · 07 ]
01

Runs inside Claude Code

No browser tab, no upload flow. Relay is an MCP server that lives in the editor you already work in.

02

Reads your actual code

AST analysis over your real .py source — not logs, not traces, not guesses. The true source of bottlenecks.

03

Ranked by time cost

Findings sorted by the milliseconds they cost, so you fix what matters instead of chasing micro-optimizations.

04

One-command fixes

Ask Claude Code to apply a fix and Relay patches the file in place, with a diff you approve first.

05

CI-ready engine

Wire the same engine into a GitHub Action and fail the build when a change regresses agent latency.

06

Your code stays yours

Everything runs on your machine through the MCP server. No account, no API key, no backend — your source never leaves the box.

07

Open source, MIT

The whole engine is on GitHub. Read the detectors, add your own, or fork it — no gate, no licence key.

Install to fix in under a minute

[ workflow ]
01
$claude mcp add relay -- uvx --from relay-arclat relay-mcp

Run one command

Wire the Relay MCP server into Claude Code. There is no account to create and no key to configure.

02
$audit this codebase for latency

Ask for an audit

Relay parses your code and returns every bottleneck, ranked by the time it costs.

03
$apply the fix for #01

Apply the fixes

Review the ranked findings, then let Relay patch the ones you approve, right in the editor.

free & open source

Install Relay

One command in Claude Code, or install the CLI and run it anywhere. Nothing to sign up for — Relay is MIT licensed and runs entirely on your machine.

  • no account, no API key, no backend
  • python 3.11+ · works standalone as a CLI
  • MIT licensed — fork it, extend it
in claude code (mcp)
$claude mcp add relay -- uvx --from relay-arclat relay-mcp
as a cli
$pip install relay-arclat

then: relay scan . · relay verify . · relay report .

Then just ask Claude Code: “Audit this codebase for latency issues”. Every audit writes a relay-report.json you can open in the dashboard.

Questions

[ faq ]

An open-source MCP server and CLI that identifies, ranks, and fixes latency bottlenecks in AI agent code directly inside Claude Code. It is a product of the Arclat studio.

Stop guessing where the latency is.

Install Relay and get a ranked audit of your agent codebase, right inside Claude Code. Free and open source.

Install Relay