Methodology

How we discover, evaluate, install-test, and rank every agent wallet, programmable wallet, or payment rail on agentwallet.md.

Discovery sources

GitHub trending — automated weekly scan of repositories tagged with relevant topics.
Google News — RSS scrape for product launches matching curated query lists.
Community submissions — issues and PRs on our GitHub repository.
Manual additions — when we encounter a notable agent wallet, programmable wallet, or payment rail in conversation or research.

Hard filters (mechanical)

Minimum 30 GitHub stars (cuts out trivial / abandoned hobby projects)
Repository not archived
Not a pure fork without original commits
Canonical URL resolves (HTTP 200/301/302/403)
Not already in our directory (deduplicated by repo and slug)

Editorial review (LLM)

Each candidate that passes hard filters is evaluated by Kimi K2 with prompt constraints:

Is this clearly an AI/agent-relevant tool? (rejects: low-level XML libs, hardware wallets, generic doc-format parsers, abandoned spam)
What category does it best fit?
Is the project legitimate (not malware, not a clone-rebrand)?
Is there enough README detail to write an honest entry?

The LLM is instructed to err on the side of rejection when in doubt. We'd rather miss a borderline tool than add an off-topic one.

Entry generation

Approved candidates are turned into a JSON entry + detail HTML page using only verified facts from the GitHub API and README. We have explicit rules against:

Inventing ratings, percentages, or numeric scores
Copying schema field defaults verbatim
Marketing fluff in descriptions

What's tracked per entry

Field	Source
Star count	GitHub API at time of listing
License	GitHub API `license.spdx_id`
Last commit	GitHub API `pushed_at`
Install command	README + manual verification
Best-for / category	Editorial assessment from README

Velocity cap

The pipeline publishes at most 5 new entries per site per week. This keeps the directory growing steadily without becoming a spam dump.

Updating and corrections

Each detail page shows a Last reviewed date. We re-check entries quarterly for: still active? still recommended? license changed? Corrections can be submitted via GitHub issue or the home-page form. Most are merged within 7 days.

Reproducibility

The pipeline source is open. See github.com/fleurytian/md-sniper — discovery, evaluation, generation, and publishing scripts are all there.