agent.reviews: tool reviews written by coding agents
agent.reviews collects tool reviews written by coding agents right after real tasks. Here is how the ratings work and how to install the skill.
TL;DR: agent.reviews is a directory where Claude Code, Codex, Cursor and Antigravity file short reviews of the tools they just used, scored on usefulness, ease and reliability out of five. Installing it takes one prompt, and the same reviews can be read free from a terminal or an MCP client.
What it is
agent.reviews collects reports that coding agents write about themselves after a task. Not human star ratings and not marketing copy. A tool like pnpm gets an entry once an agent installs it, locks a monorepo, and reports back on what worked and what got in the way.
As of 2026-11-02 the site indexes 4,395 tools and 191,848 reviews. The reviewers are Claude Code, Codex, Cursor, Antigravity, OpenCode and Gemini CLI, and reviews land under the tool they describe, in categories that range from source control and deploy to databases, AI models, observability, sandboxes, testing and browser automation.
Every tool page carries the same set of numbers: an overall score, the three sub-scores, a histogram of star counts, the share of tasks that completed, the most common problems agents hit, and which agent wrote each review.
The same company runs a separate leaderboard where one task is repeated many times and the agent picks a tool on each round. Picks measure first choice, reviews measure day-to-day use. The two answer different questions.
Why it matters
An agent picking a tool today reads the same docs a human would, then burns tokens rediscovering that a CLI needs an obscure flag or a package manager collides with an existing lockfile. Every one of those dead ends is a detail nobody wrote down.
agent.reviews captures them at the moment they happen. The score matters less than the complaint. One uv review from Claude Code on October 26 records that --project does not change the working directory, so pytest needs a path argument. Changelogs never mention that kind of thing, and a new agent otherwise pays full price to learn it.
The same installation also reads reviews. A second skill checks what other agents said about a tool’s setup and auth before you add it to a project.
How the ratings work
Every review carries three numbers from 1 to 5.
| Score | Question |
|---|---|
| Usefulness | Did it do what the task needed? |
| Ease | How much effort did setup and use take? |
| Reliability | Did it behave the way the agent expected? |
The overall rating averages those three, with missing scores left out. The average maps onto a word: Excellent from 4.3, Great from 3.8, Average from 2.8, Poor from 1.8, Bad below that.
The rubric underneath runs from 1 (Poor, major problems stopped useful progress) through 3 (Workable, useful progress with noticeable friction) to 5 (Excellent, worked clearly and consistently in this task).
Sorting carries one guard worth knowing: tools with five or more reviews come first, so a single five-star report cannot top a category. Thin tools show their rating early, before the sample size catches up.
What the numbers look like today
| Tool | Rating | Reviews | Tasks completed |
|---|---|---|---|
| Git | 4.7 Excellent | 20,125 | 99% |
| uv | 4.7 Excellent | 1,286 | 99% |
| pnpm | 4.6 Excellent | 1,455 | 99% |
| FastAPI | 4.6 Excellent | 1,838 | 98% |
| Claude API | 4.3 Excellent | 2,976 | 68% |
| Resend | 3.9 Great | 879 | 61% |
| Playwright | 3.8 Great | 237 | 65% |
| Clerk | 3.8 Great | 150 | 56% |
Counts pulled from the individual tool pages on 2026-11-02.
The completion column is worth reading next to the score. A 4.3 with 68% of tasks completed is a different proposition from a 4.7 with 99%, and the site shows both. Clerk at 3.8 with 56% completion says more about what agents run into with managed auth than the score alone does.
Getting it running
The install is one prompt aimed at your coding agent. It adds two skills, prints a sign-in link, and writes a first review.
npx -y skills add https://agent.reviews/skills -g -y -a universal -a claude-code
npx -y @armature-tech/agent-reviews login --force
The agent then opens the sign-in link for you. For a lighter setup, one review of Armature and reviews only when you ask, end with npx -y @armature-tech/agent-reviews automatic declined. Both work in any coding agent that can run a terminal command, including Claude Code, Codex, Cursor, Antigravity, Gemini CLI and OpenCode.
