How to use this directory
What the parts of a profile mean, how to find the right agent, and what to do when something looks wrong.
Finding an agent
All Agents is the full catalogue, filterable by category and by publisher. If you would rather start from a job than a category, the profession filters group agents by the work they support. Comparisons puts closely related agents side by side, Glossary explains the vocabulary, and Insights summarises what the catalogue as a whole looks like right now.
Every page exists in English and Chinese; the language toggle keeps you on the same agent rather than sending you back to the homepage.
Reading a profile
A profile leads with what the agent does and who published it, then the scored review, then the practical detail: use cases, setup, day-to-day usage, comparisons with similar agents, strengths and limitations, and a short FAQ.
The badges near the title are factual, not editorial. “Official” means the source repository belongs to the named organisation, established from public provenance. The star count and licence come from the source repository. “Updated” is the repository's last commit, not the date the review was written.
Making sense of the score
The FARS score out of 100 combines weighted dimensions — Trust & safety carries the most weight at 29, then Adaptability and Convention at 18 each, Reliability at 14 and Effectiveness at 13. Each dimension shows its own reasoning on the page, so you can disagree with a specific judgement rather than the whole number.
Alongside the score, the profile states its evidence level: high means key paths were checked and reproduced with independent corroboration; low means the review rests mainly on static reading of the author's own material. A low evidence level is a signal to test it yourself, not necessarily a bad agent.
Some agents are marked Not recommended regardless of their arithmetic total, because a red-line issue could not be reasonably resolved. The methodology page lists what those red lines are.
Before you run an agent
An agent takes actions: it writes files, calls APIs, spends money on tokens, and sometimes holds long-lived credentials. Whatever the score says, read the source repository, check what access it asks for, start it against a test account and a scratch directory, and keep a way to stop and undo it.
Never paste real credentials into an example configuration you have not read, here or anywhere else.
An agent is missing, or a profile is wrong
Candidates are discovered from GitHub by rule, so a very new or very obscure repository may not have been picked up yet, and a profile may lag behind a repository that changed last week.
To suggest an agent, report an error, or ask for your own repository to be corrected or delisted, write to [email protected] with the URL. Maintainer requests about their own repository are honoured.
Anything else
The methodology page documents the full scoring standard. The privacy page explains exactly what this site collects. For anything not covered, [email protected] reaches a person.