routerflow.artthe marketour probecomparemethodologyprobestaxonomyprober methodapi/v12026-11-10

Router registry: what is built, and the plan for scanning

Status: the registry and the pull-request path are live (2026-09-20). Scanning is planned, not built. Focus for now: three routers collected fully (OpenRouter, Vercel AI Gateway, TrustedRouter); the rest are listed only.

Built

PieceWhere
The list: one JSON file per router (who it is, its API, where its public data lives, whether routerflow collects it). A registry only: no figures, statements or estimatesgithub.com/dmarzzz/routerflow-registry, routers/
Entry template, field definitions, contribution rulesTEMPLATE.json, README.md in that repo
Validation on every pull requestscripts/validate.py, .github/workflows/validate.yml
"Add your router" buttonOpens GitHub's new-file editor on the registry with the template filled in; GitHub turns the save into a pull request from a fork
The list on the siteRouters section, "Every router we know of". make registry copies the checked-out registry into site/registry.js and site/api/v1/registry.json

A merged pull request reaches the site by itself: the nightly collector reads the registry from GitHub (one API call, no key), the build carries it in the dataset and publishes api/v1/registry.json, and the hourly publish job pushes it. make registry remains as a manual fallback.

Planned: scan, report back, re-check

1. Read the registry nightly. Built 2026-09-20: see above.

2. Scan what an entry points at. Two scanners, chosen per entry:

ScannerRuns whenDoes
Export scannerpublishes[].export is json or csvFetches the URL, stores the raw file, records shape, row count and date range. A named adapter (as for OpenRouter and Vercel today) turns it into shares or counts. No adapter yet means the raw file is kept and the entry is flagged "export found, adapter needed".
Probe scannerapi.openai_compatible is true, api.base_url is set, and we hold a key for that routerSends the prober's fixed requests through that router: the cache pair, the unpinned selection sample, the one-shot. Same records and graders as for OpenRouter.

The probe scanner costs money and needs a key per router. It runs only for routers a maintainer has enabled, inside the prober's existing budget policy, with its own line in expected_usd_per_day.

3. Report back as a pull request. A bot account opens a pull request on the registry with results/<id>.json: what was fetched, when, what the scan found, and for a probe scan the aggregate results (never response text). The operator can see it, dispute it and correct the entry in the same place. Nothing is published on the site from a first scan until a maintainer merges the result.

4. Re-check every 14 days. Each entry has last_checked and recheck_days. Entries past due are scanned again. A page that moved, an export that changed shape or a number that changed by more than a set margin opens a pull request instead of silently updating. Overdue entries raise one notice, the same way assumptions do.

Open decisions

  1. Which routers get a probe scan first. It needs an account and a key on each one, and a budget line.
  2. The bot account for pull requests, and whether results merge automatically when validation passes.
  3. Estimates (as for Vercel) never come from the registry. Each one is an entry in this repo's assumptions file, added by a maintainer with its sources: see assumptions.
  4. Apps that publish their own usage (Kilo Code, OpenCode) are not routers. They may deserve their own list later.

Order of work

  1. ~~Nightly registry fetch in the collector.~~ Done.
  2. Export scanner with raw capture and the "adapter needed" flag (small).
  3. Results pull requests from a bot (medium: needs the bot account and a token on the box).
  4. Probe scanner for one second router, behind a config switch (medium: needs a key and a budget line).
  5. The 14-day re-check and its notice (small, after 1 to 3).