To read reviews without installing anything, there is a CLI:
npx -y @armature-tech/agent-reviews lookup "<tool>"
The same package exposes compare, search and category, and prints JSON. The sign-in token stays on the machine, so the agent never handles it.
There is an MCP server as well, at https://agent.reviews/mcp, with submit_agent_review for publishing and lookup_tool, compare_tools, search_tools and list_category for reading. In ChatGPT or Claude you add it as a connector and sign in when prompted.
Automatic reviews are the part that needs a decision. The skill asks once whether your agent should review the tools it uses after every task, and if you agree it writes a standing rule into your personal instructions file: ~/.claude/CLAUDE.md for Claude Code, ~/.codex/AGENTS.md for Codex, ~/.gemini/GEMINI.md for Gemini CLI and Antigravity, ~/.config/opencode/AGENTS.md for OpenCode. It never writes into a repository, since that would switch reviews on for everyone working there.
What gets shared
The public record is deliberately thin. A review names the tool, how the agent connected to it, a broad task such as “deploy a preview”, what worked, what got in the way, the result, and the three scores.
Code, prompts, logs and conversation are never part of a review. Every review passes moderation first, and one containing a key, a token, an email address or an internal address is held back. Agents also skip a tool they reviewed in the last 30 days unless the task turned up something new.
Verified means the person signed in. The agent hands over a Google or email sign-in link on its first review, and one sign-in covers every agent on that computer. Nobody is named publicly. Without a sign-in the review still publishes, marked unverified, after ten minutes.
Where the signal gets thin
Agents are generous graders. The founder said on the launch thread that they rarely hand out a 1, so the bottom of the scale stays empty and a 4.6 reads lower than it would on a human scale. Read the numbers as a band, not a measurement.
The reviews also describe the agent’s experience rather than the tool’s general quality. An agent running with tight guardrails frames a tool differently from one running loose, and the review carries that setup with it. The skill that reads reviews before a choice is explicit about this and tells agents to treat those reports as a side note to their own research, never as the reason to pick a tool.
Coverage is uneven too. Git has more than 20,000 reviews while many tools sit at a few dozen, and the early contributors are people with opinions about agent tooling.
Then there is the commercial question sitting underneath. A directory where agents choose tools is an obvious place to sell placement. The founder said on the thread that paid ranking is not where this is headed. It is still the thing to watch.
Source and accuracy notes
- Project page: https://agent.reviews/ — visited directly on 2026-11-02, including the agent-facing docs at
/install.md,/how-it-works.md,/skill.mdand/llms.txt - Review payload schema: https://agent.reviews/schema.json — exposes the
agent-review.compact.v1format - Made by Armature: https://armature.tech — the same team that ships an MCP analytics product
- HN launch thread: https://news.ycombinator.com/item?id=49995539 — posted 2026-10-07 by the founder, 73 points at time of writing
- Ratings and counts in the table were read from each tool’s own page on 2026-11-02
- Pricing: free to read and free to review, as stated on the product page
Related Posts
dev-tools
Automotive Skills Suite for AI Engineering
Evaluate Automotive Skills Suite for APQP, ASPICE, HARA, safety-plan, and DIA workflows with setup notes, governance risks, and SME review guidance.
5/28/2026
dev-tools
awesome-agentic-ai-zh Roadmap Guide
Explore awesome-agentic-ai-zh as a Chinese agentic AI learning roadmap, with setup notes, track selection, study workflow, and evaluation guidance.
5/28/2026
dev-tools
Baguette iOS Simulator Automation Guide
Set up Baguette for iOS Simulator automation, web dashboards, device farms, gesture input, streaming, and camera testing with Xcode caveats.
5/28/2026