Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion .claude-plugin/marketplace.json
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
{
"name": "perfloop",
"description": "Perfloop for Claude Code and Codex: always-on performance engineering for teams where performance is the product.",
"description": "Perfloop for Claude Code and Codex: performance improvements, bug fixes and security repairs in your code, each with proof.",
"owner": {
"name": "Perfloop",
"email": "hello@perfloop.ai"
Expand Down
4 changes: 2 additions & 2 deletions .claude-plugin/plugin.json
Original file line number Diff line number Diff line change
@@ -1,8 +1,8 @@
{
"name": "perfloop",
"displayName": "Perfloop",
"version": "1.0.0",
"description": "Perfloop does always-on performance engineering on your code. A performance change comes back only with a measurement against the base. Your agent can submit a branch and steer the work.",
"version": "1.1.0",
"description": "Perfloop finds performance problems, bugs and security flaws in your code. Each change comes with proof: a measurement or a tested assertion. Your coding agent can steer it.",
"author": {
"name": "Perfloop",
"email": "hello@perfloop.ai",
Expand Down
49 changes: 36 additions & 13 deletions README.md
Original file line number Diff line number Diff line change
@@ -1,18 +1,31 @@
# Perfloop

The Performance Machine. A stream of proven pull requests, and software that
stays fast.

Perfloop does always-on performance engineering on your system. It runs
performance experiments and only brings back a performance change with a
measurement against the base. Your engineers set the goals. This repository
connects your coding agent to Perfloop's MCP server at
`https://app.perfloop.ai/mcp`.

With it, your agent can submit a branch or a pull request and read the ranked
hypotheses Perfloop finds on the change. It can read a Case's evidence and
start or stop work on it. It can also steer the work with feedback. Perfloop opens pull requests.
It cannot approve or merge one, and it cannot deploy.
Always-on engineering. Every change proven.

Perfloop finds performance problems, bugs and security flaws in your code.
Perfloop measures each performance change it proposes against the original code.
A bug fix or security repair must fail its test on the original code and pass
after the change. Your plan and each repo's settings in Setup decide which kinds
of work Perfloop does. Setup is a page in the Perfloop app. Your team chooses
the repos Perfloop works on and what it should improve. This repository connects
your coding agent to Perfloop's MCP server at `https://app.perfloop.ai/mcp`.

With the MCP server connected, your agent can submit a branch or a pull request
and read the ranked hypotheses Perfloop finds on the change: possible problems,
not yet proven. A Case is Perfloop's record of one finding and the work on it.
Your agent can read a Case and its evidence, and start or stop that work. It can
also send Perfloop feedback on the work. Perfloop opens pull requests. It cannot
approve or merge one, and it cannot deploy.

Security work is defensive: Perfloop finds security flaws in the repos you
connect and proves fixes for them. It is available on the [Pro and Enterprise
plans](https://perfloop.ai/pricing), once the Perfloop team turns it on for your
workspace. Only members of your workspace can see a security Case, even for a
public repo. When Perfloop reviews a pull request on a public repo, the review
leaves out security findings. Perfloop does not open a pull request or issue for
a security fix on a public repo. Download the fix from the Case; if someone else
maintains the repo, send it to them privately. Read the [Security
guide](https://perfloop.ai/docs/work/security) before sharing a finding.

## First run

Expand Down Expand Up @@ -55,6 +68,16 @@ then start the server and sign in when VS Code asks.
Every other agent: the steps are at
[perfloop.ai/docs/work/mcp](https://perfloop.ai/docs/work/mcp).

## Skill

`skills/perfloop/SKILL.md` tells your agent how to connect to Perfloop and
use it. Claude Code loads it with the plugin. For other agents, install only
the skill from perfloop.ai:

```sh
npx skills add https://perfloop.ai
```

## What this plugin sends

The plugin holds no code and no credential. It only tells your agent where
Expand Down
4 changes: 3 additions & 1 deletion llms-install.md
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,9 @@ their Perfloop account through OAuth.

