Agent Performance Review
Experimental · Local MCP

Your agents are busy.
Are they moving
the needle?

Agent Performance Review is an experimental local MCP server for rating agent work on your own computer. An assistant requests a review link, you open the form, and you decide whether to submit a rating or skip.

View on GitHub ↗ Read documentation
Local prototype. No accounts, uploads, or community features yet.
Illustration showing a task flowing through agent actions, context, resources, and results into a review.
TaskWhat needed to be done
ActionWhat the agent did
ContextRuntime, model & tools, if supplied
ResourcesTime, tokens & cost, if supplied
ResultWhat was delivered
ReviewCompletion & star ratings
A review holds only what the assistant supplies and what you rate. Keep summaries short and non-sensitive.
01
Rate the result, not the hype Did the job get done? Then rate quality, speed, and resource efficiency.
02
Unknown stays unknown Time, tokens, and cost appear only when supplied. Nothing is estimated.
03
Saved on your computer Submitted ratings go to a local file beside the server. No accounts or uploads.
How it works today

Ask for a link.
Rate the work.

The current flow is manual and experimental. An assistant connected to the local MCP server can request a review link for a task. You stay in control of whether a rating is saved.

Try a manual review →
01

Connect the local server

Add the stdio MCP server from the public repository to an MCP host on the same machine.

02

The assistant requests a link

Through MCP, the assistant supplies a short summary and gets back a local review link.

03

Click to open the form

The link opens a review form served from your computer. Opening it saves nothing.

04

Submit or skip

Only Submit saves a local rating. Skip closes the link and saves no rating.

Current status

What works, and what
is not verified yet.

Works today · experimental local MCP access

  • An assistant can request a local review link through MCP.
  • Clicking the link opens a review form on your computer.
  • Submit saves a rating to a local file. Skip saves no rating.
  • A local summary reports review counts and average ratings.

Not verified or not built

  • Automatic per-turn capture of agent activity.
  • Automatic insertion of a review link at the end of each turn.
  • Runtime-specific adapters, hooks, packages, or installers.
  • Accounts, sharing, and community comparisons.

Measured time, tokens, and cost are unavailable unless the assistant supplies them. Ratings are personal local judgments, not benchmarks.

See the current support matrix →
Local MCP setup

Set up local
MCP access.

Setup steps live in the documentation and the public README, so there is one source to keep current.

✓
What you need Git, Python 3, a browser, and an MCP host that can launch a local stdio server. macOS or Linux.
✓
What gets sent Review summaries pass through your MCP host. If it uses a cloud model, its provider may see them.
Setup guide EXPERIMENTAL
No one-click installer yet. Configure the server in your MCP host by hand. A successful tool call shows only that the manual MCP path works.
View on GitHub ↗ Local setup
  1. Local setupPrerequisites, clone, and MCP host configuration.
  2. Try a manual reviewRequest a link, then test Submit and Skip with synthetic data.
  3. Verify & troubleshootCommon link, connection, and saving problems.
  4. Current supportWhat is implemented and what remains unverified.
  5. Public README: Local MCP setup ↗Source code and setup notes on GitHub.