How we assess
Counts, not grades.
Every number on this site is a count of things we checked against a written rule, read from the vendor’s public pages. Here is the whole method.
01 · Read from public docs
Read from public docs
Every entry is read from the vendor’s unauthenticated surface: documentation, pricing page, repository, well-known files. No login, no sales call, no vendor questionnaire. What the public cannot see, we do not claim.
- 01Read the public pages
- 02Apply the written rule
- 03Record the evidence URL
- 04Count
02 · Four words, no grade
Four words, no grade
Each capability is marked with one of four words against a written rule. A product’s number is how many it passes out of how many were assessed; you will never see an 87/100 of ours.
- SupportedSupported: present and usable as documented
- Partially supportedPartially supported: present but thin, gated, or not as described
- Not supportedNot supported: absent, though the product could have it
- Not assessedNot assessed: not applicable, or not checked yet
03 · The denominator always shows
The denominator always shows
Products carry different denominators: a database has no SSH surface to assess, a sandbox does. So 8 of 11 is honest for one product and not comparable to 6 of 8 for another. We never show a bare ratio or a percentage.
A sandbox
8 of 11
A database
6 of 8
04 · “Not assessed” is not “no”
“Not assessed” is not “no”
Entries scanned before a capability joined our registry say nothing about it: a gap in our data, never a gap in the product. Not-assessed rows stay visible, dimmed, so you can see what we have not checked.
- MCP serverSupported
- CLIPartially supported
- Snapshots / forkingNot supported
- GPU computeNot assessed
- Egress controlsNot assessed
1 of 3 assessed · 2 not assessed
05 · Every claim cites its evidence
Every claim cites its evidence
Each row keeps the note from the scan and the URL it was read from. Often it is the vendor’s own wording; where something is absent, the note says what was looked for and where.
“Install the CLI with one command, then log in from your terminal.”
Source docs.example.com/cli
- ① status
- ② the rule it was judged against
- ③ the note
- ④ the source URL
06 · Isolation, as the docs name it
Isolation, as the docs name it
We record the isolation primitive the public docs name and place it on one ladder, from a language sandbox that shares the host process to a full virtual machine. When the docs do not say, the entry reads “Not disclosed”, never a guess.
- Virtual machinehypervisor, or nothing at all
- Firecracker microVMshares only the hypervisor
- Kata / gVisorhardened or user-space kernel
- Containershares the host kernel
- Language sandboxshares the host process
Not disclosed · the public docs do not say
07 · Works with: advertised, not inferred
Works with: advertised, not inferred
A coding harness appears on a product only when the vendor advertises support for it, with the page that says so. Nothing is inferred from an API shape or a community post.
- Claude Codeadvertised · source linked
- Cursoradvertised · source linked
- inferred never shown
08 · Pricing from the pricing page
Pricing from the pricing page
Tiers are read from the vendor’s pricing page into a machine-readable shape: price, unit, free tier, card required, overages. The “from” price is the cheapest priced tier; “not disclosed” stays not disclosed.
Free
$0
Pro
$20/mo
Enterprise
contact
09 · Agent readiness: can an agent read it, can it drive it
Agent readiness: can an agent read it, can it drive it
Two counts, each over a fixed list of checks of the public surface. Read: llms.txt, an AI-crawler policy in robots.txt, a sitemap, structured data, markdown on request, discoverable docs, no bot wall on public content. Drive: a documented API and spec, an MCP server, SDKs, auth and self-serve credentials, webhooks, a CLI, an error model, rate-limit docs.
Read
- llms.txt
- AI-crawler policy
- sitemap
- structured data
- markdown on request
- discoverable docs
- no bot wall
Drive
- API + spec
- MCP server
- SDKs
- auth
- self-serve credentials
- webhooks
- CLI
- error model
- rate-limit docs
10 · What this atlas refuses to do
What this atlas refuses to do
- Grade: no 0–100 score, no stars, no ranking.
- Guess: when the docs do not say, the entry says not disclosed.
- Take the vendor’s word for it: a correction needs a public source we can link.
- Claim what needs a login: everything is read from the pre-login surface.
- Hide its age: every product page shows when it was scanned.
11 · Dispute it
Dispute it
Something wrong? Every product page has “Suggest a correction”; pick the section, tell us what it should say and where you saw it. We check the public source and fix the entry.