The user needs a Perfloop account in a workspace with a connected GitHub
repo. If they have none, they get an account at https://perfloop.ai, and an
Admin connects GitHub and selects repos in Setup in the Perfloop app.
Admin connects GitHub and selects repos in Setup in the Perfloop app. After
sign-in, an Admin's agent can also select repos with the `setup` and
`setRepositories` tools.

## Add the server

Expand Down
14 changes: 7 additions & 7 deletions plugin.json
Original file line number Diff line number Diff line change
@@ -1,8 +1,8 @@
{
"$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
"name": "perfloop",
"version": "1.0.0",
"description": "Perfloop does always-on performance engineering on your code. A performance change comes back only with a measurement against the base. Your agent can submit a branch and steer the work.",
"version": "1.1.0",
"description": "Perfloop finds performance problems, bugs and security flaws in your code. Each change comes with proof: a measurement or a tested assertion. Your coding agent can steer it.",
"author": {
"name": "Perfloop",
"email": "hello@perfloop.ai",
Expand All @@ -22,8 +22,8 @@
"com.openai": {
"interface": {
"displayName": "Perfloop",
"shortDescription": "The Performance Machine.",
"longDescription": "Perfloop does always-on performance engineering on your system. It runs performance experiments on your code and only brings back a performance change with a measurement against the base. Your engineers set the goals.\n\nIt also finds correctness bugs. Each correctness claim proves one stated assertion about the code with the change. When a change repairs a defect, the proof also shows the assertion failing before the change, for example a test that fails first and passes with it.\n\nWith Perfloop connected, your agent can submit a branch or a pull request, and Perfloop lists ranked hypotheses on the changed code, each with a size estimate. Read a Case's evidence and measurements. Start or stop work on it, and steer it with feedback. Decide what waits in the Inbox, such as a pull request that is ready to open.\n\nPerfloop opens pull requests. It cannot approve or merge one, and it cannot deploy. Every write is recorded under your account.\n\nYou need a Perfloop account with a connected GitHub repo. You sign in with that account, and your role in the workspace decides what the agent can read or change. Tools that start work say so, because that work is paid from the workspace's Usage balance.",
"shortDescription": "Always-on, proven engineering.",
"longDescription": "Perfloop finds performance problems, bugs and security flaws in your code. Perfloop measures each performance change it proposes against the original code. Your team chooses the repos and goals Perfloop works on.\n\nEach correctness claim proves one stated assertion about the code with the change. A bug fix or security repair must fail its test on the original code and pass after the change. Your plan and each repo's settings in Setup decide which kinds of work Perfloop does. Security work is defensive: Perfloop finds security flaws in the repos you connect and proves fixes for them. It is available on Pro and Enterprise, once the Perfloop team turns it on for your workspace. Only members of your workspace can see a security Case, even for a public repo. When Perfloop reviews a pull request on a public repo, the review leaves out security findings. Perfloop does not open a pull request or issue for a security fix on a public repo. Download the fix from the Case; if someone else maintains the repo, send it to them privately. Read https://perfloop.ai/docs/work/security before sharing a finding.\n\nWith Perfloop connected, your agent can submit a branch or a pull request, and Perfloop lists ranked hypotheses on the changed code, each with a size estimate. Read a Case's evidence and measurements. Start or stop work on it, and steer it with feedback. Decide what waits in the Inbox, such as a pull request that is ready to open.\n\nPerfloop opens pull requests. It cannot approve or merge one, and it cannot deploy. Every write is recorded under your account.\n\nYou need a Perfloop account with a connected GitHub repo. You sign in with that account, and your role in the workspace decides what the agent can read or change. Tools that start work say so, because that work is paid from the workspace's Usage balance.",
"developerName": "Perfloop",
"category": "Developer Tools",
"capabilities": [
Expand All @@ -32,7 +32,7 @@
"File feedback that steers the work",
"Read the model Perfloop keeps of your system",
"Create an Initiative and steer its research",
"Submit a branch or pull request and read ranked performance and correctness hypotheses on it",
"Submit a branch or pull request and read ranked hypotheses about performance, bugs and security",
"Decide what waits in the Inbox"
],
"websiteURL": "https://perfloop.ai",
Expand All @@ -41,7 +41,7 @@
"termsOfServiceURL": "https://perfloop.ai/terms",
"defaultPrompt": [
"Which Perfloop Cases on my repo are verified, and what did each one measure?",
"What performance hypotheses did Perfloop find on my open pull request?",
"What performance problems, bugs and security flaws did Perfloop find on my open pull request?",
"What is waiting for my decision in the Perfloop Inbox?"
],
"brandColor": "#3B82AF",
Expand All @@ -52,7 +52,7 @@
"logoDark": "./assets/logo-dark.png"
},
"publication": {
"release_notes": "Initial release: the Perfloop MCP server, with tools to submit a branch or pull request, read Cases and their evidence, steer and start work, and decide Inbox items."
"release_notes": "Adds the Perfloop agent skill and describes performance, bug and security work, each with proof."
}
}
}
Expand Down
58 changes: 58 additions & 0 deletions skills/perfloop/SKILL.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,58 @@
---
name: perfloop
description: Work with Perfloop, which finds performance problems, bugs and security flaws in the user's code and proves changes. Use when the user asks about Perfloop Cases or the Perfloop Inbox, or wants a branch or pull request checked for these problems.
---

# Perfloop

Perfloop finds performance problems, bugs and security flaws in the user's code.
Perfloop measures each performance change it proposes against the original code.
A bug fix or security repair must fail its test on the original code and pass
after the change. The plan and each repo's settings in Setup, a page in the
Perfloop app, decide which kinds of work Perfloop does. A Case is Perfloop's
record of one finding and the work on it.

Security work is defensive: Perfloop finds security flaws in the repos the user
connects and proves fixes for them. It is available on Pro and Enterprise, once
the Perfloop team turns it on for the user's workspace. Only members of the
user's workspace can see a security Case, even for a public repo. When Perfloop
reviews a pull request on a public repo, the review leaves out security
findings. Perfloop does not open a pull request or issue for a security fix on a
public repo. Tell the user they can download the fix from the Case and, if
someone else maintains the repo, send it to them privately. Before you or the
user share a finding, read https://perfloop.ai/docs/work/security.

Perfloop opens pull requests. It cannot approve, merge or deploy.

You use Perfloop through its hosted MCP server.

## Connect

1. The user: get a Perfloop account. Sign up at https://perfloop.ai, or accept
an invitation to a workspace.
2. An Admin, in the Perfloop app (not you): connect a code host, such as GitHub,
and select repos in Setup.
3. You: add the remote MCP server `https://app.perfloop.ai/mcp` with the
Streamable HTTP transport. Do not add an API key or an `Authorization`
header.
4. Your MCP client opens the sign-in page itself, using the server's OAuth
metadata. The user signs in with their Perfloop account in the browser.

When the user who signed in is an Admin, you can also choose which repos
Perfloop works on, with the `setup` and `setRepositories` tools. This works for
GitHub and for Cursor Origin. Cursor Origin is another code host, which the
Perfloop team turns on per workspace.

The steps for each coding agent are at https://perfloop.ai/docs/work/mcp.

## Use

- The server's instructions and its guide resource give each job as tool
calls, in order. Follow the guide for the job the user asked for, and do
only that job.
- Tools that start work say so in their descriptions, because that work is
paid from the workspace's Usage balance. Ask the user before you call one.
- Report results as the tools return them, with the Case link. Do not invent
numbers.

The full agent guide is at https://perfloop.ai/docs/work/mcp.
Loading